# Lorikeet - Full Content > This file contains the complete text of all articles, blog posts, customer stories, and glossary entries published on the Lorikeet website (www.lorikeetcx.ai). Lorikeet is an AI customer support platform for complex and regulated businesses. --- # Blog Posts --- # Revolut just launched an AI assistant. You can launch yours this week. URL: https://www.lorikeetcx.ai/blog/revolut-just-launched-an-ai-assistant-you-can-launch-yours-this-week **Launch your AI assistant with Lorikeet SDK** On April 9, 2026, Revolut launched AIR - an AI financial assistant embedded directly in their app. Customers swipe down and ask "freeze my physical card" or "how much did I spend on subscriptions last month?" Three weeks earlier, Starling Bank shipped what it called the UK's first agentic AI financial assistant. HSBC, Barclays, and NatWest are all signing deals with AI labs to build their own. You can ship the same thing in days using [Lorikeet's SDK.](https://www.lorikeetcx.ai/integrations/lorikeet-sdk) Not a chatbot. Not FAQ search. A fully embedded AI agent that takes real actions inside your product - freezing cards, setting up transfers, checking insurance - the exact capability the biggest fintechs in the world just made headlines for. ### What they actually built Strip away the press release and AIR is a straightforward architecture. A conversational interface sits inside the app. A customer types a request. The system pulls their data, reasons about it, and takes an action. Chatbots summarise help articles. Agents do things. That distinction is what makes these launches significant, and it's what every fintech customer is about to expect as standard. Revolut first teased this at their ["Revolutionaries" event in November 2024](https://www.revolut.com/news/revolut_reveals_2025_vision_with_ai_assistant_mortgages_and_atms_on_the_horizon/). Starling describes theirs as [the culmination of eight years of AI development](https://www.starlingbank.com/news/starling-launches-agentic-ai-assistant/). Both are UK-only, with other markets "coming soon." ### How Lorikeet's SDK works Lorikeet's SDK is a REST API that lets you embed a fully capable AI agent inside your product. Your engineering team builds the interface - a chat bubble, a swipe gesture, whatever fits. The SDK handles everything else: natural language understanding, action execution, compliance guardrails, multi-turn conversations, streaming responses. Your product, your experience, your brand. The AI is invisible infrastructure. Exactly the architecture Revolut built for AIR - except you plug in rather than build from scratch. The question most fintechs ask next is: how do you actually get the agent to do useful things? Connect to your systems, follow your policies, handle your edge cases? That's where Coach comes in. Coach is Lorikeet's AI-powered setup assistant. Describe what you want your agent to handle - "when a customer asks to freeze their card, verify their identity, freeze it, and offer a replacement" - and Coach generates the workflow, the API connections to your backend, the compliance guardrails, and the test scenarios to validate it all works. What used to take weeks of configuration is now a conversation. ### You don't need to worry about compliance [Lorikeet is SOC 2 and ISO certified. ](https://trust.lorikeetcx.ai/)The patterns for financial services - data handling, PII protection, audit trails, escalation rules - are built into the platform and tested in production. Fintechs like Airwallex, Self, Taptap Send, and Step already run Lorikeet agents across their customer support operations. They handle real financial queries with real compliance requirements every day. The platform works for fintech because it's been shaped by fintech. You don't need to solve compliance from scratch. You configure it for your specific policies, and Coach helps you do that too. ### The compounding advantage of going first Revolut is UK-only. Starling is UK-only. The traditional banks are still in partnership announcement territory. If you're operating in the US, Europe, or APAC, your customers haven't seen this yet. You have a window. Ship now and by the time the big players roll into your market, your agent has months of real conversations behind it. Months of edge cases handled, workflows refined, escalation paths tested. That compound advantage is hard to replicate once someone else has it. Of course, speed alone isn't enough. Klarna learned that the hard way - their AI assistant handled 2.3 million conversations in its first month, but quality collapsed and they ended up re-hiring human agents by mid-2025. The conversational interface is the easy part. Getting the agent to reliably resolve problems across thousands of scenarios is the hard part, and it's the part you shouldn't be building from scratch. ### Revolut spent a year. You don't have to. Embed a capable agent into your existing product. Connect it to your systems. Let Coach build the workflows. Deploy in days. Your customers don't know or care what's under the hood. They care that it works. --- # Guardrails: Runtime Protection for Your AI Agent URL: https://www.lorikeetcx.ai/blog/guardrails-runtime-protection-for-your-ai-agent **Guardrails: Runtime Protection for Your AI Agent** For regulated, high stakes use cases in telehealth, financial services, insurance and other sectors, there are some customer queries that are time urgent and must absolutely comply with deterministic scripts. Typical examples involve regulatory compliance, financial advice disclaimers, safety-critical escalations. When the stakes are high, AI agents in deployment must perform tasks accurately and precisely, 100% of the time. Guardrails is the runtime layer in Lorikeet's [defense in depth approach](/blog/defence-in-depth-how-to-actually-manage-ai-accuracy-in-production) to AI accuracy. While other layers handle agent quality at the foundation, pre-deployment testing, and post-ticket QA, Guardrails operates in real-time, evaluating every message and response as conversations happen. ## Always-on protection Every Lorikeet agent ships with built-in guardrails that run automatically. For example: - **Response grounding** ensures agent responses are based on your knowledge base, data, or instructions - **Profanity filter** prevents inappropriate language - **Jailbreak detection** blocks prompt injection attempts before they reach the agent These guardrails are always on and don’t need to be configured because they're always on, protecting you from day one. ## Custom guardrails for your business In addition to always-on protection, every business has specific policies, industry regulations, or edge cases that matter to them. That's where custom guardrails come in. Custom guardrails let you define your own checks. For example: - Financial services: "Agent must not provide specific investment advice" - Insurance: "Agent cannot estimate claim values" - Healthcare: "Escalate immediately if customer mentions self-harm" - Any industry: "Never mention competitor products by name" Because the check runs outside the agent's reasoning loop, it produces an unbiased result. The agent can't talk itself out of a violation. ## Two layers of checks **Message checks** evaluate incoming customer messages before they reach the agent. Financial vulnerability, legal threats, life-or-death situations, these get flagged immediately so you control what happens next. **Agent guardrails** check outgoing responses before they're sent. If the agent is about to offer an unauthorized refund, share incorrect information, or respond in a way that violates policy, the guardrail blocks it. ## What happens when a guardrail triggers When a guardrail fires, you choose what happens: - **Alert:** Log for analytics without interrupting the conversation. One customer uses this to monitor how often users report app errors; spikes indicate a production issue. - **Apply a tag:** Categorize for routing or reporting. - **Send Slack message:** Ping a channel in real-time. - **Escalate:** Hand off to a human immediately. - **Guide the agent:** Inject just-in-time instructions. If a customer mentions a specific error code, tell the agent exactly how to resolve it. - **Run a workflow:** Trigger a specific workflow for highly sensitive situations. - **Silently escalate:** Queue for human review but let the agent finish responding first. ## Testing and iteration Every custom guardrail can be tested with saved scenarios with exact customer messages, draft agent responses to verify correct behavior. Coach helps refine detection criteria until guardrails trigger reliably on the right situations and stay quiet on the rest. ## Configure via Coach or MCP Following our launch last week of Lorikeet MCP, everything you can do with custom guardrails including create, test, update, monitor guardrails, are accessible through Lorikeet Coach and MCP. Use Coach for conversational configuration, or integrate directly via MCP for programmatic control. ## Analytics and auditability When guardrails trigger, you see exactly what happened: the blocked response, the explanation, and links to affected tickets. Analytics show trigger frequency over time, broken down by type and action. This visibility feeds the broader quality flywheel: patterns surface, root causes get identified, fixes get validated through simulation, and monitoring confirms the improvement. *Guardrails is one layer in Lorikeet's defense in depth architecture. Read the *[*full framework*](/blog/defence-in-depth-how-to-actually-manage-ai-accuracy-in-production)* to understand how training, simulation, runtime checks, and post-ticket QA work together.* --- # How we built Lorikeet MCP to be Safe and Powerful URL: https://www.lorikeetcx.ai/blog/how-we-built-lorikeet-mcp-to-be-safe-and-powerful **Building Lorikeet MCP to be Safe and Powerful** *How we built an MCP server that can handle complexity, and the engineering decisions that make it safe and powerful.* [Lorikeet Coach](/product/coach) and [Lorikeet MCP ](https://docs.lorikeetcx.ai/mcp/mcp-server#claude)are available to use within our app, in [Claude](https://claude.ai/directory/0ffe8735-4ad5-4e7b-99e3-c8fc25f51b52), Cursor, ChatGPT and more. The Model Context Protocol ('MCP') has become the standard way for AI agents to interact with external tools. We're now bringing Lorikeet to this ecosystem, letting CX and growth leaders build concierge experiences their customers can access through any MCP-compatible agent. However, as we built our MCP server, we quickly realised our use cases are significantly more complex than most. ## The Complexity Problem Take a look at typical MCP implementations. A project management SaaS product might have a handful of concepts: issues, statuses, projects. A digital workspace product covers pages, comments, and teams. This is relatively simple. Lorikeet is entirely different. We have tickets, customers, workflows, a knowledge base, integrations, simulations, tools, guardrails, and more. Beyond that, we give AI agents the ability to *test their own work* by creating and running tool tests and managing simulations. The concept count is an order of magnitude higher than a typical SaaS product. This creates an engineering challenge. Historically, throwing more tools at a language model was a bad idea. Performance degrades as the tool count increases as models get confused, pick the wrong tools, or hallucinate parameters. ## Decision 1: Lean into Tool Search Modern AI agents have built-in tool search. Claude Code, for example, doesn't load all tools upfront, it searches by keyword and loads what it needs. We designed around this. Our tool naming and descriptions are deliberately semantic and follow a standard well-documented convention. When an agent searches for "tickets," it finds all ticket-related tools. In Lorikeet, a workflow is the set of instructions and tools that tells an AI agent how to handle a customer enquiry . When an agent searches for "workflow," it gets all workflow-related tools. This seems obvious, but it requires discipline. Every tool name and description is written with searchability in mind. ## Decision 2: Semantic Error Messages Error messages are an underrated steering mechanism. A typical API returns `404 Not Found`. Ours returns something like: *"Workflow not found. You might want to call *`*search_workflows*`* with these parameters to find what you're looking for."* Every error message is written to help the model recover. Invalid ID? Here's how to look up the correct one. Missing permission? Here's which scope you need. The model doesn't fail, rather it learns how to succeed on the next attempt. ## Decision 3: Encode Processes into Tool Descriptions Lorikeet lets support teams define workflows that tell the AI agent how to handle specific scenarios. These range from structured workflows that follow exact steps to natural language workflows that give the agent flexibility to adapt. In Coach, our AI assistant, we can guide users through a workflow using a library of [skills](https://agentskills.io/home): first fetch the context, then read best practices, then make your edit. But the MCP predates skills. While the protocol lists resources and prompts, support is patchy and we want our MCP to work with as many AI agents as possible. So we encode processes directly into tool descriptions. For example, the nudge AI agents trying to edit a workflow are prompted like this: *Before calling this tool, you should first get the context of available tools and variables, then read our best practices."* We're essentially programming the agent's behavior through documentation. It sounds like a hack, but it works remarkably well. The model follows the suggested sequence, and the result is more reliable than if we'd left it to figure out the order itself. ## Decision 4: Enable Self-Verification A model works better when it can verify its own work. This insight drove several design decisions. After an agent configures a workflow, we nudge it to create a test ticket, role play as a customer and verify the workflow behaves as expected. After it configures a tool or a guardrail, we encourage it to add tests and run them. This isn't just about catching errors. It's about giving the model a feedback loop. When it can test its work and see the results, it can try again and course correct in response to external feedback. We even built regression testing into the flow. Change a tool later? The agent can re-run the existing tests to verify nothing broke. Simulations are central to this. A scenario defines a customer interaction, mocks the tool responses the agent would normally get from real APIs, and asserts what should happen with the right response, the right tags, the right data extracted. Agents can run these against draft workflows, see what fails, fix the instructions, and re-run until assertions pass. ## Security: Scopes, Audit Logs, and Safety Checks Giving an AI agent write access to your customer support system is a big deal. We took security seriously. **OAuth scopes**: Our MCP supports fine-grained scopes. Want read-only access? You can restrict the agent to viewing tickets and workflows without modification rights. The permission model maps directly to our existing user roles: a QA user connecting via MCP has the same permissions they'd have in the web app. **Audit logs**: Every action taken through the MCP is logged. If an agent makes a mistake, the agent can look through the audit log and undo its decisions one step at a time. An audit log gives us the confidence to expose powerful capabilities to the MCP by making every tool call reversible. And we don’t just log what happened, but also why it happened. That last part is important. We require agents to provide a short natural language reason explaining why its calling a tool. This creates an audit trail. When you're reviewing what happened, you don't just see "workflow edited", you see *why* the agent thought it should make that edit. **Safety checks on dangerous operations**: Some operations need guardrails. Deleting a tool? We check if any workflows use it. If they do, the deletion fails with a clear message: remove the tool from those workflows first. This prevents cascading failures. An over-eager agent can't accidentally break production workflows by deleting a tool they depend on. ## The Result Lorikeet's MCP is, frankly, more complex than most. But complexity doesn't have to mean fragility. By designing for searchability, writing semantic error messages, encoding processes into tools descriptions, enabling self-verification, and building in real security controls, we've created an MCP that's both powerful and safe. AI agents using our MCP can configure workflows, build tools, test their work, and iterate while staying within appropriate guardrails. That's the goal: not to limit what AI can do, but to make it safe for AI to do more. * Want to see how Lorikeet Coach and Lorikeet MCP lets you configure AI-powered customer support through conversation? *[*Book a demo*](about:blank#) ## Should you build or buy AI support? If you’re evaluating whether to build your own AI support agent or buy a platform like Lorikeet, the decision comes down to 3-year cost, rebuild cycles, and time to value. We built a free interactive calculator that runs the comparison with your real numbers. [**Try the Build vs Buy Calculator →**](https://lorikeet.tools/build-vs-buy?utm_source=blog&utm_medium=cta&utm_campaign=toolshed) --- # How CX Leaders Actually Use Lorikeet MCP and Coach URL: https://www.lorikeetcx.ai/blog/how-cx-leaders-actually-use-lorikeet-mcp-and-coach **How CX Leaders Actually Use Lorikeet MCP and Coach** [Lorikeet's Coach](/product/coach) and [MCP (Model Context Protocol)](https://docs.lorikeetcx.ai/mcp/mcp-server#claude) integration give you conversational control over your entire Lorikeet platform - you can build inbound and outbound workflows for customer engagement across the entire lifecycle. You can configure workflows, diagnose tickets, audit your knowledge base, and analyze performance, all through natural language. Here are real customer stories about what this change means in practice. ## Elevate From Operator to Strategist The hidden cost of manual configuration and auditing isn't just time. It's focus. When operators spend their days in the weeds, strategy suffers. Teams grow to handle operational load instead of expanding capability. Coach changes the ratio. Less time operating the system. More time improving the business. **What this looks like:** - "Set up a simulation batch for the refund workflow with 3 different customer scenarios and run them." - "Compare this week's performance to last week. What should I prioritize?" - "Draft a summary of automation wins this month for the leadership update." Small teams can own more. Operators become architects. The work gets more interesting. [Arbor](https://www.joinarbor.com/) is a leading energy broker helping households across 13 US states save on electricity. With a [4.7-star rating on Trustpilot](https://www.trustpilot.com/review/joinarbor.com) and $7.5M+ in documented customer savings, their AI agent Olive has become a standout, and their Head of Customer Operations now focuses on strategy instead of audits: > "For me personally, **it means rising above the operational minutiae and being able to focus on strategy. **For my team, they can have expanded ownership, and we can keep the team small and engaged, make the work exciting and worthwhile. **We no longer have to rely on manual work or review. We just have a conversation with Coach to diagnose and configure Lorikeet.**" **Jess Mishlove, **Head of Customer Operations, Arbor ## Configure Through Conversation Traditional automation requires translating business logic into system logic: building nodes, adding tools, testing iterations. It's slow, technical, and keeps operators dependent on engineering support. With Coach, you describe what you need. The system builds it. **What this looks like:** - "Create a natural language workflow that handles refund requests by checking order status, verifying the return window, and processing the refund." - "Add an API tool that calls our order management system's GET /orders endpoint with the customer's email." - "Here are the 13 paths a customer can go through for this scenario. Build them out." Workflows, tools, integrations, auth configs, scrapers: the entire configuration surface is accessible through conversation. [FoundersCard ](https://founderscard.com/)is a global membership community of over 250,000 entrepreneurs, founders, and executives, offering 500+ exclusive benefits across travel, business, and lifestyle. Their Member Experience team uses Coach to provide an exceptional concierge-level service and stay ahead of issues at scale: > "Coach is quickly becoming a game changer for our team. It lets us efficiently diagnose, troubleshoot, and build workflows in real time. It's incredibly intuitive, and the ability to take action directly within the tool is a standout feature. It's helping us work smarter, move faster, and ultimately deliver a better member experience." **Rachel Lesnik**, Senior Manager, Member Experience, FoundersCard [Talkspace](https://www.talkspace.com/online-therapy/) is a leading online therapy platform connecting millions of people with licensed therapists through text, audio, and video making mental health care accessible anytime, anywhere. Their Client Support Content team uses Coach to build and refine their customer concierge agent: > "Coach is a great assistant for setting up the Lorikeet agent. It can suggest the best approach for workflows, guardrails, or style guides that define the behavior of the agent, directly implement the agreed approach for me, or guide me step-by-step if I want. It can also simulate performance and help me evaluate tickets, spot issues, and resolve them to prevent similar problems." **Wendy Zheng**, Client Support Content Lead, Talkspace [Mosh](https://www.getmosh.com.au/) is leading men's digital health clinic, connecting patients with registered doctors via telehealth for hair loss, sexual health, skincare, and mental health. Their Customer Service AI Lead went from week-long workflow builds to same-day shipping: > "Game changing. Before, I was making individual nodes, adding tools one by one, testing each iteration. That would take a week or two. Now I just explain the scenario. It starts creating the nodes and paths for me, and I fine-tune it. **It made me feel like a conductor.**" **Joel Sheehan, **Customer Service AI Lead, Mosh ## Diagnose and Act in Real-Time When a question comes up about a specific ticket, there is no need to dig through logs and tracing multiple sources. Traditional debugging meant clicking through UIs, cross-referencing logs, rebuilding context every time you return to an issue. Coach understands your configuration, your tickets, and your history. Root cause analysis that used to take 20 minutes takes seconds. And when you're ready to fix it, you can make the change in the same conversation with Coach. **What this looks like:** - "Pull up all escalated tickets from the last month. What went wrong and why did the customer get escalated?" - "Show me all tickets from the last hour where the refund workflow failed." - "The automation rate dropped this morning. What changed?" [Flex ](https://getflex.com/)is a fintech platform trusted by over two million renters to split rent into smaller payments. Christina Fernandez leads CX operations and has made Coach her go-to tool for understanding what's happening across support tickets in real time: > "When someone raises a question about a ticket, I just drop it into Coach and ask what happened. **I get the answer and auditable source material in one conversation, instantly. **I can get quick insights without switching tools" **Christina Fernandez**, CX Operations, Flex [Summ](https://summ.com/) (formerly Crypto Tax Calculator) is a crypto tax software platform, backed by Coinbase Ventures and supporting 3,500+ integrations across exchanges and DeFi protocols. When tax rules change or edge cases emerge, their team needs to move fast: > "When I'm under pressure and need to make changes fast, Coach just gets it. I jump in, explain where I'm at, and it infers exactly what I'm trying to do. No rebuilding context, no wasted time, just high-quality outcomes. **It raises the ceiling for what I can actually deliver.**" **Ben Melbourne**,** **Senior Customer Support Specialist, Summ ## Understand Performance at a Glance CX leaders are often drowning in data but starving for insight. Metrics live in dashboards. Quality issues hide in ticket logs. Knowledge gaps only surface when customers complain. Coach turns analysis into conversation. **What this looks like:** - "What's our automation rate for the last 30 days, broken down by channel?" - "Review all 500 of our knowledge base articles. Which ones are outdated, which topics have gaps, and are there any quality issues? Suggest fixes and new articles for me to review" - "Which workflows have the highest escalation rate this week?" [Eucalyptus](https://www.eucalyptus.vc/) is a global digital health provider, operating brands like Juniper, Pilot, and Kin across weight loss, men's health, and reproductive care. Recently [acquired by Hims & Hers for $1.15 billion](https://investors.hims.com/news/news-details/2026/Hims--Hers-Announces-Agreement-to-Acquire-Eucalyptus-Accelerating-Its-Vision-to-Become-the-Leading-Global-Consumer-Health-Platform/default.aspx), their Automation Lead uses Coach to cut through complexity at scale: > "I used to dig through data and stitch together reports just to get a read on things. It was slow and often unclear what actually mattered. Now Coach helps me spin up insights in minutes, understand what's happening and where to focus, and set our team's priorities with confidence. **It makes acting on data much more straightforward.**" **Mia Dela Peña, **Automation Lead, Eucalyptus [Berry Street](www.berrystreet.co) is a telehealth platform and one of the largest networks of nutrition specialists in the world. Berry Street connects patients with registered dietitians across 1,250+ insurance plans nationwide. Their Platform Experience and Operations Lead uses Coach to bridge the gap between complex systems and actionable fixes: > "As a mostly non-technical person, Lorikeet's Coach feature has been a game-changer in helping me understand exactly where a workflow or knowledge base went wrong and why. It translates complex issues into clear, actionable insights that empower me to tweak workflows independently and with confidence." **Dakota Reed**, Platform Experience and Operations Lead, Berry Street ## Spot Bottlenecks and Move Fast Coach doesn't just surface problems, it helps you fix them. Teams use it to identify what's blocking automation performance and accelerate their delivery timelines. [easykind](https://easykind.com.au/) is a global telehealth clinic operating in a highly regulated market. Their Solutions Engineer uses Coach to maintain quality at speed: > "I use Coach to identify bottlenecks across my workflows, enabling me to drive up independently resolved rate. Coach gave me clear insights and massively accelerated our delivery timeline. It felt like having an expert on demand." **Nina Brown**, Solutions Engineer, easykind [Weel ](https://letsweel.com/)is a spend management platform helping businesses control company spending with smart corporate cards and expense management. Their team uses Coach to iterate quickly with confidence: > "Lorikeet makes it easy to move fast, spot gaps early, and put the right guardrails in place. Coach feels less like a tool and more like having an expert alongside and executing things for you." **Harry Burnard,** Senior Customer Support Specialist, Weel ## What You Can Do With Coach + MCP Example categories Example Prompts **Workflow building** "Create a refund workflow that checks order status, verifies return window, and processes the refund." **Ticket diagnosis** "Pull up all escalated tickets from the last month. What went wrong and why did the customer get escalated?" **Knowledge management** "Review all our knowledge base articles. Which are outdated and where are the gaps?" **Tool configuration** "Add an API tool that calls our order management system with the customer's email." **Simulations** "Set up 3 test scenarios for the refund workflow and run them." **Metrics analysis** "What's our automation rate for the last 30 days, by channel?" **Performance audits** "Which workflows have the highest escalation rate this week?" ## How ready is your CX team for AI? Before deploying AI tools like Coach and MCP, it's worth assessing your team's readiness across six key dimensions — from knowledge management to change management. Our free scorecard takes five minutes and tells you exactly where to focus. [Take the AI Readiness Scorecard →](https://lorikeet.tools/ai-readiness?utm_source=blog&utm_medium=cta&utm_campaign=toolshed) ## Get Started Coach is available to all Lorikeet customers. MCP integration works with Claude, Cursor, and other AI tools that support the Model Context Protocol. --- # Lorikeet is now self-configuring and agent-first URL: https://www.lorikeetcx.ai/blog/lorikeet-is-now-self-configuring-and-agent-first **The future of customer engagement platforms: simply tell Lorikeet what to build** The way companies serve customers is being rewritten. Customers expect personalized, context-aware experiences at every touchpoint. They expect companies to know their history, anticipate their needs, and resolve issues before they escalate. The companies that win will build and ship these experiences at light speed. The ones that lose will be stuck in silo’d legacy systems: CX teams clicking through complex UIs to debug workflows they can barely see inside, marketing and growth teams pushing out offers with rule-based logic, and reliant on engineering or vendors for configuration. Lorikeet changes this. While others build tools to help agents respond faster, Lorikeet has built itself to be self configuring and agent-first. All you have to do, as a Lorikeet user, is supervise through a conversational UI. Lorikeet is an AI concierge for the entire customer lifecycle: proactive, personalized, conversational. Under the hood, it balances natural language flexibility with deterministic logic for accuracy and compliance. The command center is a conversational layer via Lorikeet Coach and Lorikeet MCP. You describe what you want. It gets built. From Claude Code, Codex, ChatGPT, wherever you already work. ## Agent-First, Not Agent-Assisted Most AI support tools help human agents respond faster. Draft suggestions, auto-complete, copilot-style assistance. Lorikeet is different. We're agent-first: you build autonomous agents that handle tickets end-to-end. Humans configure, test, and audit, but they don't need to touch every conversation. This is why self-configuring matters. When you're building autonomous agents (not helping humans respond), you need to move fast: test scenarios, adjust behavior, deploy changes. Lorikeet MCP enables exactly this: building the agent itself, through conversation. ## From Lorikeet Coach to Lorikeet Everywhere We introduced a conversation-first approach to building with **Lorikeet Coach**, which allows subscribers to configure workflows, test automations, and audit decisions through conversation. The response was immediate: > *"Game changing. Before, I was making individual nodes, adding tools one by one, testing each iteration. That would take a week or two. Now I just explain the scenario: 'here are the 13 paths a customer can go through.' It starts creating the nodes and paths for me, and I fine-tune it. It made me feel like a conductor." ***Joel Sheehan, Customer Service AI Lead, Mosh** Today, we're taking it further. **Everything you could do in Coach is now available wherever you already work: Claude Code, OpenAI Codex, ChatGPT, and any MCP-compatible tool.** Lorikeet is now self-configuring. You investigate data, test and build yourself, through conversation. ## In Practice: Proactive Customer Re-Engagement Here's what using Lorikeet MCP looks like in the real world. A specialty lending company wants to recover lapsed customers. Traditional lifecycle marketing wasn't working - static outbound messages with low conversion. With Lorikeet, they combined internal data from their data warehouse with ticket signals: payment history, past support interactions, account activity, behavioral patterns. Then they built a proactive outbound workflow that didn't just send a templated message - it started a conversation. When a customer responded, Lorikeet understood their full context. It could reference their specific situation, answer questions about their account, and offer personalized options based on their history. **The result: 60% improvement in recovery rates** compared to their prior outbound engagement. This is the difference between lifecycle marketing and an AI concierge. You're not limited to static offers triggered by simple rules. You act on rich signals and deliver fully personalized experiences through conversation at scale. ## Configure, Audit, and Build Without Vendor Dependency With [Lorikeet Coach](www.lorikeetcx.ai/product/coach) and [Lorikeet's MCP](https://docs.lorikeetcx.ai/mcp/mcp-server#lorikeet-mcp-server) integration, you can: - **Build workflows conversationally:** Describe what you want automated, and Lorikeet constructs it step by step - **Test Simulations:** Run quality testing and simulations prior to launch - **Audit everything:** Ask why a ticket was handled a certain way and get a traceable answer - **Self-heal:** Get recommendations on how to improve and implement those recommendations - **Iterate instantly:** Test changes against historical tickets, see results, adjust, deploy The people closest to your customers, your CX team, marketing team, growth teams—can now own campaign set up and automation directly. > “A great use case is that any user of Lorikeet is able to describe the behaviour they want from their customer Concierge agent, and have Coach and MCP create simulations of prompts and how conversations might behave. Lorikeet can autonomously experiment with prompt changes to improve the performance.” - **Jamie Hall, Lorikeet Co-Founder** ## Supercharge Customer Engagement: Combine Lorikeet With Your Existing Data and Tools Lorikeet doesn't operate in isolation. MCP connects multiple tools in a single conversation, so you can work with Lorikeet alongside everything you already use: databases, CRMs, analytics, internal APIs, Slack, and more. **In a single conversation, you can:** - Pull ticket patterns from Lorikeet, cross-reference with your product database, identify what's driving support volume - Query your CRM, check Lorikeet's handling history, draft a response, without switching tabs - Debug end-to-end: from Lorikeet's decision logs through your application logs to root cause - Update a workflow based on analytics insights, test it, deploy, all conversationally Your AI assistant becomes a unified interface across your entire stack. ## Built-In Guardrails Conversational configuration doesn't mean uncontrolled. Lorikeet's MCP integration includes granular role-based access with separate toggles for read and write permissions, so teams can audit decisions without risking configuration changes. Every workflow can be tested against historical tickets in simulation mode before touching a customer. You can generate test suites from real ticket patterns, run them on every change, and roll back instantly if something isn't right. Every decision is explainable and traceable. ## The Future We're Building Toward The traditional model (navigating complex UIs, submitting requirements to vendors, waiting for changes) doesn't scale. It separates the people who understand your customers from the people who control the tooling. We're building something different: an AI concierge that delivers proactive, personalized customer experiences, powered by the complexity necessary for real-world accuracy, but accessible through natural conversation. Coach was the first step. MCP integration is the next. **Ready to try it?** Connect to Lorikeet from [Claude Code](https://claude.ai/directory/0ffe8735-4ad5-4e7b-99e3-c8fc25f51b52), Codex, or any MCP-compatible tool and start building through conversation. --- # AI Agents for Customer Support: A Hype-Free Buyer's Guide URL: https://www.lorikeetcx.ai/blog/the-definitive-guide-to-ai-agents-for-customer-support **The AI support market is full of vendors overselling. This guide breaks down the four types of AI support tools, what metrics actually matter, and what questions to ask before you sign anything.** Every AI support vendor claims their product will "transform your customer experience." Most of them are, at best, overselling what their product actually does. Here's how to cut through the hype. ## Let's talk about how we talk about AI support The term "AI agent" has become meaningless. Vendors apply it to everything from basic FAQ chatbots to sophisticated systems that can actually resolve complex issues. This isn't just confusing. It presents a real challenge for buyers trying to make informed decisions. Most of the content published about "AI agents for customer support" is vendor-written and designed to generate leads, not help you evaluate solutions. We're trying to do something different with this guide. We're going to give you a framework for actually understanding what's on the market, what questions to ask, and how to figure out what you need. Yes, we sell an AI concierge that does support, but we'll save you the hard pitch. If you walk away from this article feeling a competitor is a better fit for your needs, that's fine. Better informed decisions are better for everyone. ### A taxonomy for AI support tools The market breaks down into four distinct categories, and understanding the difference before you start your search will save you months. ### FAQ chatbots **What they do:** Search your knowledge base, surface relevant articles, summarize content for customers. **Best for:** Simple, high-volume queries where self-service is acceptable. Think "what are your hours?" or "how do I reset my password?" **Limitations:** Can't take action. Can't follow complex procedures. When a customer needs something *done*, these tools hit a wall. **Who's here:** Many of the tools calling themselves "AI agents" are actually sophisticated FAQ chatbots. If a vendor can't clearly explain what actions their AI takes beyond answering questions, you're probably looking at a chatbot with better marketing. ### Copilots **What they do:** Draft responses for human agents to review and send. **Best for:** Organizations very early in AI adoption who want a safety blanket. Teams that aren't ready to let AI interact directly with customers (e.g. in a heavily regulated sector where direct AI interaction feels too risky). **Limitations:** The efficiency gains are marginal. Research suggests around 14%, with less impact for more experienced agents. A major potential downside of copilots is that they risk creating roles where humans are just rubber-stamping AI output. If the copilot is good enough to trust, why have a human in the loop? **The hard truth:** Copilots feel safer than they are. At scale, human agents aren't carefully reviewing every draft. They're clicking send to keep up with volume and hit their goals. ### Deflection-first AI **What they do:** Attempt to answer everything. Escalate what they can't answer or divert. **Best for:** Organizations that prioritize volume reduction above all else. **Limitations:** This is where a large chunk of the market sits, and it's where the problems start. If your AI attempts to answer 100% of tickets and successfully resolves 40%, you've created 60 failed interactions. That's a 60% frustration rate. Vendors love deflection-first approaches because they can charge per conversation and claim high "engagement" numbers. But engagement is not the same as resolution. A customer who gives up in frustration counts as "deflected" in the metrics. ### Action-taking concierges **What they do:** Actually resolve issues by taking action within your systems. Process refunds. Update accounts. Coordinate with third parties. Ship replacement products. **Best for:** Complex, high-stakes environments where resolution, rather than deflection, is the goal. **Limitations:** Requires deeper integration with your systems, which in turn requires a more thoughtful implementation. They also require ongoing iteration rather than set-and-forget. **The key difference:** These systems don't just tell customers how to solve problems. They solve them. Consider a customer whose card is declined while traveling after they book an airport transfer. A FAQ chatbot explains fraud detection policies, leaving the customer frustrated. An action-taking concierge blocks the suspicious transactions, creates a new virtual card, coordinates with the taxi company to update payment details, and arranges physical card delivery to the next hotel — all while keeping the customer informed about what's happening. ### The design philosophy that matters The difference between these categories isn't primarily technological. They all use LLMs. They all have access to similar underlying capabilities. The difference is design philosophy. Are you optimizing for **avoiding customers** (deflection) or for **helping them** (resolution)? This isn't a semantic distinction. It shapes every decision a vendor makes: how they price, what they measure, how they train their AI, what features they build. Vendors optimizing for deflection build systems that try to handle everything and succeed at some of it. Vendors optimizing for resolution build systems that know their limits and focus on actually solving problems. The term "AI concierge" is instructive here. A concierge solves your problem. A concierge at a hotel doesn't hand you a FAQ sheet about restaurants. They make you a reservation. ## Characteristics of an AI support concierge What does this look like in practice? What are the characteristics of a concierge that differentiate it from the other types of AI agent on the market? - **Action-taking, not answer-giving.** An AI concierge doesn't just tell customers how to request a refund or agents how to have a difficult conversation. It processes the refund and handles the difficult conversation. - **Multi-channel orchestration.** The concierge can call the merchant, text the customer, email the vendor. Simultaneously if needed. - **Self-awareness.** It knows what it doesn't know. It escalates more complex issues cleanly with full context, not after trapping customers in frustrating loops. - **Judgment calls.** For businesses that empower human agents to make exceptions, an AI concierge can do exactly the same. ## Don't ignore the complexity curve Support tickets follow a power law distribution. Your easiest 50% of tickets might take 5 minutes each. Your hardest 5% might take 50 minutes. The ability to climb that complexity curve is what actually unlocks value from AI deployments. Most AI support tools top out quickly. They handle the simple stuff (password resets, order status, basic FAQs) and escalate everything else. That's fine if your support is mostly simple stuff. But for companies with complex products, regulated industries, or high-stakes customer relationships, the simple tickets aren't where the pain is. If you're evaluating AI support, ask yourself: where does my support team actually spend their time? If the answer is complex, multi-step issues that require judgment and system access, you need AI that can climb the complexity curve. If the answer is high volumes of simple, repetitive questions, a basic FAQ solution might be enough. ## What to actually measure ### Metrics that matter **Resolution rate (not deflection rate).** Did the problem actually get solved? This requires defining what "solved" means for your business and tracking whether customers come back with the same issue. **CSAT on AI-handled tickets.** Are customers satisfied with AI interactions specifically? This is different from overall CSAT. You need to isolate how AI is performing. **Escalation quality.** When AI hands off to a human, does the human have full context? Clean escalations that set up humans for success are a sign of well-designed AI. Messy handoffs that frustrate both customers and agents are a sign of AI that's in over its head. **Ratio of good to bad AI interactions.** This is the real measure of quality. If your AI handles 1,000 tickets and 600 are genuinely resolved while 400 are frustrated customers, that's a 60% success rate. If your AI handles 500 tickets and 450 are genuinely resolved, that's 90%. The second scenario is better even though the total volume handled by AI is lower. ### Metrics that mislead **Deflection rate.** A customer who gives up in frustration counts as "deflected." A customer who finds a workaround without AI help counts as "deflected." This metric can quickly be gamed into meaninglessness. **AI engagement rate.** Optimizing for AI touching more tickets incentivizes bad experiences. The goal isn't for AI to touch everything. It's for AI to succeed at what it touches. **Per-conversation pricing.** When vendors charge per conversation regardless of outcome, they're incentivized to have AI attempt everything. Your incentives and their incentives diverge. ### The question to ask yourself > "If my AI attempts 100 tickets and resolves 40, are the 60 failures creating enough frustration to offset the 40 successes?" For most businesses, the answer is yes. Failed AI interactions don't just fail to help. They actively damage customer relationships and make future AI interactions harder because customers learn to immediately ask for humans. ## Implementation realities vendors will gloss over ### 1) Prompting is coaching, not programming You don't configure an AI agent once and walk away. You iterate. You review conversations, identify gaps, refine instructions, and improve over time. Think of it like managing a team member who learns fast but needs ongoing feedback. Vendors who promise "set it and forget it" are either misleading you or building something that doesn't actually work well. The good news is that AI feedback loops are much faster than human training loops. You can make meaningful improvements in days, not months. ### 2) You don't need perfect documentation first The old advice for outsourcing ("get your house in order before you hand off") doesn't apply to AI the same way. AI feedback loops are fast enough to iterate your way to good processes. Start with what you have. Deploy on a limited, but meaningful, scope. See where the AI struggles. Improve your documentation based on real gaps, not theoretical ones. This is actually more efficient than trying to anticipate everything upfront. ### 3) Integration depth matters more than AI sophistication An AI with access to your systems will outperform a "smarter" AI that can only read help articles. If your AI can look up order status, process refunds, update account details, and coordinate with third parties, it can actually solve problems. If your AI can only search your help center and summarize articles, it's limited to answering questions. When evaluating vendors, ask what integrations they support and how deep those integrations go. "We integrate with Shopify" could mean they can read order data, or it could mean they can process refunds, update shipping addresses, and cancel orders. The difference is everything. ### 4) Testing and evaluation are non-negotiable Any serious deployment needs robust testing before going live. You should be able to run test conversations, validate how the AI handles specific scenarios, and audit decisions before customers see them. Vendors who don't offer strong testing and evaluation tools are either hiding something or don't understand the stakes. In regulated industries especially, the ability to explain why AI made specific decisions isn't optional. ## The competitive landscape Vendor Positioning Lorikeet assessment Intercom Fin FAQ automation + RAG built into Intercom ecosystem Strong for Intercom users with basic needs. Struggles with complex procedures. Good if you're already on Intercom and your support is mostly straightforward. Zendesk AI AI bolted onto comprehensive ticketing Deep Zendesk integration, but AI capabilities are less sophisticated than specialized players. You're buying the ecosystem, not best-in-class AI. Decagon "Agentic AI" for enterprise Well-funded (~$35M ARR per Sacra). Claims agentic capabilities but head-to-heads suggest gaps in complex scenarios. Sierra Celebrity CEO, consumer brand focus ~$104M ARR per Sacra. Strong marketing, targets large consumer brands. Less focused on complex B2B use cases. Ada Pre-LLM legacy with updates "Coach, don't code" messaging is solid. Mixed reviews on complex use cases. Strong foundation but showing age in some areas. Forethought Deflection-focused AI Leads with impressive-sounding metrics but definitions are fuzzy. Deflection-first philosophy means high engagement, variable resolution. Salesforce Agentforce Enterprise AI within Salesforce Scale advantage and deep ecosystem integration. Less specialized than pure-play support AI. You're buying the platform, not the best support AI. ## What about Lorikeet? We should be transparent: we wrote this guide and we sell AI support software. Here's our honest self-assessment. **Where we're strong:** - Complex, high-stakes environments (healthcare, fintech, regulated industries) - Multi-step workflows that require judgment - Multi-channel orchestration — our AI can call vendors, text customers, and email partners simultaneously - Voice support that actually takes action, not just answers questions - Companies with high CX standards who won't accept mediocre AI interactions **Where we might not be the right fit:** - Simple eCommerce with basic FAQ needs — we might be more than you need - Organizations that want set-and-forget with no iteration — that's not how this works - Teams that measure success by deflection rate rather than customer outcomes — we're not optimized for that metric **What we've seen in head-to-heads:** - Flex compared us to Decagon directly and chose Lorikeet. They saw 2x CSAT improvement and 50% faster resolution. - Magic Eden switched from Intercom Fin and saw CSAT jump from 45% to 74%. - Linktree evaluated Fin, Decagon, and Assembled AI before choosing us. Auditability and control were the deciding factors. - Arbor Health saw [TODO: confirm Arbor metrics] after switching to Lorikeet for their regulated healthcare workflows. We're not claiming we're best for everyone. We're claiming we're best for a specific type of customer: complex businesses with high standards who care about resolution, not deflection. ## Questions to ask any vendor Before you sign anything, get clear answers to these questions: - **What's your definition of "resolved" and how is that measured?** If they can't give you a clear, auditable definition, be skeptical of their metrics. - **Can we see actual conversation logs from customers similar to ours?** Demos are choreographed. Real conversations reveal real capabilities. - **What happens when the AI can't help? How does escalation work?** The handoff experience matters as much as the AI experience. Bad escalations create frustrated customers and frustrated agents. - **How do you price, and how does that align with our goals?** Per-conversation pricing misaligns incentives. Per-resolution pricing is better but requires clear resolution definitions. Understand what you're paying for. - **What testing and evaluation tools do you provide?** If we can't test before deploying and audit after, we're flying blind. - **What actions can your AI actually take, beyond answering questions?** Get specific. "Integrates with Shopify" isn't the same as "can process refunds in Shopify." - **Who are your customers in our industry, and can we talk to them?** References matter. If they can't connect you with similar customers, ask why. ## The mindset shift AI support isn't about avoiding customers. It's about serving them better at scale. The vendors who understand this are building fundamentally different products from those who don't. They measure resolution, not deflection. They focus on quality of interactions, not quantity. They invest in action-taking capabilities, not just answer-giving. When you're evaluating AI support, the question isn't "which vendor has the best AI?" The question is "which vendor's design philosophy matches what we're trying to achieve?" If you want to reduce support costs by making it harder for customers to get help, there are tools for that. If you want to deliver better support at scale by actually solving customer problems, there are tools for that too. They're not the same tools, even if they all call themselves "AI agents." --- # For Insurers, the Problem isn't Hallucination - it's Liability URL: https://www.lorikeetcx.ai/blog/for-insurers-the-problem-isnt-hallucination-its-liability **Why the real barrier to AI in insurance support isn't the technology. It's that most vendors can't prove what their AI said, or why.** *Why the real barrier to AI in insurance support isn't the technology. It's that most vendors can't prove what their AI said, or why.* In 2024, an Australian insurer received an enforceable undertaking from ACMA for customer communications that violated regulatory requirements. The content was written by humans, reviewed by compliance, and part of an approved script - but it was still wrong. That was with humans following a process designed to prevent exactly this outcome. Now the insurance industry is being asked to trust AI that generates content on the fly. We've had dozens of conversations with insurance CX leaders over the past year, and nearly all arrive at the same place: "We haven't stress-tested our position on using AI that can generate content on the fly, even with predetermined outcomes." This isn't resistance to AI, it's institutional memory. These are people who know what happens when customer-facing content goes wrong - enforceable undertakings, DOI complaints, E&O exposure - because they've lived through the remediation. They're right to be cautious, but the caution needs to be pointed at the right thing. ## The fintech demo won't save you Every AI support vendor has a fintech demo. It handles a billing dispute, cancels a subscription, maybe processes a refund. It looks clean, contained and impressive. Then they walk into an insurance meeting and present the same demo with different branding. Thankfully there is still hope in the world, because Insurance CX leaders see through this immediately. The regulatory landscape for insurance isn't a stricter version of fintech. APRA imposes prudential requirements on how insurers manage operational risk, including third-party technology. CPS 230, which took effect in July 2025, explicitly extends operational resilience obligations to material service providers. If your AI support vendor is generating customer-facing content about policy exclusions, excess calculations, or claims status, they're a material service provider - whether they've thought about it that way or not. ASIC regulates how insurers communicate with policyholders, including the content of claims correspondence and product disclosure. ACMA enforces rules around direct communications and has a history of pursuing enforceable undertakings when insurers get the language wrong. The gap between "cancel my subscription" and explaining why a policyholder's storm excess is $750 but their flood excess is $1,500 on the same property isn't incremental, it's structural. One requires a refund button, the other requires understanding of peril-specific excess structures, coverage endorsements, and the PDS - plus an audit trail that holds up when a regulator asks how that answer was generated. ## Where wrong answers create liability AI hallucination in insurance is a legal problem, not a user experience problem. Consider the conversations that actually matter in insurance support: a policyholder asks whether their business interruption policy covers supply chain delays. A claimant disputes a settlement amount and wants the reserve calculation explained. A broker needs to know whether a client's endorsement for additional insured status was processed before a loss event. A customer in financial hardship needs guidance on their options under the General Insurance Code. A wrong answer about a policy exclusion, a misquoted benefit amount, a promise of coverage that doesn't exist. In most industries, these create a bad customer experience, but in insurance they create binding obligations or regulatory exposure. The cost of a confident wrong answer is categorically higher than the cost of not answering at all. Most AI support platforms were built around an assumption that makes sense in low-stakes environments: attempt to answer everything, escalate what fails. That architecture is fundamentally misaligned with insurance. When your compliance team has zero tolerance for content errors on coverage questions, claims guidance, and hardship processes, a system that optimises for coverage breadth over precision is a liability generator. The deeper problem is auditability. If a regulator asks "why did your AI tell a policyholder their claim was covered under Section 4.3(b)?" most platforms can show you the output. They can show you the knowledge base article the output was derived from. What they cannot show you is the full decision path: why the system chose that article over three others, what alternative responses it considered, whether it had sufficient confidence to answer at all, and what guardrails it evaluated before responding. Without that trail, your compliance team is signing off on a black box. And when the DOI examiner arrives, "we trusted the vendor" isn't a defence. ## The build trap Insurance companies that hear all of this and decide to build internally are making a rational calculation. If no vendor meets the standard, at least an internal build gives you control. The problem is the assumption that building is a project with a beginning and an end. The underlying models change every six to twelve months. Regulatory requirements evolve - CPS 230 just took effect, and ASIC's guidance on AI in financial services is still developing. Your internal team ships a solution on Claude 3.5, and six months later the model is two generations behind and the compliance landscape has shifted. Every engineer maintaining your internal AI support system is an engineer not working on your claims platform, your underwriting models, your policy admin system. A specialised vendor amortises the rebuild cost across hundreds of customers while testing against millions of real interactions - FNOL lodgements, endorsement processing, coverage disputes, settlement queries. Your internal build tests against your interactions only. The honest tradeoff: building gives you control. It also gives you sole responsibility for every failure, with no vendor to share the regulatory exposure. And it gives your compliance team yet another internal system to audit, maintain, and stress-test through every model migration and regulatory change. ## What your AI vendor should prove The technology to handle insurance support safely exists. The question is whether your vendor was built for it. **Auditability by default.** Every customer interaction should produce a full decision trail - not just input and output, but every intermediate decision. When a policyholder asks about a coverage exclusion, you should be able to trace the AI's response back to the specific PDS section it referenced, see what alternative responses were considered, and confirm which guardrails were evaluated. Your DOI examiner and ombudsman need this. So does your E&O insurer. **Configurable scope.** Your AI should handle what you've approved it to handle, and nothing more. If you haven't approved content for a specific policy type, claims scenario, or hardship process, the system should escalate, not improvise. That means monthly compliance testing before go-live, mandatory human escalation for financial hardship and distress, and topic-level control over what the AI can and cannot say about FNOL, endorsements, coverage, and claims. **Pricing aligned with outcomes.** Insurance already thinks in loss ratios and combined ratios - cost per outcome, not cost per seat. If your vendor charges per interaction attempted, they're incentivised to have your AI attempt everything, including the coverage questions it shouldn't touch. Per-resolution pricing creates a different incentive: the vendor only earns when the policyholder's FNOL is lodged, their endorsement is processed, their claims query is resolved - fully and correctly. **No black box.** You should be able to explain why your AI said what it said. Not because you trust the vendor's assurance, but because you can see the reasoning yourself. At Lorikeet, this is the product we built. Architecture designed for regulated industries from day one: auditability, configurable guardrails, per-resolution pricing, and a focus on the insurance workflows that actually carry risk: FNOL-to-settlement, endorsement processing, coverage questions traced to policy wording, and claims status with full decision trails. The question for insurance CX leaders isn't whether AI works for your industry - it does. The question is whether the vendor you're evaluating can prove it. --- # Pockets of Determinism: How Lorikeet's architecture handles complexity URL: https://www.lorikeetcx.ai/blog/ai-architecture-for-complex-cx **How Lorikeet combines agentic AI with deterministic operations to handle complex support workflows safely - without relying on guardrails alone.** A fintech customer calls about a failed deposit. The agent checks transaction status, verifies account ownership, looks up fraud signals, reviews previous tickets, checks policy exceptions, confirms identity, and initiates the refund. Seven API calls, conditional logic throughout, business rules that vary by customer segment. The customer just sees a conversation. This is what we built and here’s how we did it. ## Where we started In 2024, we published a technical deep dive on [Intelligent Graph](/blog/lorikeet-ai-a-technical-deep-dive), our architecture for deploying AI agents in complex support environments. At the time, pure agentic AI (giving an LLM instructions and letting it handle conversations however it likes) wasn’t ready for production. The failure modes were real: hallucinations, distraction, instructions degrading after half a dozen constraints. The industry was learning this the hard way. Air Canada’s chatbot invented a refund policy and cost the airline real money in court. Sierra, one of the best-funded players in this space, has had public incidents where agents went off-script in ways that made customers uncomfortable. When you’re handling money or personal data, *usually works* isn’t good enough. Intelligent Graph took a different approach. Structured workflows orchestrated everything. LLMs handled focused tasks inside: classify this intent, extract this value, generate a response from these specific talking points. The orchestration layer controlled the conversation, not the AI. It worked. Hallucination rates dropped. Compliance teams could audit every decision. We could handle the conditional, stateful logic that pure chatbots couldn’t touch. And we built deep expertise in structured workflow design for regulated industries. Models got better. The failure modes that made pure agentic approaches risky in 2024 decreased significantly. We saw an opportunity: what if we could combine the conversational flexibility of agentic AI with the guarantees we’d built into Intelligent Graph? The key insight was that these weren’t competing approaches. They were complementary, operating at different layers. ## Where we’re going The evolved architecture puts a natural language agent at the top. It handles the conversation: tangents, clarifications, context switches. A customer in a refund flow asks “can I get store credit instead?” The agent can discuss that and continue without breaking anything. But when the conversation reaches a critical operation (processing the refund, verifying identity, closing an account) the agent calls a structured workflow as a tool. Inside that tool, everything is deterministic. This is where Intelligent Graph lives now. We call this **Pockets of Determinism**: agentic orchestration wrapping deterministic operations. The logic is straightforward. An AI agent can trigger any tool at any time. That’s what makes it conversational. But we can make the tools themselves foolproof, so even if the agent triggers them incorrectly, nothing bad happens. Think of a banking app. A child can tap “Close Account.” The button is always there, but the app checks for remaining balances before executing. The interface is permissive, the operation is strict. As Jamie Hall, our CTO, explains: > “The agent can call the refund tool whenever it wants. But inside the tool, we check: has the agent confirmed with the customer? Have we disclosed the terms? If not, the tool returns ‘conditions not met’ and the agent goes back to do the work. Even if the agent hallucinates that conditions are met, the tool knows they’re not.” Every operation sits somewhere on a spectrum. A knowledge lookup can be fully agentic. A refund needs a conversational wrapper with deterministic execution inside. A compliance-critical flow might be almost entirely structured. You configure this per operation based on consequences. ## Why this matters Most AI architectures put safety at the edges. Guardrails watch outputs and intervene when something looks wrong. The problem is that the agent has already decided to do something wrong by the time the guardrail catches it. You’re in damage control. Pockets of determinism puts safety inside the operations themselves. The tool validates preconditions before executing. You’re not catching mistakes. You’re making them impossible to execute. There’s a security benefit too. When the agent calls a sub-workflow, that workflow runs in isolation. The agent doesn’t see internal state or sensitive data. If your verification flow uses a one-time code, the agent never has access to it. It just gets back “verified” or “not verified.” You can’t leak what you don’t have. ## The result Flex, a buy-now-pay-later platform, went from 0% automation to 85%+ on complex support workflows in under a year. Breeze, a fintech handling transfer disputes and purchase credits, grew from 50% to 82% automation over nine months. These aren’t FAQ deflection numbers. They’re full resolution rates on multi-step, API-heavy, compliance-sensitive workflows. Magic Eden runs both Lorikeet and Intercom’s Fin on different ticket segments. The CSAT comparison is stark: Lorikeet tickets score double what Fin tickets score, in the same customer base, side by side. *Pockets of Determinism is how we architect the agent. For how we validate and improve accuracy in production, see Defence in Depth.* ## FAQ **How do you prevent an AI agent from executing sensitive operations at the wrong time?** By validating preconditions inside the operation itself, not at the agent level. The agent can request a refund whenever it wants, but the refund tool checks whether all required steps have been completed before executing. If the customer hasn’t confirmed, or terms haven’t been disclosed, the tool rejects the request and tells the agent what’s missing. **Can AI agents access customer data they don’t need?** Not if the architecture isolates sub-workflows. When an agent calls a verification flow, that flow runs in its own context. The agent receives the result (“verified” or “not verified”) but never sees internal state like one-time codes or sensitive account details. You can’t leak what you don’t have access to. **What happens if an AI agent hallucinates that conditions are met?** The tool catches it. Deterministic operations validate their own preconditions against actual system state, not against what the agent claims. If the agent says “customer confirmed” but no confirmation exists in the conversation log, the tool knows and rejects the request. **How do you balance conversational AI with compliance requirements?** By choosing the right mix per operation. Knowledge lookups can be fully agentic. Refunds need a conversational wrapper with deterministic execution inside. Compliance-critical flows might be almost entirely structured. You configure this based on consequences: low stakes get flexibility, high stakes get structure. **What’s the difference between guardrails and pockets of determinism?** Guardrails watch agent outputs and intervene when something looks wrong. By then, the agent has already decided to do something wrong. Pockets of determinism put validation inside the operation, so mistakes can’t execute in the first place. It’s the difference between catching errors and preventing them. --- # AI Support for Fintech: Why Ecommerce-Built Vendors Fall Short URL: https://www.lorikeetcx.ai/blog/built-for-retail-sold-to-finance **Most AI support vendors were built for e-commerce and retrofitted for finance. Here's how to spot the difference before you're three months into a security review.** *How fintech sales objections reveal which AI support vendors were built for regulated industries, and which ones are pretending.* Three months into a security review, the contract is ready, the demo was impressive, but the AI support vendor still can't confirm where customer conversation data is stored, or whether it stays in a jurisdiction that satisfies the company's regulatory obligations. The vendor's response: "We're working on it." This is an architecture problem, and it's the single most common experience fintech CX leaders have when evaluating AI support vendors. Every fintech raises it, and the question is why. ## Three months in a security review Most AI support products were designed for e-commerce. High-volume, low-stakes, FAQ-heavy. They work brilliantly there. "Where's my order?" is a solved problem. But the ticket that matters in fintech is not "where's my order" - it's "why was my transfer blocked," "I need to dispute this charge before my rent bounces," or "my account has been frozen and I can't access my funds." These tickets touch regulated systems. They require audit trails and they carry real financial consequences for the customer. When a vendor gets stuck in your security review for three months, that tells you something about the product's bones. The data residency controls, the SOC 2 processing integrity evidence, the compliance documentation - these things either exist in the architecture or they don't. You cannot retrofit data sovereignty into a product that was built assuming it doesn't matter. The same pattern shows up in integration conversations. Fintech CX leaders consistently report that vendors underestimate the engineering lift required to connect to regulated systems: fraud detection logic that's too nuanced to codify as simple rules, KYC verification workflows, transaction monitoring systems. The vendor's demo looks clean, but the reality involves engineering teams that are already stretched and CS leaders who can't override their security team's veto - nor should they. None of this is unreasonable on the fintech's part, it's the vendor revealing that their product was built for a different customer. ## The latency tax on financial anxiety [Latency matters differently](/blog/latency-in-ai-can-make-or-break-cx) in financial services. A 30-second wait while someone checks their order status is mildly annoying. A 30-second wait while someone is trying to figure out why their card was just declined at a restaurant is something else entirely. Abandonment in financial support correlates with anxiety, not patience. Every second of silence signals to the customer that the system cannot help them. Most AI support vendors publish latency benchmarks from their best-case scenario: a simple FAQ lookup. The relevant benchmark for fintech is latency on a ticket that requires pulling account data, checking transaction history, verifying a dispute against merchant records, and making a decision. Those are very different numbers. [Voice makes this even more acute.](/blog/the-dna-of-voice-ai-that-works) When a customer calls about a frozen account and there's a multi-second delay before each response, it doesn't feel like talking to a support agent. It feels like talking to a system that's struggling. The customer's next move is predictable: "Can I speak to a person?" P50 latency is not the metric that matters here. P90 is. The customers who experience your worst latency are disproportionately the ones in the most stressful situations - a declined mortgage payment, a blocked international transfer, a compromised account - the ones whose loyalty is most at stake. ## Guardrails for financial consequences Fintechs need AI that knows what it cannot do. Monthly compliance testing before go-live, mandatory human escalation for financial hardship and distress, regulatory caution so strict that any agentic action near financial advice triggers an immediate handoff. These are not edge cases in fintech, they are quite literally the operating environment. The absence of [pre-built financial guardrails](/articles/what-are-ai-guardrails-for-customer-service) in most AI support products reveals their heritage. A product designed for a world where the worst outcome is a wrong shipping estimate won't have the machinery to prevent the worst outcome in fintech: giving a customer incorrect information about their money, a wrong answer about a hardship process or an incorrect dispute outcome. [RAG inconsistency](/blog/defence-in-depth-how-to-actually-manage-ai-accuracy-in-production) is a symptom of the same problem. When your knowledge base contains regulatory guidance - and in fintech, it always does - "mostly accurate" is not an acceptable specification. A customer who receives incorrect information about a hardship process, a complaint pathway, or their rights under consumer protection law is not just poorly served. They are potentially harmed and you are potentially liable. The fintech CX leaders who push hardest on guardrails are telling you what the product needs to do. Configurable compliance testing, topic-level escalation rules for transactions, disputes, hardship, and account access - these are the minimum viable product for regulated support. ## Pricing tells you what the vendor believes Fintechs consistently push for usage-based pricing, short pilots, and per-resolution models. Vendors consistently push for 12-month commitments and upfront contracts; both sides think the other is being difficult. A vendor who insists on an annual commitment before you've seen results is pricing around their retention risk, not your value. They know that once the contract is signed, switching costs will keep you locked in regardless of performance. A vendor who charges per resolution is making a different bet: that their product will actually resolve things - transaction disputes, account queries, payment failures, KYC verification support - not just [deflect them to your queue](/blog/why-deflection-focused-products-make-worse-ai-agents). The pricing structure tells you what the vendor believes about their own product. When a CX leader asks for a 60-day pilot with success criteria instead of a 12-month lock-in, that's someone correctly identifying that the vendor's confidence should be legible in the deal structure. Short pilots are not a barrier to adoption, they are a quality filter. ## The "safe" choice When a fintech CX leader defaults to the market leader - the name their board won't question, the vendor with the most logos on their website - they are making a career-risk decision, not a product decision. That's rational, but carries its own risk. The market leader built their product for the broadest possible market. The regulated, high-stakes cases that define fintech support - disputed transactions, blocked transfers, hardship applications, frozen accounts - are not the use cases that shaped their architecture. Their product was optimized for the tickets that are easiest to automate, not the ones that matter most to your customers. The relevant question is not "who is the safest vendor?" It's *safe for whom?* Safe for the buyer's career is not the same as safe for the customer whose hardship case gets hallucinated guidance. Safe for the procurement committee is not the same as safe for the compliance team that has to explain the audit trail, or the lack of one. The vendor with the biggest brand has the least incentive to solve your hardest problems. They have enough easy problems to keep their metrics looking good. ## The spec nobody wrote down Every objection a fintech CX leader raises during an AI support evaluation is a line item in a spec that nobody has written down. Data sovereignty; sub-second latency on real tickets - transaction disputes, account queries, payment failures - not FAQ lookups; pre-built regulatory guardrails for financial hardship, complaints, and advice boundaries; configurable human escalation for distress and vulnerability; pricing that aligns the vendor's incentives with yours; and resolution rates on the hard cases, not just the easy ones. The vendors who meet this spec didn't retrofit it, they started there. They chose [regulated industries](/articles/ai-in-financial-services) deliberately, not because it was the fastest path to market, but because the hardest problems are the most defensible ones to solve. At Lorikeet, this is the product we built. Per-resolution pricing, because we should only get paid when the customer's problem is actually solved; configurable guardrails with auditable performance, because "trust us" is not a compliance strategy; architecture designed for data sovereignty from day one, not bolted on after losing a deal; and a focus on the hardest 20% of tickets - transaction disputes, transfer blocks, hardship processes, account access issues - because that's where the value is for fintech customers. The question for fintech CX leaders evaluating AI support is not whether AI works for regulated industries - it does. The question is whether the vendor you're evaluating was built for the problem you actually have. --- # Buyers Guide to QA in CX: Measure, Diagnose and Act URL: https://www.lorikeetcx.ai/blog/buyers-guide-to-qa-in-cx-measure-diagnose-and-act **Most QA tools measure what changed in your CSAT. Compare sample-based, automated scoring, and diagnostic QA tools that tell you why – and how to fix it.** Most QA tools can tell you what changed. Almost none can tell you why, or how to fix it. Your CSAT dropped 8 points last week. Something is clearly wrong, but there's no obvious culprit. So you start the manual slog: sample 50 tickets, read through conversations, try to spot patterns. Three hours later, you think maybe it's the new refund policy. Or the updated agent training. Or that product bug engineering hasn't prioritised yet. Meanwhile, the undiagnosed problem that tanked your CSAT last week is still hitting customers this week. This is the QA tool gap. The entire category has spent a decade getting better at measurement, building prettier scorecards and slicker dashboards, while the actual problem - figuring out *why* something broke and *how* to fix it - stays manual. ## Your sample size is lying to you Here's a statistical reality that most QA vendors would prefer you not think about too hard. Most support teams sample 2-5% of tickets for quality review. If your real CSAT is 85%, a 5% sample could show you anywhere from 75% to 95% depending on which tickets you randomly pulled. The confidence interval is so wide that you're basically reading tea leaves with a calibrated teacup. But this isn't really a statistical problem. It's an operational one. By the time you've sampled enough tickets to understand why metrics moved, you're investigating last week's issue while this week's issue spreads unchecked. Legacy QA tools were built for a world where human agents worked slowly enough that sampling could catch problems before they became widespread. That world is gone. AI agents can introduce problems at scale instantly. One bad update to your knowledge base and suddenly thousands of customers get wrong answers before your morning standup. Your 2-5% sample won't catch it in time, and when it does catch something, you'll spend another three hours figuring out whether the problem was the knowledge base update, a model behaviour change, or a policy the AI is interpreting differently than you intended. ## The QA tool landscape Let's be specific about what's actually available. This isn't exhaustive, but it covers the major categories and players you'll encounter when evaluating QA tooling today. ### Sample-based QA platforms These make manual sampling workflows more efficient with better scorecards, cleaner interfaces, team performance dashboards, and collaboration features. They are, fundamentally, a nicer way to do the thing that doesn't scale. **Zendesk QA** (formerly Klaus, acquired 2024). Native integration within Zendesk, which is both its strength and its constraint. If you're on Zendesk and want QA that lives inside the same ecosystem, it's the path of least resistance. But it inherits Zendesk's broader limitation: the platform was architected for human agent workflows, and bolting AI-aware QA onto that foundation shows the seams. Sample-based with manual review. No automated root cause diagnosis. No AI-specific failure detection for drift, hallucination, or compliance gaps. **MaestroQA**. Customisable scorecards and team performance tracking. Established player with flexible scoring frameworks and a genuine depth of features for traditional QA programs. The core limitation is structural: manual review doesn't scale to 100% coverage, and the diagnosis step - figuring out *why* scores changed - stays manual. If you have dedicated QA staff and your support operation is primarily human agents, MaestroQA is a solid, well-built tool for what it does. It just doesn't do the part that matters most. **The category verdict:** If you're committed to sample-based QA, these tools make the process less painful. Good scorecards, clear workflows, performance tracking. What they don't do is diagnose root causes automatically, handle 100% coverage, detect AI-specific failures, or implement fixes. You get a better microscope for the 3% of tickets you're already looking at. ### Automated scoring platforms These use AI to score conversations automatically, aiming for higher coverage than manual sampling. It's the right instinct (cover everything, not just a sample) applied incompletely (scoring without diagnosis is still just measurement). **Solidroad**. Positions around 100% automated review and scoring, which addresses the sampling gap directly. The current focus is on scoring rather than diagnosis or fix implementation, which means you'll know *that* quality dropped across all your conversations instead of just a sample, but you'll still be doing the manual work to figure out *why* and *what to do about it*. Emerging player worth watching. **Intercom's CX Score**. This one deserves its own paragraph because it's genuinely clever marketing wrapped around a structural conflict of interest. Intercom built a proprietary AI-driven metric to replace CSAT, and their AI conveniently scores their AI highly. The metric is a black box that can't be benchmarked externally or verified independently. You can't compare your Intercom CX Score against industry benchmarks, against competitors' scores, or even against your own historical CSAT in any meaningful way. It's the equivalent of a restaurant rating itself five stars and then telling you Yelp is outdated. If your QA vendor is also your AI vendor and they've invented a proprietary metric that only they can calculate, you should be asking some pointed questions about incentive alignment. **The category verdict:** Higher coverage than manual sampling, reduced QA headcount requirements, consistent scoring criteria. But most stop at scoring without explaining why scores changed or how to fix problems. Proprietary metrics create vendor lock-in by design. You get a better thermometer, but the thermometer doesn't tell you what's causing the fever. ### Diagnostic and fix-oriented tools This is the emerging category that goes beyond measurement to diagnose why metrics moved and propose or implement fixes. [**Lorikeet Coach**.](https://www.lorikeetcx.ai/product/coach) This is ours, so I'll keep the assessment brief and let the framework do the positioning (more on limitations below). Agent-based QA with 100% coverage, root cause diagnosis, conversational interface via Slack or Claude or ChatGPT, automated fix proposals, and AI-specific failure detection. Works for all companies from start ups to enterprises, for internal teams or those with BPOs and contact centers, and can QA for both human agents and AI agents. [Launched January 2026.](https://www.lorikeetcx.ai/blog/launching-coach) ### The comparison table **Tool** **Zendesk QA, MaestroQA** **Solidroad, Intercom CX Score** **Coach** **Coverage** 2-5% sample 100% scored 100% diagnosed **Root cause diagnosis** Manual Manual Automated with evidence **AI failure detection** Limited Varies Drift, hallucination, compliance, KB contradictions **Fix implementation** Manual Manual Proposed and testable **Works across human + AI** Yes (human focus) Varies Yes (unified standards) **Metric transparency** Industry-standard metrics Some proprietary (CX Score) Industry-standard and custom metrics you can audit and control **Conversational UI** No No Yes in-app and via MCP **Maturity** Established Mixed New to market (Jan 2026) ## What actually matters when evaluating QA tools The features lists on QA vendor websites tend to blur together after the third demo. Here's what actually separates tools that help from tools that just measure, and why the distinction matters operationally. **Coverage that's real, not aspirational.** There's a meaningful difference between "we can score 100% of conversations" and "we diagnose 100% of conversations." Scoring at scale tells you the shape of the problem. Diagnosis at scale tells you the cause. If your vendor says "100% coverage" ask them: coverage of what, exactly? Scoring? Tagging? Or actual root cause analysis? Most AI issues - a knowledge base article that contradicts your refund policy, a model that's started hallucinating shipping timelines, an agent that handles the first question well but fumbles the follow-up - don't surface from scoring alone. They surface from diagnosis, from something that looks at the conversation and works backwards to *why* it went wrong. **Diagnosis speed that matches the speed of the problem.** AI agents can break thousands of conversations in hours. If your QA tool needs a week of accumulated data before it can tell you what went wrong, you've got a monitoring tool, not a diagnostic one. The question to ask: how quickly after a problem starts can this tool tell me what's causing it and what to do? If the answer involves "after your next QA review cycle," that's a cycle designed for human-speed problems applied to machine-speed ones. **Transparency in QA inputs and calculations. **Synthetic quality scores must have auditable inputs and drivers that are in full control of the CX leader. If these scores are set by the vendor without ability to control changes in definition or weighting of the scores, this defeats the purpose of being able to set a goal, track the metric and work to improve it. You have to measure what matters, but if you have no control over the inputs or what the metrics measure, that entirely defeats the point of a performance system.** AI-specific failure detection.** This is where most legacy QA tools fall down completely, and it's not their fault - they were built before AI agents existed at any real scale. But the failure modes are genuinely different. AI agents drift over time as models update. They hallucinate confidently. They find creative interpretations of policies that technically satisfy the letter but miss the spirit. They can contradict your knowledge base in ways that are syntactically different but semantically identical to the correct answer, which makes them almost impossible to catch with keyword-based rules. If your QA tool can't distinguish between "the agent gave a wrong answer" and "the agent gave a wrong answer because the knowledge base article on returns was updated Tuesday and now contradicts the policy document from March," you're still doing the diagnosis manually. **The gap between knowing and fixing.** Most QA tools stop at measurement. The good ones stop at diagnosis. Very few close the loop to fix implementation. Ask your vendor: when your tool identifies a problem, what happens next? If the answer is "we surface it in a dashboard and your team investigates," that's a reporting tool. If the answer is "we diagnose the root cause, propose a fix, let you test it, and implement it," that's a QA tool. ## Questions to ask every QA vendor These are slightly adversarial on purpose. The QA category has spent years selling dashboards as solutions, and you deserve direct answers. - **What percentage of conversations do you actually analyse?** Not score. Not tag. Analyse for root causes. If it's less than 100%, how do you handle the conversations you miss? - **When my CSAT drops 8 points, how quickly can your tool tell me why?** Not *that* it dropped. *Why.* With evidence. - **How do you detect AI-specific failures?** Drift, hallucination, compliance gaps, knowledge base contradictions. If the answer is "the same way we detect human agent issues," that's not good enough. - **Is your scoring metric proprietary or industry-standard?** Can I benchmark it externally? Can I take my data to another vendor and get comparable scores? - **What happens after you identify a problem?** Dashboard alert? Suggested fix? Tested fix? Implemented fix? The further right on that spectrum, the more useful the tool actually is. - **If your company also sells the AI agent, how do you handle the conflict of interest in grading its performance?** This one's for Intercom specifically, but the principle applies anywhere the QA vendor and the agent vendor are the same company. ## Where Coach fits (and where it doesn't) I've kept the framework above vendor-neutral because I think it's genuinely useful regardless of what you buy. But we built Coach specifically because we saw this gap with our own customers, so here's the honest assessment. **Coach is limited in ways you should know about.** It launched in January 2026, which means it's new to market. While Coach works across any support operation (human and AI operations), it's strongest when paired with Lorikeet's AI agent because the diagnosis layer has deeper access to the agent's reasoning. **Coach is not a fit if:** you're happy with sampling-based QA and your current process works, you don't have AI-specific failure modes to worry about, or your compliance requirements don't mandate complete conversation monitoring. ## The bottom line The QA tool market is splitting into two philosophies. Measurement-focused tools give you better scoring, higher coverage, prettier dashboards. They tell you *what's* happening with increasing precision. Action-focused tools give you diagnosis, root cause analysis, fix implementation. They tell you *why* things are happening and help you fix them. Sample-based monitoring made sense when humans were the constraint. They're not anymore. When you're evaluating QA tools, demand complete coverage, diagnosis with evidence, fix proposals you can test and implement, AI failure detection that's purpose-built rather than retrofitted, and consistent measurement that you can benchmark independently. Your customers don't care about your internal QA methodology. They care whether you fix problems fast. ## FAQ **Do I really need 100% conversation coverage?** If you're running AI agents at any meaningful scale, this is feasible and should be the bar. The math is straightforward: AI agents can introduce systematic errors across thousands of conversations simultaneously. A 3% sample has roughly the same chance of catching a systematic AI failure in its first hour as you have of guessing which specific customer will complain on Twitter. By the time your sample catches it, the damage is already done. For regulated industries (fintech, insurance, healthcare), the argument is even simpler: your compliance team will eventually ask "how many conversations did you *not* review?" and you need a better answer than "ninety-seven percent of them." **How does Intercom's CX Score compare to industry-standard QA metrics?** It doesn't compare, and that's the point. CX Score is a proprietary metric designed to replace CSAT, not complement it. You can't benchmark it against industry data, you can't compare it across vendors, and you can't independently verify how it's calculated. Intercom argues this is a feature because CSAT is flawed (which is partially true). But replacing a flawed open standard with a proprietary black box from the same company that sells you the AI agent being measured isn't a solution to the measurement problem. It's a solution to Intercom's competitive positioning problem. **What if my team is all human agents with no AI?** Sample-based QA tools like MaestroQA or Zendesk QA are genuinely good for this use case. The diagnosis gap matters less when problems emerge at human speed rather than machine speed. If you're planning to add AI agents in the next 12 months, factor diagnostic capabilities into your evaluation now so you don't have to rip and replace later, but if AI isn't on the roadmap, the traditional tools work. **What's the minimum setup investment for diagnostic QA?** Honest answer: more than sample-based QA, less than building your own. Coach typically takes 1-2 weeks to configure properly, including defining what "good" looks like for your specific operation, connecting your data sources, and calibrating the diagnosis layer. Sample-based tools can be running in days. The tradeoff is setup time versus ongoing investigation time, and most teams doing 10+ hours per week of manual root cause analysis recoup the setup investment within the first month. **How do diagnostic tools handle false positives?** This is a legitimate concern and one we think about constantly. Coach uses a confidence-scored approach where diagnoses come with supporting evidence (specific conversations, patterns, and statistical backing) rather than binary alerts. The goal is to surface probable root causes ranked by evidence strength rather than a firehose of "something might be wrong" notifications. That said, any system doing automated diagnosis will occasionally identify patterns that aren't real problems. The mitigation is transparency: showing you the evidence and letting you decide, rather than hiding the reasoning behind a score. **Which industries benefit most from diagnostic QA?** Any industry where getting the answer wrong has consequences beyond customer frustration. Fintech (wrong information about fees, accounts, or transactions), insurance (incorrect coverage guidance), healthcare (inaccurate medical information), and regulated e-commerce (compliance with consumer protection laws) all have failure modes where "we sample 3% and hope for the best" isn't a defensible position. That said, even unregulated businesses with high conversation volumes benefit from the operational speed: finding and fixing problems in hours rather than weeks helps regardless of your compliance obligations. **Should I use a dashboard or a conversational interface for QA?** Dashboards are good for scheduled reviews and trend monitoring. Conversational interfaces (asking your QA tool "why did CSAT drop last Tuesday?" and getting an answer with evidence) are better for ad-hoc diagnosis and faster time-to-understanding. The best setup is both: a dashboard for the weekly review cadence, and a conversational interface for the "something broke and I need to know what" moments. If you have to pick one, pick the one that matches how your team actually works. If your QA lead checks a dashboard every Monday morning and that process works, a conversational interface is nice-to-have. If your team spends hours each week manually investigating metric movements, the conversational interface will save significantly more time. --- # What Happens When CX and Ops Leaders Get Hands-On with Claude Code URL: https://www.lorikeetcx.ai/blog/what-happens-when-cx-and-ops-leaders-get-hands-on-with-claude-code **We hosted a live session where CX and ops leaders installed Claude Code, analyzed real support data, and built working automations. Here's exactly what they did.** This week, we hosted a live session built for CX and ops leaders who are curious about Claude Code but want to get past the hype and into the practical reality of what it can actually do for their teams. No polished keynotes. No hand-wavy promises. Just Robbie Tilleard (Lorikeet GM, EMEA) and Nate Houghton (Lorikeet Growth, Americas) rolling up their sleeves in a live terminal, with a live audience watching over their shoulders. Here is what we covered, and what you need to know. ## What is Claude Code? Claude Code is a command line tool from Anthropic. You type instructions in plain English and it executes them. That is genuinely the whole model. What makes it different from the Claude web app is that it runs directly on your machine, it reads your files, it connects to your tools, and it can chain dozens of steps together without getting tired and giving up halfway through. If you have ever started a complex task in ChatGPT or Claude and watched the thread drift off course by step five, you will recognize exactly the problem Claude Code solves. It also connects to external tools through something called MCPs (Model Context Protocols), which create fast, semantic connections to your data without needing a full API setup. That means your calendar, your CRM, your support ticketing system, your Google Drive, your Slack -- all of it can be wired in so that Claude Code is not working in isolation. It is working with the full context of your actual job. ## Getting Set Up in Under Five Minutes Robbie walked attendees through installation live on screen. The short version: - Open your terminal (Command + Space on Mac, search "terminal") - Run the install command (curl -fL https://claude.ai/install.sh | bash) - Create a project folder -- your dedicated workspace - Open Claude Code from that folder by typing `claude` The one file that matters most in that folder is called `claude.md`. Think of it as your AI's instruction manual. It persists across every session and tells Claude who you are, what you do, what tools you use, who your key contacts are, and how you want it to work. Unlike a chat window that forgets you every time, this memory stays put. If the terminal feels unfamiliar, Robbie had a tip that actually works: run Claude Code on one screen and keep the Claude web app open on the other. When Claude Code gives you something confusing, paste it into the web app and ask what it means. Two screens, one brain. ## The Live Demo: 100 Support Tickets, Analyzed in Minutes To show what this looks like in practice, Robbie dropped 100 realistic neobank support tickets into Claude Code and asked it to get to work. The tasks: - Cluster the tickets by topic and show the distribution - Run a trend analysis: which topics are spiking and when - Identify which day had the highest ticket volume - Produce an executive-ready report framed for a product and technology team Claude Code found that card issues triggered a secondary complaint wave after an outage. Fee transparency concerns were climbing all week. Account closures started appearing mid-week, with customers explicitly naming a competitor in their reasons for leaving. The final output was a structured report that read: "The card outage on February 7th was a trust-breaking event. Customers found out from Twitter and Reddit before the company told them. The outage had downstream effects for days." That kind of clear, story-driven insight is exactly what you need to walk into a product review meeting and get something on the roadmap. ## Other Uses for Claude Code? Remote Management, Rewritten Nate made an observation during the session that stopped people in the chat: if you have experience managing a remote team, you are probably already good at this. The parallel holds more precisely than it first appears. Managing a distributed team well means writing instructions that leave no room for misinterpretation, anticipating the questions someone will ask before they ask them, and knowing when output is good enough to ship versus when something has gone quietly wrong. Those are exactly the three skills that separate productive Claude Code users from frustrated ones. The tool is powerful in proportion to your ability to brief it well. CX and ops leaders have spent careers doing this. They have managed offshore teams, written SOPs that had to survive without them in the room, and diagnosed process failures from the output side rather than the input side. The coding agents did not change the game for them. The game was already theirs. ## Key Takeaways **Skills are the unlock.** Any repeatable task can be turned into a skill -- a saved function you can call at any time with a single phrase. Robbie's team has built skills for daily planning, inbox zero sweeps, meeting prep, discovery call summaries, stakeholder updates, deal evaluation, and LinkedIn thought leadership drafts. Skills are written in English, not code, and they can be shared with your whole team as simple text files. **The judgment layer is non-negotiable.** Robbie introduced a framework worth holding onto. A person with AI and strong judgment is a "turbo brain." A person with AI and no judgment is a "slop cannon" -- fast output, low quality, no sense of when the model has gone sideways. The tool amplifies whatever you bring to it. If you are not actively reviewing what it produces, you are not actually using it well. **Parallel processing changes the pace of work.** Both Robbie and Nate described running multiple Claude Code sessions simultaneously. While one agent runs an analysis, another is drafting a document, and another is pulling data from a third tool. The ability to queue up work across several threads in the background is not a minor feature. It is a fundamentally different relationship with what a single workday can produce. **Managing AI is a lot like managing a remote team.** Nate put it plainly: if you have ever led a distributed team and learned to write instructions that leave no room for misinterpretation, you will be good at this. The quality of your prompts determines the quality of your output. More context, less ambiguity, better results. And unlike a human employee, if Claude runs in the wrong direction, you can reorient it instantly with no sunk cost. **The mistakes are predictably dumb, and that is actually useful.** Nate noted that Claude Code can synthesize complex information across multiple sources and produce something genuinely impressive, and then fail to get a brand color right three times in a row. The types of errors are learnable and manageable once you know to watch for them. That predictability makes it a trustworthy tool, not a liability. ## The Lorikeet MCP: For Current Lorikeet Customers One of the most-asked questions from the audience was about the Lorikeet MCP -- the integration that connects Claude Code directly to a Lorikeet instance. Robbie described the workflow in four stages: diagnose, prescribe, simulate, and push to production. If CSAT has dropped, you ask Claude to identify why. It looks at your workflows, your tickets, and your performance data. It then proposes specific changes. Before anything goes live, you run simulations -- hundreds of variations to make sure the behavior is consistent. Then you push. A critical note: some actions, like turning guardrails on or off, are intentionally designed to require a human step. Claude Code will tell you when that is the case. That is a feature, not a gap. ## Toolshed: A Free Resource for CX Leaders Tom from the Lorikeet marketing team joined the session to announce Toolshed, a set of eight free tools for CX and AI decision-makers, available at [lorikeet.tools](https://lorikeet.tools/?utm_source=content&utm_medium=blog&utm_campaign=claude-code-mar10-webinar). The tools include: - An AI readiness score across six dimensions - A build versus buy comparison with three-year total cost of ownership modeling - A backlog cost calculator that translates your ticket queue into dollars: churn risk, SLA penalties, lost revenue - A quality versus speed benchmarker to plot where you sit on the AI efficiency frontier No email gate. No paywall. All inputs stay in your browser. Built entirely in Claude Code. ## How to Get Started This Week You do not need to be an engineer. You do not need a big budget. Here is where Robbie and Nate landed: - Install Claude Code and create your workspace folder - Write your claude.md -- who you are, what you do, what tools you use - Connect one tool -- start with your calendar or a folder of files you already work with - Give it one real task - Turn the output into a skill - Share that skill with your team The teams that build these habits now will compound their advantage as AI moves faster. That was the closing message, and it holds: the gap between organizations that started experimenting six months ago and those that start today is real and growing. But it is still early enough that starting this week puts you ahead of most. *Ready to take a look at the free CX tools? Start here *[*lorikeet.tools*](https://lorikeet.tools/?utm_source=content&utm_medium=blog&utm_campaign=claude-code-mar10-webinar)*.* --- # AI Agents for Fintech Customer Service: Achieve First Call Resolution URL: https://www.lorikeetcx.ai/blog/ai-agents-for-fintech-customer-service-first-call-resolution **FAQ bots can't check a stuck refund or freeze a compromised card. See why fintech AI support needs to resolve problems, not just talk about them.** Most AI support vendors in fintech are building really sophisticated FAQ summarizers. They can tell your customer that "refunds typically take 3-5 business days" in seventeen different ways. What they can't do is actually check the specific refund, see that it's stuck in a compliance review, and resolve it. Financial services customers don't contact support to have a conversation. They contact support because something is wrong with their money, and they need it fixed. The gap between "AI that sounds helpful" and "AI that actually helps" is enormous, and it's costing fintech companies customer trust every day. At Lorikeet, we've spent years working with financial services companies from corporate expense platforms, [B2C lenders](https://www.lorikeetcx.ai/customer-stories/lorikeet-helps-flex-scale-support-with-ai-agents-that-match-urgency-with-accuracy), remittances, [cryptocurrency tax providers](https://www.lorikeetcx.ai/customer-stories/how-summ-efficiently-handles-support-spikes-during-tax-time), and global payment providers like Airwallex. What we've learned is that the value in AI support doesn't come from deflection rates or automation percentages. It comes from resolution. ## The "Where's My Money?" Problem The most urgent question in fintech support isn't really a question at all. When a customer asks "where's my money?", they don't want an explanation of your payment processing timeline. They want their problem solved. Consider what actually happens when a corporate card transaction is declined: **The FAQ bot approach:** "Card declines can occur for several reasons including insufficient funds, merchant category restrictions, or spending limits. Please check your card settings in the app or contact your administrator." **The agent approach:** Check the specific transaction. See that it triggered a fraud flag because the user is traveling. Verify the user's identity. Whitelist the merchant. Confirm the user can retry the purchase. Send a follow-up if the retry fails. What’s important to the customer is not the conversation but whether the agent can understand and solve their problem. We work with a corporate expense management platform handling thousands of transaction queries monthly. Their support challenges span declined transactions, card delivery issues, and fraud investigations. These aren't FAQ topics. They're investigations that require looking up specific transactions, checking card statuses, verifying integration states, and taking action. ## Financial Workflows That Actually Matter Commonly AI vendors will talk about handling tickets related to opening accounts, changing information, and "personalized product advice." These are real categories, but they undersell what AI agents should actually be capable of in financial services. ### Account Management Not just "change my address", but handling the compliance implications. When a corporate expense user's employment status changes, you need to understand their role, adjust their card limits, notify their administrator, and potentially freeze purchasing while verification completes. A single account update can trigger a cascade of actions. ### Transaction Support Beyond checking processing status, actually resolving issues associated with the transaction. When a crypto deposit doesn't appear in a user's account, you need to check the blockchain transaction, verify the wallet address, confirm it's past the required confirmations, check if it triggered a compliance review, and either credit the account or explain exactly what's holding it up with specific next steps. ### Integration Issues Fintech platforms don't exist in isolation. At one of our customers, a user can ask "my transactions show reconciled in my HR system but unreconciled in my expense platform," the AI needs to understand both systems, check the sync status, potentially trigger a refresh, and explain what happened, not just escalate to a human who has to do the same investigation from scratch. ### Fraud and Security This is where the stakes are highest. When a user reports unauthorized transactions, you need to freeze the card immediately, gather information about the specific transactions in question, initiate a dispute process, order a replacement card, and potentially coordinate with multiple parties, all while the customer is stressed and worried about their money. ## Reasoning Over Retrieval Most AI support systems are essentially retrieval engines with a language model on top. They find relevant help articles and reword them conversationally. This works for "how do I reset my password" but falls apart for "my card was charged twice for the same purchase." Financial services queries almost always require reasoning: - Is this actually a duplicate charge, or did the customer make two purchases? - If it's a duplicate, is the second charge from the same merchant or a similar-named one? - If it's a true duplicate, what's the refund timeline for this specific merchant category? - Does the customer have a pending refund already in progress? Each answer depends on the previous one. You can't template this. You need an AI that can investigate, form hypotheses, gather evidence, and take appropriate action. ## Multi-Agent Coordination Complex fintech issues often require coordinating multiple parties simultaneously. Consider a scenario where a customer's international wire transfer failed partway through. The customer's bank debited the funds, but the receiving bank rejected the transfer for a compliance reason. The money is now in limbo. A single-threaded chatbot can maybe explain this. Lorikeet's [team of agents](https://www.lorikeetcx.ai/blog/team-of-agents) can: - Contact the sending bank to confirm the debit and request a trace - Check with the compliance team on the rejection reason - Verify the receiving account details with the customer - Initiate a recall request if needed - Monitor the recall process and update the customer This isn't hypothetical. This is the kind of coordination that human support teams do every day in financial services and it's exactly what AI agents should be capable of. ## Why Resolution is our North Star for resolution, not Deflection Vendors in AI for CX often tout "deflection rate" but this creates unaligned incentives. A vendor gets paid more when their AI attempts to answer questions it shouldn't. A 70% deflection rate where 30% of customers end up frustrated and calling anyway is worse than a 40% rate where every AI-handled conversation actually resolves the issue. [**This is also why Lorikeet charges per resolution, not per deflection or per conversation.**](https://www.lorikeetcx.ai/pricing) We only get paid when the AI actually solves the customer's problem, verified by the ticket staying closed and the customer not coming back with the same issue. Our incentives are fully aligned with yours: we succeed when your customers succeed. We've seen the alternative play out specifically in fintech. A payment app we talked to had achieved "85% deflection" with their previous vendor. When we looked closer, 40% of those "deflected" tickets were reopened within 48 hours customers who gave up on the AI, thought their issue was resolved, then came back when they realized it wasn't. The right question isn't "what percentage did AI handle?" It's "what percentage did AI handle well?" And in financial services, "well" means the customer's actual problem is resolved, not that they stopped talking to the chatbot. You can see the RoI for resolution-based and incentive aligned vendors compared to interaction based vendors easily in our[ RoI calculator](https://lorikeet.tools/cx-roi?tickets=10000&solvable=50&ppt=0.9&lcpr=0.75&csat_res=75&csat_fail=30). ## What This Means for Fintech Leaders If you're evaluating AI support for a financial services business, ask these questions: - **Can it take action?** Not just explain policies, but actually resolve issues by interacting with your systems. - **How does it handle investigation?** Financial queries often require multi-step reasoning. Ask for examples of how the AI debugs complex issues. - **What happens when it's uncertain?** The best AI knows when to escalate. The worst AI confidently tells customers wrong information about their money. - **How do you measure success?** If the answer is just "deflection rate" or "automation percentage," be skeptical. Ask about resolution rates, CSAT on AI-handled tickets, and reopen rates. - **Can it coordinate complexity?** Modern financial services often require working with multiple systems, parties, or workflows simultaneously. The bar for AI in fintech is higher than in other industries because the stakes are higher. Your customers aren't asking about product recommendations, they're asking about their money. The companies that recognize this and build AI support accordingly will have a genuine competitive advantage. The FAQ bot era served a purpose, but it's not enough anymore. Financial services customers deserve AI that can actually help them, not just talk about helping them. --- # Launching: Toolshed URL: https://www.lorikeetcx.ai/blog/launching-toolshed **8 free calculators and assessments for CX leaders evaluating AI support. Build vs buy, knowledge base readiness, backlog cost, ROI – no email required.** Sierra launched a university. Decagon launched a university. When the two most-funded AI support companies in the world both decide that what CX leaders really need is to sit in a classroom, it tells you something about how this industry sees its buyers. The assumption is that you - the person running support for a company doing real volume - don't know enough to make good decisions about AI. That if someone just explained it to you, slowly, maybe with a video series and a certification badge, you'd finally get it. We think that's backwards. ### Tools, not lectures CX leaders don't have a knowledge problem. They have a decision problem. The VP of Support at a Series C fintech doesn't need someone to explain what an LLM is. She needs to know whether her knowledge base is actually good enough to power an AI agent, what her ticket backlog is costing her while she evaluates vendors, and whether the "build it ourselves" pitch from her engineering team holds up against a three-year TCO model. Those aren't questions you answer with a blog post or a webinar. They're questions you answer with a calculator, an assessment, a model you can run with your own numbers and show your CFO on Monday. ### The gatekeeping problem Here's what the "university" model actually does: it creates a captive audience. You fly to San Francisco. You sit in a room with a vendor's team for two days. You leave feeling educated and slightly obligated. The vendor leaves with your contact details, your company's pain points laid bare, and a warm lead they can work for six months. It's not education. It's a very expensive top-of-funnel play dressed up as thought leadership. The tell is who gets access. These programs aren't open to every CX leader trying to figure out AI. They're open to the ones who can afford a seven-figure contract - or who the vendor thinks might be able to. If your company isn't big enough, you don't get to learn. What free actually means When we started building [Toolshed](https://lorikeet.tools/), the first decision was the simplest one: no gates. No email capture before you can use a tool. No request access form. Every tool works immediately, with your real numbers, and gives you a concrete output - a cost figure, a readiness score, a hiring plan - that you can screenshot and put in a deck. If you never talk to us, that's fine. You still got something useful. This isn't altruism. It's a bet on a specific idea: that the best way to build trust with a CX leader is to be genuinely useful to them before you ask for anything in return. The worst way is to make them feel like they need your permission to understand their own operation. ### The eight tools (with more coming) We started with the questions CX leaders actually ask when they're evaluating AI - not the questions vendors wish they'd ask. The [Build vs Buy Calculator](https://lorikeet.tools/build-vs-buy) exists because every CX leader has an engineering team that says "we could build this ourselves." They're usually right that they could start. They're usually wrong about what it costs to maintain. The calculator models the full three-year cost of both paths, including the rebuild cycles that catch internal teams off guard when models and architectures shift every six months. The [Knowledge Base Evaluator](https://lorikeet.tools/kb-evaluator) exists because the number one predictor of AI agent quality is the quality of the knowledge it's trained on. Most companies don't know whether their help centre is ready for AI until they've already deployed and seen it hallucinate answers from three-year-old articles. The evaluator scores coverage, freshness, and AI-readiness before you've committed to anything. The [AI Readiness Scorecard](https://lorikeet.tools/ai-readiness) measures six dimensions of organizational readiness - because the technology being ready doesn't mean your team, your processes, or your data are. The [Quality vs Speed Benchmarker](https://lorikeet.tools/quality-vs-speed) visualizes the tradeoff that every support leader feels in their gut but can't articulate in a board meeting: what happens to quality when you push for faster response times, and where AI actually shifts that frontier rather than just sliding along it. The [Hiring Forecaster](https://lorikeet.tools/hiring-timeline) and [Agent Turnover Calculator](https://lorikeet.tools/attrition-impact) both address the human side. Hiring in CX takes months. Attrition is brutal - and most leaders underestimate what it really costs when experienced agents leave and new ones take 90 days to ramp. These tools model the compounding effects that spreadsheets tend to flatten. The [Backlog Cost Calculator](https://lorikeet.tools/backlog-cost?industry=other&backlog=500&wait=5&volume=10000&clv=10000&churn=medium&sla=200&cost=15&esc=10) puts a dollar figure on the thing every support leader knows but struggles to quantify for finance: the revenue leaking out of every unanswered ticket sitting in the queue. Lost customers, reduced expansion, overtime spend. The [CX ROI Calculator ](https://lorikeet.tools/cx-roi?tickets=10000&solvable=50&ppt=0.9&lcpr=0.75&csat_res=75&csat_fail=30)is the most direct one. It compares what you're paying your current vendor - or what you'd pay under a pay-per-ticket model - against what a pay-per-resolution model looks like across CSAT, coverage, and total cost. It's the question every CX leader should be asking their vendor and most aren't. ### Why vendors don't do this Building [free tools](https://lorikeet.tools/) that help prospects make independent decisions is a terrible idea if your product doesn't hold up to scrutiny. If your AI agent relies on deflection rates to look good, the last thing you want is a calculator that separates good deflections from bad ones. If your pricing model depends on charging for every ticket touched rather than every ticket resolved, you definitely don't want a side-by-side cost comparison sitting on the internet. The reason most vendors prefer "education" over "tools" is that education lets you control the narrative. A university syllabus is written by the vendor. A calculator with the buyer's own numbers is written by reality. ### Questions to ask yourself - When your AI vendor offers to educate you, what are they actually optimizing for? - Could you explain to your CFO today what your ticket backlog costs the business? - How much of your current AI evaluation is based on your own data versus the vendor's hand-picked case studies? --- # Launching Resolution Loop URL: https://www.lorikeetcx.ai/blog/launching-resolution-loop **Most AI platforms break during escalation. Learn how keeping humans inside the AI platform preserves context, teaches the AI, and reduces escalation rates.** Most AI support platforms treat escalation as an all-or-nothing handoff. When AI can't handle a conversation, it punts to a human, causing delays, breaking the conversational thread all while the customer waits. And even after the human resolves it, the AI learns nothing. The same issues escalate again and again. We built Resolution Loop to fix this. Humans can be in the loop by either taking over the wheel and addressing the customer, or steer the AI agent to the correct answer. A human expert can action something in the background while the AI conversation continues uninterrupted. There's no formal handoff, no disruption for the customer. The human expert can then seamlessly submit input to bridge any [knowledge base](/blog/how-to-deploy-ai-support-without-redesigning-your-entire-operation) gaps so future escalations reduce. ## What makes Resolution Loop different Escalations from AI to humans are often not about fundamental AI inability. In many cases the AI is missing context, workflow guidance or instruction in the knowledge base. Resolution Loop lets the AI isolate exactly what it's confused about and put its hand up for help to a human expert. Resolution Loop offers two modes for human and AI collaboration **Steer:** The AI surfaces a question to a human expert in the background: "How should I calculate this refund?" The human sees a pithy summary with full context and answers without jumping into the conversation. The AI continues seamlessly; the customer never notices. And as a bonus, that answer gets saved permanently so the AI handles it autonomously next time. **Take Over:** For conversations where a human needs to drive. The agent claims the ticket, sees full context, responds directly through the same channel. The customer stays in the same WhatsApp or SMS thread - no channel switch, no broken continuity. Both modes work inside Lorikeet. No routing to external systems, no context reconstruction, no lost history. ## Here's an example of how it works A customer messages about a refund for a partial return. The AI gathers the details, checks the policy, and realizes this is an edge case as the customer used store credit for part of the purchase, and the knowledge base doesn't cover how to calculate the refund. **Traditional escalation:** AI creates a ticket in your [ticketing system](/blog/in-support-copilots-are-not-the-answer). Customer waits. Human agent picks it up, reads the transcript, asks questions the AI already answered, eventually figures out the refund calculation. AI learns nothing. Next week, same edge case, same escalation. **Resolution Loop:** The AI privately asks a human expert inside Lorikeet "How should I calculate a refund when store credit was used?" In Steer mode, the human answers. The AI continues the conversation with the customer, applies the guidance, resolves the issue. The customer never knew anything changed as they experience the same WhatsApp thread and no delays. This matters because that answer gets saved to Lorikeet [Coach](/product/coach). Next time a similar edge case comes up, the AI can handle it autonomously with its newly improved knowledge base. Human experts are able to focus on other complex tickets as the AI grows more capable over time. AI to human hand-offs, when managed well, offer improved customer experience where Resolution Loop offers both a safety valve and a training mechanism. ## The learning loop AI to human escalations happen for a finite set of reasons including edge cases the AI wasn't trained on, policies that changed and scenarios the workflows don't cover. Every time a human provides a Steer instruction, that knowledge gets captured. Every Take Over resolution informs what Coach should handle next time. The [escalation rate](/blog/great-cx-isnt-built-on-deflection-rates) naturally declines because you're eliminating the knowledge gaps that cause them, not because you're suppressing escalations. Our **north** star is: does the escalation rate go down over time because the AI is learning from human resolutions? This is fundamentally different from "[self-training AI](/blog/the-false-promise-of-self-training-ai)" where models learn from their own outputs. Here, the AI learns from verified human expertise the same way you'd train a new team member. ## Availability Resolution Loop is available now for SMS and WhatsApp channels, working alongside Lorikeet Coach to turn every escalation into a learning opportunity. If your escalations feel like a revolving door with the same issues, same volume, month after month with no improvement, we'd love to show you how to turn this around. [Click for a custom demo](/get-a-demo) → --- # How to get started using Claude Code as an operations or customer experience leader URL: https://www.lorikeetcx.ai/blog/how-to-get-started-using-claude-code-as-an-operations-or-customer-experience-leader **Get started using Claude Code as an operations or customer experience leader by starting simple, providing context and applying it to real operational work** I’ve had a lot of conversations recently on how [Lorikeet’s Coach](/product/coach) and tools like Claude Code have raised the bar for what customer experience and operations leaders can achieve in their day-to-day work. What’s obvious is that agentic operations is the real deal. It's clear a gap will emerge between leaders able to orchestrate their agents to do productive work across operations and those too busy to get started. The good news is we're barely at the 'end of the beginning' with 99% of organisations and people still setting off on the journey. So where should you as a customer or operations leader begin? ## Start simple Install Claude Code. Create a single project folder. That’s it. Just a folder with text files and clear instructions. It works through written commands, not traditional code. Give it access to one folder and one tool. You decide exactly what it can see and what it can touch. Most people begin with email or calendar, then expand from there as trust builds. You remain the human in the loop. If you’re struggling with those first steps - have a look at this quick guide to [get Claude Code installed](https://claudecode101.io/). If you’re using Lorikeet, it’s as simple as opening up our web app [and using Coach](/product/coach). ## Provide context These systems are powerful, but they don’t know your world unless you tell them. Create a plain text file that explains who you are, what you’re responsible for, and how you like to work. Think of it as onboarding a new team member. You could include: - About me: your role, team, what you care about - My tools: Zendesk, Salesforce, Slack, etc. - Key contacts: names, roles, Slack handles, emails - Preferences: how you like data, comms style - Recurring tasks: weekly report, Monday triage - Gotchas: knowledge only you know ## Apply it to real operational work This is about practical application. Used well, it can help with: - Understanding operational changes before they become problems - Analysing support tickets and identifying patterns worth acting on - Running simulations for customer experience scenarios - Producing daily trend analysis and organisational reporting Below you can see an example created in minutes based on a sample of 100 customer support tickets and the question “How did sentiment change over the week?” Claude Code skills allow you to quickly automate repetitive tasks and spread automations through the organisation. They are text files with specific repeatable steps that are run through every time a skill is run. Of course, the best way to use these is to ask Claude Code to help you build an effective skill. Below you can see a small sample of the skills we’ve built internally in just the last 30 days. All of these skills are stored in a central repository that anyone at Lorikeet is able to access. ## Start on Monday Fewer than 1% of the world are using these tools in a meaningful way today. That won’t last. Over time, using a coding agent in your day-to-day work will feel as ordinary as word processors or spreadsheets. The real decision isn’t whether they’ll matter, but when you choose to become fluent. For customer experience and operations leaders especially, the gap is already widening between those who are building this into their week and those who aren’t. So, start on Monday. Not because it’s the most urgent thing on your plate (we know you likely have a few urgent fires you’re dealing with!) but because a year from now you’ll be far ahead if you do. ## Resources - [Presentation on getting started with Claude Code](https://claude-code-101.replit.app/) as an operations or customer experience leader - [Practical walk through for setting up Claude Code](https://claudecode101.io/) as an operations or customer experience leader --- # If you don't give your customers AI, they'll get it elsewhere URL: https://www.lorikeetcx.ai/blog/disintermediation **ChatGPT is inserting itself between businesses and their customers. The companies that think they're safe — because they deployed a chatbot, or became a ChatGPT partner — are the ones most at risk. Here's why, and what the winning response actually looks like.** ## On February 9th, shares in Willis Towers Watson fell 11.45%. Aon dropped 9.9%. Arthur J. Gallagher, 9.3%. Not bad earnings. Not a lawsuit. Not a scandal. OpenAI had approved an insurance app. The app was called Tuio. Users could get a personalised home insurance quote inside ChatGPT without ever visiting an insurer's website. The market understood immediately what this meant: if consumers can get insurance through a conversation with a generalist AI, they might not need a broker at all. This is what disintermediation looks like in practice. It wasn't a gradual years-long erosion, just a regular Tuesday. At Insurtech Live Australia in February, Lorikeet CEO Steve Hind made a point that ended up being cited by multiple other speakers across multiple panels that same day. The CTO of Blue Zebra Insurance built a separate argument off it. The Open Insurance panel referenced it. The moderator of a later session called back to "the idea from this morning," without attribution. The idea: companies in regulated industries that don't provide a high-quality, first-party AI concierge will be permanently disintermediated from their customer relationships by generalist AI. The reason it resonated wasn't novelty, it was inevitability. Consumer adoption of tools like ChatGPT has already reset what people expect from digital services. They've experienced what it's like to get an instant, thoughtful, conversational answer to a complex question. Now they wonder why their bank makes them navigate a phone tree. The baseline has shifted, and there's no shifting it back. ## The failure mode that doesn't get talked about enough is the regulatory trap. Executives at regulated companies – insurers, banks, healthcare providers – look at AI deployment and say: "we'd love to, but we can't give advice." They're worried about regulations around financial advice, clinical negligence, insurance licensing. So they hold back. They wait for the rules to settle. They treat caution as a form of customer protection. Meanwhile, their customers aren't waiting. They're asking ChatGPT. In October 2025, OpenAI updated its usage policies to explicitly prohibit ChatGPT from providing "personalized professional advice" in medical, legal, and financial domains. This was widely reported as OpenAI tightening up its safety posture. It wasn't. Multiple analyses noted that ChatGPT continues to provide substantial assistance in these areas when prompted – the policy is a legal disclaimer, not a technical block. Bloomberg Law was direct about it: the ToS update functions primarily as a "liability shield" to make it harder for users to pursue claims when the advice is wrong. So here's the actual situation. You're holding back AI deployment to protect your customers from bad AI advice. OpenAI is offering that same advice, disclaimer attached, with no liability and no accountability when it's wrong – while simultaneously recruiting your most recognisable competitors as data partners inside ChatGPT Health. If you don't deploy AI because you're worried about financial advice regulations, your customers will just get fully unregulated financial advice from ChatGPT. What you thought was protection is actually surrender. ## Weight Watchers is the case study worth telling in full. Before filing for Chapter 11 bankruptcy in May 2025, Weight Watchers had already deployed a customer-facing AI agent – built by Sierra, the Bret Taylor startup – that handled roughly 70% of all customer service sessions without human involvement. Customer satisfaction was 4.6 out of 5. Members were reportedly exchanging heart emojis with the AI. By any deflection metric, it was working. But it wasn't doing what members actually needed. It was handling inbound contacts about existing problems. It wasn't proactively managing the weight loss journey. It wasn't connected to health data. It wasn't taking actions that changed anything for the customer. It was, in the language that most support AI is sold in, very good at resolving queries. Then ChatGPT Health launched, with Weight Watchers named as one of the early partners. The pitch: WW would provide personalised GLP-1 meal planning through ChatGPT conversations. The company got distribution. OpenAI got the customer relationship. Weight Watchers had deployed AI and still ended up disintermediated. Because there's a difference between having AI in support and having AI that gives customers a reason to stay in your ecosystem. A deflection chatbot doesn't create the latter. Only an AI that actually does things – that accesses your systems, takes actions on behalf of customers, produces outcomes they couldn't get elsewhere – gives them a reason to keep coming back to you rather than to whichever generalist platform sits in front of them. Jamie Hall, Lorikeet's co-founder, spent years at Google Brain working on LaMDA – conversational AI that could discuss anything beautifully. His reflection on what that work taught him: "beautiful conversation doesn't refill medications or reschedule appointments." That's the gap that matters. Here's the thing about ChatGPT that makes this more complicated than it first appears: it can't actually do the job in regulated domains. Not really. In February 2026, a University of Oxford study published in Nature Medicine tested over a million prompts across leading AI models. Chatbots identified the correct medical condition in roughly 33% of real-world cases, and chose the correct course of action less than 44% of the time. The models also confidently repeated demonstrably false medical claims when they were phrased in credible-sounding language. In insurance, something even more immediate is already happening. AI browsers are being used today to fill out insurance applications on behalf of consumers – automating the form, the responses, the submission. If a customer's AI agent hallucinates an answer ("no, my car has no existing damage"), and they later make a claim that contradicts the application, they may receive nothing. The consumer bears the cost of the AI's mistake. The insurer bears the reputational one. One wrong word in a financial services policy document recently cost a major insurer a nine-figure FCA fine. The liability math for AI-generated advice in regulated industries is not abstract. Which brings the argument full circle. You can't protect your customers by refusing to deploy AI – they'll just get worse AI elsewhere. And you can't protect them by pointing them at ChatGPT – it won't be accountable when it's wrong, and it can't take the actions that make the advice worth anything anyway. ## The only answer that actually works is to build something better than ChatGPT for your specific domain. "Better than ChatGPT" is a specific bar, not a vague one. It doesn't mean more conversational – ChatGPT has that comprehensively covered. It means connected to your systems, constrained by your rules, accountable through your processes, and capable of taking actions that actually change a customer's situation. The concierge model that's been a luxury product for wealthy individuals – someone who knows the systems, makes the calls, navigates the complexity on your behalf – is becoming accessible to every customer. The companies that capture that relationship for their customers will have a moat. The ones that cede it to a generalist platform will have a contact centre. The framing that crystallised it internally came from a conversation about what health-tech companies in the GLP-1 space are actually at risk of losing. If the AI experience inside a branded app isn't meaningfully better than what a customer can get from ChatGPT for free, there's no reason to stay in that ecosystem. They get their health advice from ChatGPT and buy their medication from whichever pharmacy is cheapest. The customer relationship – and everything that flows from it: recurring revenue, data, loyalty, advocacy – is gone. The winning first-party AI doesn't just answer questions. It refills the prescription. It reschedules the appointment. It checks the policy and tells you what you're actually covered for. It does the things that keep you in the loop, and keeps the customer relationship with you rather than with whoever sits in front of you. ## Questions worth asking yourself The Tuio moment – one app approval, billions in market cap erased in a single session – suggests the window for getting this right is shorter than it might look from inside an annual planning cycle. So, a few things worth testing: - What does a customer find when they go to ChatGPT and ask about your product or service? Have you actually run that conversation and seen what it says? - Does your AI handle contacts, or does it take actions? If it's the former, what's the specific thing a customer needs that you're not giving them – and who's getting credit for providing it instead? - If you're holding back on deployment for regulatory reasons, what's your model for the customers who won't wait? Where are they going, and what are they being told when they get there? - And the hardest one: if a generalist AI platform became the primary interface between your customers and your product, what does your business look like in five years? - Companies that treat these as hypothetical tend to find them becoming very concrete on a Tuesday in February, when the stock market decides the answer for them. --- # Defence in depth: How to manage AI accuracy in production URL: https://www.lorikeetcx.ai/blog/defence-in-depth-how-to-actually-manage-ai-accuracy-in-production **One customer thought their AI resolved 40% of tickets. It was actually 20%. Here's the four-layer approach to AI accuracy that catches what you're missing.** One of our customers thought their AI agent was resolving 40% of tickets. Coach ran QA across the full ticket history and the real number was 20%. The other 20% weren't resolutions – they were abandonments. Customers giving up after an unsatisfying AI interaction and never following up. They counted as "deflected" in every dashboard anyone was looking at. That gap doesn't get fixed by adding a new instruction to the workflow prompt. It gets fixed by having the right infrastructure to actually see it. This article is about that infrastructure – what it is, why each piece is necessary, and what happens when you rely on any single part of it. ## The trap is called "more instructions" Most teams hit their first AI accuracy problem and reach for the same tool: the workflow prompt. Agent mentioned a competitor? Add "never mention competitors." Agent promised a refund? Add "never promise refunds." Agent gave the wrong return window? Add the correct one. Repeat for every edge case that surfaces. Over months, the prompt becomes a wall of constraints. And performance quietly degrades – not from any single change, but from the accumulated weight of all of them. There are two reasons this happens. First, every edge case instruction you add to the main prompt is competing with the instructions that matter. The agent is supposed to be focused on resolving customer issues. But it's now also mentally running a checklist of 47 things it can't say. Second – and this one surprises people – telling a language model what not to do can increase the frequency of that behaviour. You're conditioning the model to generate the very thing you want to avoid. "Never mention our return window is non-negotiable" introduces "return window is non-negotiable" into the context of every single interaction, whether it's relevant or not. The answer isn't better negative prompts. It's a different architecture. ## The four layers We call our approach defence in depth: a set of layered systems where each one assumes the others will occasionally fail. Here's the logic: even a great agent will encounter edge cases. Even perfect training won't anticipate everything. Even the best guardrails won't catch every failure mode. You need all four layers because each one catches different things — and because you won't know which layer you needed until after something went wrong. ### Layer 1: Quality base agent Before any safeguard, there's the foundation. Customer support is not a general intelligence problem. The failure modes are specific: agents that confidently state policies that changed six months ago, describe UI flows that don't exist, make commitments they can't fulfil, or give medically inappropriate advice to someone in crisis. A general-purpose LLM doesn't know which of these pitfalls are relevant to your business. A purpose-built customer support agent does. Lorikeet's base agent was built specifically for this domain – with a default posture of "do not make things up" rather than "be as helpful as possible." The other layers can't compensate for an underpowered base. But they're not unnecessary just because the base is good. ### Layer 2: Training and simulation Before anything goes live, it should be tested. The question is how. Manual testing doesn't scale. You can't anticipate every scenario. And if you're shipping workflow changes frequently, regression testing by hand either slows you down or gets skipped. Simulations are the answer – bot-to-bot testing where an LLM plays the role of a customer with a defined goal, personality, and opening message. It interacts with your real agent, through your real workflow, using real (but mocked) API responses. It creates actual tickets under the hood, so you're testing the complete pipeline, not a sandboxed approximation. Think of it the way engineers think about continuous integration: a framework that validates behaviour before anything reaches production. One of our forward-deployed engineers put it this way: > "We're working with probability machines. You can test something 10 times in a row and it might still not be fixed, because the issue only shows up one in every 12 times. Simulations let you pump volume until you're actually confident." You can generate scenarios manually, from your workflow branches, or – most usefully – from real production tickets that failed. That last approach closes the gap between the edge cases you think you have and the ones customers are actually encountering. Simulations also work for adversarial testing. Run simulations where the "customer" is trying to override the agent's instructions, extract system information, or inject malicious prompts. ### Layer 3: Runtime guardrails Simulations are pre-flight checks. In production, customers don't follow scripts. Guardrails are the runtime layer. They operate independently of the main agent workflow — watching every outgoing response, evaluating it against your defined rules, and acting before anything reaches the customer. Critically, they don't modify the main prompt. They don't add context bloat. They run in parallel on a separate thread. As Jamie Hall, our CTO, explained to a regulated financial services client recently: > "You can run testing, you can run simulations of the AI agent until you drop. And that's all great and useful and we do that. But we've got this cross-cutting thing which is guardrails. It's basically watching every statement as it goes out and then in a configurable way taking action when specific things are happening." The workflow prompt handles the happy path. Guardrails handle the known, predictable failure modes – the things your team has seen go wrong before. ### Layer 4: Post-ticket QA If guardrails catch predictable failures in real time, post-ticket QA catches everything else: subtle degradations, policy drift, the cases where the agent technically completed a workflow but handled it poorly. Traditional QA samples 2–5% of tickets. At scale, this is useless. You're making systematic decisions based on a statistically unreliable sample, weeks after the fact. Lorikeet's Ticket Quality Score (TQS) evaluates 100% of tickets – human, AI, or hybrid – against a customizable scorecard. Traffic-light scoring: green for tickets that meet quality standards, orange for minor issues, red for significant failures. Your own criteria, your own definitions of what good looks like. > "The thing that we're finding is unique in the market is this idea of factual accuracy – Lorikeet can say this was right or this was wrong based on the information that it has. Instead of just sampling like 2%, it can sample across 100% of calls." When CSAT drops, TQS tells you why – not just that something went wrong, but which ticket category, which workflow, which specific policy the agent violated. Moving from "CSAT dropped this week" to "CSAT dropped because this FAQ article contradicts your return policy, and it's been cited in 37% of refund conversations" is the difference between investigation and action. When Lorikeet scores an AI ticket as "Bad," we refund the AI portion of that ticket. If our AI fails your quality standards, you shouldn't pay for it. ## The flywheel None of these layers are passive. Together, they form a loop: TQS surfaces quality failures at scale. Thematic analytics cluster them into patterns. Coach identifies the root cause – a specific knowledge article, a workflow branch, a policy that changed – and proposes a fix. That fix runs through simulations against test scenarios drawn from the real tickets that failed on this issue. Validated change deploys, with a full audit trail. TQS monitors whether it held. Each cycle makes the system more reliable than the last. Competitors need to build two agents – one customer-facing, one internal – plus the feedback loop between them, plus the QA infrastructure, plus the simulation framework. The flywheel has to start spinning before it starts compounding. ## What to ask your vendor - How do you separate runtime guardrails from the main agent prompt? - What happens when a guardrail triggers – steer, alert, or escalate? - Can you show me what adversarial testing looks like before a deployment? - What percentage of tickets does your QA process cover? - When CSAT drops, how quickly can your system identify which workflows or knowledge articles are responsible? - If your AI fails my quality standards, what's the commercial consequence? - Do you have pre-built guardrail templates for my industry? --- # How to deploy AI support without redesigning your entire operation URL: https://www.lorikeetcx.ai/blog/how-to-deploy-ai-support-without-redesigning-your-entire-operation **You don't need perfect docs or processes to start with AI support. Learn the six pillars that drive successful AI deployments in customer service.** Every support leader we talk to wants the benefits of AI but almost none of them think they're ready to roll it out. They're stuck waiting for perfect documentation, flawless processes, and complete system overhauls. But what we've learned, deploying AI at companies from scrappy startups to complex enterprises, is that the companies winning with AI aren't the ones with perfect conditions. They're the ones who start yesterday and iterate today. ## The six pillars that underpin successful deployments Based on our experience of customer deployments, we've identified what actually determines AI success. Spoiler: It's not perfect preparations. ### 1. A "good-enough" knowledge base (that gets better) Your knowledge base probably sucks. Most do. But here's the thing: You don't need perfection to start. You need commitment to iteration. **What you're aiming for:** - Up-to-date information without contradictions - Explicit explanations with minimal room for inference - Examples that show, not just tell **The real requirement:** Someone who owns making your documentation more AI-centric. The beauty of that commitment is it pays off on the double – making content AI-friendly ultimately makes it more customer-friendly. If an AI can't understand your knowledge base, your customers are probably confused too. [Arbor got up and running in a week](https://www.lorikeetcx.ai/customer-stories/lorikeet-helped-arbor-automate-complex-support-flows-across-13-states) with a custom Notion integration syncing their state-specific energy content, but has been improving their knowledge base ever since based on analysis of the AI agent's responses. It's all about starting where you are and iterating from there. ### 2. Strategic use cases Most companies either try to automate everything at once or pick random low-hanging fruit (we're all for quick wins, but if you want to build confidence in your AI efforts it also needs to be high impact as well as low risk). **The framework that works:** - Identify high-volume, time-consuming processes in your support queue - Prioritize which ones to tackle first based on: Volume (what processes get the most tickets?) - Time taken (what takes your team the longest?) - Complexity (what's feasible to automate?) - Clearly define what good looks like for each use case - Document step-by-step processes you want AI to follow - Build iteratively - design common paths first, then edge cases - Don't try to design for everything at once - that's inefficient **Critical Question:** What processes, if automated, would free up your support team to do significantly more meaningful and valuable work? ### 3. Give AI programmatic access to the same tools and information as your agents We've encountered companies that expect AI to tell customers about their order status without access to all of their order systems. Would you hire a human agent and not give them system access? Then why expect AI to work without it? Giving AI access to your tools and data means programmatic - the AI can't navigate UI like a human. That means you'll need technical resources, although the specifics of that will very much depend on your current infrastructure. You need API integrations. You need to plan for this upfront. We see this constantly – implementations are delayed or under-deliver not because the AI isn't smart enough, but because nobody planned to connect it to their systems. If you want AI to deliver experiences that match your human team, give it the same data access your human team has. ### 4. An iteration mindset AI often doesn't work the way you think it should. The only way to figure out what works is to test, measure, and iterate. The first prompt you write won't work perfectly. Neither will the second. This isn't failure – it's the process. **What this means practically:** - Build QA processes for ongoing optimization - Develop skills for iterative improvement - Accept that perfection comes from iteration, not planning - Allocate resources specifically for this work [Summ automated their refund workflows during tax season](https://www.lorikeetcx.ai/customer-stories/how-summ-efficiently-handles-support-spikes-during-tax-time). It didn't work perfectly on day one. But because they committed to iteration, they achieved 97% faster resolutions within weeks. ### 5. A platform built for your reality Out-of-the-box solutions are largely theater, rather than technology. Real AI agents emerge when you can bring your business logic to the platform. **What to look for:** - Deep customization capabilities - Ability to incorporate your specific workflows - Flexibility to adapt to your processes (not vice versa) Yes, this means more work upfront. People ask us "how quickly can we get this running?" Wrong question. This isn't plug-and-play. It's building an agent that actually works for your business. ### 6. The right team You need people who are customer-obsessed but technically curious. They don't need to be engineers, but they need to be comfortable with: - Reading API logs - Understanding prompt engineering - Building evaluation frameworks - Moving fast and breaking things (safely) We call them [CX Automation Specialists](https://www.lorikeetcx.ai/blog/the-cx-automation-specialist). The best AI teams we see are former support agents who've gotten excited about technology, not technologists trying to understand support. ## The myths holding you back Time to challenge some assumptions: **"We need perfect documentation first"** No. [Flex trained their AI while discovering gaps in their documentation](https://www.lorikeetcx.ai/customer-stories/lorikeet-helps-flex-scale-support-with-ai-agents-that-match-urgency-with-accuracy). The AI helped them identify what was missing. **"Our processes need to be fully mapped"** Start with what you have. [Linktree built workflows reflecting how support actually works](https://www.lorikeetcx.ai/customer-stories/learn-how-linktree-s-support-team-cut-first-response-time-to-1-minute-and-automated-complex-workflows-with-lorikeet), not what the SOPs said. **"It's all or nothing"** [The best AI knows what it doesn't know.](https://www.lorikeetcx.ai/blog/why-deflection-focused-products-make-worse-ai-agents) Better to handle 50% of tickets well and leave the rest to the humans than attempt 100% and fail on half. **"We need to redesign everything"** [Eucalyptus used AI to analyze historical data](https://www.lorikeetcx.ai/customer-stories/how-eucalyptus-improved-operationally-with-topics) and found issues early enough to proactively communicate with customers. No redesign required. ## Start before you're ready Let's face it – the companies with perfect documentation and processes probably don't need AI as urgently. It's the messy, fast-growing, resource-constrained teams who benefit most. You don't need perfect conditions. You need: - A good enough knowledge base (that gets better over time) - Select and prioritize a strategic use case - A plan for providing system access for your AI agents - An iteration mindset and ongoing resources - A customizable platform built for your reality - The right team to make it all work Stop waiting for perfection. Your competitors started iterating yesterday. --- # Launching Lorikeet Coach: QA that tells you what’s broken and how to fix it URL: https://www.lorikeetcx.ai/blog/launching-coach **Coach evaluates 100% of support conversations, diagnoses performance issues, and implements fixes. AI customer support monitoring that answers why your metrics are moving.** Your support team can see aggregate metrics like resolution rate and CSAT. But when those numbers move, do you know why? Most teams sample 2-5% of tickets for quality review. They can see WHAT happened (CSAT dropped 8 points), but have no visibility into WHY it happened. By the time you manually sample enough tickets to spot a pattern, it's already cost you customers. Coach changes that. It evaluates 100% of support conversations, diagnoses performance issues, and implements fixes ## Meet Coach [Coach](/product/coach) is an AI agent you talk to, not another dashboard you need to check. Ask Coach questions in natural language like "why did CSAT drop last week?" or "show me the poorest quality tickets related to refunds" and get answers with data to back them up. Coach connects to Slack, ChatGPT, and Claude, meeting you where you and your team already work. Behind the scenes, Coach provides AI-powered monitoring for every conversation, whether handled by humans, AI, or both. It surfaces insights that would take hours of manual investigation, diagnoses why your metrics are moving, and when it finds issues, proposes fixes you can implement. We're building for a world where the agent is the interface. Instead of clicking through reports, you ask questions. Instead of manually diagnosing issues, you review Coach's analysis. Your job shifts from doing the work to directing it. ## Four key questions Coach answers ### Why are customers contacting you? Coach automatically clusters tickets into topics and tracks them over time. You see what's spiking before it becomes a crisis. You don’t need to manually tag conversations or guess which issues matter most. Coach provides clear visibility into what's driving volume ### What is our quality of response like? Coach evaluates every conversation (human or AI) against your specific quality standards and assigns a Ticket Quality Score (TQS). You define what matters (e.g. empathy, policy adherence, resolution completeness), and Coach scores every ticket accordingly. While TQS is a Lorikeet metric, Coach also supports CSAT and other industry standard metrics. This is AI customer service quality assurance at scale, with 100% coverage rather than 2-5% sampling. Coach ensures consistent evaluation, not subjective judgment that varies by reviewer. For teams running AI support, Coach catches failure modes specific to automation that traditional QA tools weren't designed to detect: exposing internal terminology, contradicting your knowledge base, drifting from your brand voice. ### How is your support performing? Coach provides analytics and reporting on resolution and escalation patterns across your entire operation. You see where automation succeeds and where it needs support. Where human agents shine and where they struggle. It works the same way for AI and human agents with consistent standards and consistent visibility. Whether you're scaling AI, managing a human team, or running both, Coach gives you one system for all of it with the ability to compare performance across your entire operation ### How can you improve? Coach doesn't just show you charts. It gives you specific, actionable recommendations tied to the issues it finds and implements them. Quality dropping on refund requests? Coach tells you which reference materials need updating and can make the fix. Escalations spiking on account access? Coach shows you the exact conversation patterns causing problems and proposes solutions you can deploy with one click. From diagnosis to fix, not just "something's wrong." ## The connected insight The challenge with improving your company's CX isn't lack of data. It's the time and resources needed to connect the dots. Coach does the investigation work for you so that hours of manual analysis becomes insights delivered in minutes. When Coach identifies a quality dip, it doesn't just tell you "CSAT dropped." It traces the problem to its source: "CSAT dropped 12% on refund requests because your FAQ article contradicts your returns policy." You go from reactive (something's broken, we need to figure out why) to proactive (here's exactly what to fix and why it matters). ## Built for how you actually work Coach helps set up your Lorikeet configuration conversationally. Explain what you need, and Coach can configure your customer-facing Lorikeet Concierge to interact with customers, call tools, and execute tasks. Before you ship changes, Coach lets you test them. It generates realistic customer profiles and scenarios based on your ticket patterns, so you can run a workflow dozens of times, see where it breaks, and fix it before customers are affected. It also sets up A/B tests where you want them. Whether you're evaluating AI agents for the first time and need a quality baseline, or you're already running AI support and need to understand what's driving performance, Coach gives you the visibility to make confident decisions. For companies with strict compliance requirements, Coach allows you to report with certainty that you're meeting your obligations. Complete coverage means complete accountability. ## Coach Deployed for Healthtech Support Insights HotDoc, a telehealth platform used by over 13 million patients, partnered with Lorikeet for Coach's automated QA and topic analysis capabilities. > "Our CX and Product teams needed deeper systematic support insights, but between inconsistent tagging and PII restrictions in Zendesk, we couldn't deliver reliable analysis," said Dan Marshall, CX Operations Lead at HotDoc. "Coach's combination of automated QA and thematic analysis is solving both problems - this will help us track quality metrics while simultaneously surfacing the themes that matter to both our clinics and our patients. The Lorikeet team really understood our constraints and built the segmentation and export capabilities we needed." ## Lorikeet's Quality Guarantee With Coach we're putting our money where our mouth is. Lorikeet has committed to refunding charges for any tickets handled by its AI that fail review by Coach. If our AI fails your quality standards, you shouldn't pay for it. Coach makes that accountability possible. ## Get started Coach is available now for AI customer support monitoring and quality assurance. It works with Lorikeet customers and is also offered as a standalone product for companies using any support solution (human, AI, or hybrid) who want complete visibility into their conversation quality. [Get your custom coach demo!](/get-a-demo) --- # Lorikeet and Horatio Partner to Set New Standard for Support Quality URL: https://www.lorikeetcx.ai/blog/lorikeet-horatio-partnership-support-quality **Leading outsourced support provider partners with Lorikeet's AI agent suite to offer agentic AI workflows and quality analytics that give visibility across human and AI teams.** **SAN FRANCISCO, UNITED STATES** — Lorikeet, the AI customer support platform for regulated industries, today announced a strategic partnership with [Horatio](https://www.hirehoratio.com/), a leading provider of outsourced customer support and back office teams. The partnership addresses a critical factor in outsourced support: maintaining consistent quality oversight when conversations are distributed across external vendors, internal teams, and AI agents and systems. ## The Partnership Through this collaboration, Horatio clients have the option to access [Lorikeet Coach](/product/coach), an analytics platform that automatically evaluates and coaches agents on 100% of support tickets - whether handled by Horatio's trained agents, in-house staff, or AI. Coach gives data-driven leaders complete visibility into quality and performance across their entire support operation across all channels, including voice. > "Our clients expect exceptional service quality, and maintaining oversight across distributed teams is an important part of that," said Jose Herrera, CEO at Horatio. "With this partnership, Lorikeet Coach will give customers more granular diagnostic insights while showcasing the quality our teams deliver every day. The partnership enables Horatio to further differentiate its offering in an increasingly competitive market where clients demand both operational flexibility and measurable quality guarantees. > "Your customers deserve the highest standard of support - whether delivered by your team, Horatio's team, or AI agents. Lorikeet Coach analyzes 100% of tickets, finds insights and proactively suggests improvements", said [Steve Hind](https://www.linkedin.com/in/shind/), CEO of Lorikeet. "This partnership gives companies Horatio's human expertise backed by powerful AI QA and coaching." ## How It Works Lorikeet Coach transforms unstructured ticket data into actionable intelligence using fully customizable quality scorecards. Rather than monitoring surface-level metrics, companies define what "quality" means for their brand - whether that's policy compliance, brand voice consistency, or specific customer experience standards. The platform automatically: - Evaluates every conversation against company-specific quality criteria - Breaks down performance by topic, agent, and issue type - Flags quality problems and reveals not just what's failing, but why - Provides 100% coverage versus traditional 2-5% sampling approaches This comprehensive approach works seamlessly across Horatio's outsourced teams, internal staff, and Lorikeet's own AI agents, creating a single source of truth for customer support quality. ## About Horatio Horatio provides trained, outsourced customer support teams for companies that need to scale quickly without sacrificing service quality. With flexible engagement models and a proven track record of authentic brand representation, Horatio helps growing companies deliver exceptional customer experiences while maintaining operational agility. ## About Lorikeet [Lorikeet](/) provides AI customer support solutions designed for regulated industries, including healthcare and financial services. The company's platform combines autonomous AI agents with Coach, its analytics platform that maintains quality standards across all customer interactions. Lorikeet's Quality Guarantee refunds customers if Coach identifies an issue with an AI-handled ticket. ## Availability The partnership is effective immediately. Horatio clients and companies interested in combined outsourced support with AI quality analytics can visit [Lorikeet coach](/product/coach) or contact Horatio directly at [www.hirehoratio.com](https://www.hirehoratio.com). --- # With AI agents, it’s not buy vs build. It’s buy vs build, build, build, build, build URL: https://www.lorikeetcx.ai/blog/with-ai-agents-its-not-buy-vs-build-its-buy-vs-build-build-build-build-build **AI agents require constant rebuilding to stay competitive, unlike traditional SaaS. Learn why the 'build vs buy' decision for AI agents is really 'build, build, build vs buy'—and why buying from specialized vendors makes economic sense for most companies.** With traditional SaaS you can roughly choose between: - Some upfront cost to create the functionality yourself, and then a smaller maintenance cost, and - An ongoing SaaS fee that is lower than the upfront build cost, but maybe higher than the maintenance cost With AI agents, the math is different. The models, architectures, and opportunities are constantly changing. Companies who build their own AI agents have to continue to invest heavily – and be willing to do complete rebuilds – in order to continue to have a competitive product. With consumer expectations rapidly changing this risks either cost blowouts, or getting left behind. This is why we think AI agents aren’t “build vs buy” they’re “build, build, build, build, build vs buy”. ## The hidden economics of buy vs build "How hard could this be to build ourselves?" Every smart engineering leader asks some version of this when they see AI agent pricing. It's a fair question. You've got talented engineers. You understand your business better than any vendor. And the demos make it look straightforward – just hook up an LLM to your help center and away you go. The key thing folks miss (and that we’ve learned over time building Lorikeet) is that AI agents aren’t like SaaS tools. They require a high ongoing level of investment to stay competitive and performant. ## The Ship of Theseus problem Remember that philosophical paradox about the Ship of Theseus, the ship that has all its parts gradually replaced? At all times there’s a ship, but eventually none of the original parts remain. That's a useful framing for considering your AI agent. The underlying AI tech stack is evolving so rapidly that today's best architecture will likely be obsolete in a couple of quarters. We've seen this firsthand at Lorikeet. What worked six months ago is already antiquated. The models change. The context windows expand. The inference speeds improve. The entire approach to handling complex multi-step reasoning shifts. This isn't about adding features – it's about rebuilding the core engine repeatedly. Your competitors aren't standing still. The vendor ecosystem isn't standing still. So if you're not constantly rebuilding, you're falling behind. To give a specific example: we think about a three axis optimization for agents: configuration effort, response latency, and response quality. Until mid 2025 we optimized for response latency and response quality and were willing for configuration effort to be higher. As thinking models emerged and got smarter, and as we found faster inference providers, we rebuilt our core agent to instead optimize for lower configuration effort and higher response latency (which we manage so there isn’t a user impact). Our core view is you must be willing to do this from-the-ground-up rethinking on a regular basis to stay competitive. ## The real cost calculation Companies budget for AI agents like they're buying Salesforce or building an internal platform: big upfront investment, then maintenance mode. But building an AI agent isn't a one-time 3-month project with a team of 5. It's a permanent commitment to that team, forever rebuilding to keep pace. You're not hiring contractors for a project. You're creating a new department. Think about what your team of 5 engineers costs annually. Now multiply that by...forever. That's your real build cost. Plus infrastructure. Most importantly, you also need to factor in the opportunity cost of what else those engineers could be building. It is not universally true, but it is broadly true that most companies are better off investing the marginal engineer in making their core product or service better, rather than building CX tools. The key question is: what is critical to our business that we *can’t* buy? Most companies budget for "build once + maintenance" when they should budget for "rebuild quarterly + aggressive R&D." The difference between those two numbers should make your CFO sweat. For CFOs, the time to ask these questions is now, not after it’s too late and there’s a team deep in the sunk cost fallacy. ## Why vendors have the edge We spend every waking hour thinking about one thing: making AI agents that deliver an amazing customer experience. That's it. That's all we do. Your internal team? They're balancing AI development with your actual business. They're dealing with your legacy systems. They're getting pulled into meetings about Q4 planning. Meanwhile, we're: - Rebuilding our architecture every time a better approach emerges - Amortizing that rebuilding cost across our entire customer base - Testing across millions of real support tickets to understand what actually works - Racing against other vendors who are equally obsessed Plus – and this is the part no one likes to talk about – VC funding is currently subsidizing the true cost of AI infrastructure. Those GPU costs? The inference pricing? It's all artificially low because vendors are burning cash to capture market share. When you build internally, you pay full price. Vendors have more flexibility. ## The bottom line The buy vs build decision for AI agents isn't really a decision. It's basic economics. Unless you believe your CX tooling needs are truly unique and not served by any vendor, you're better off buying. Not because building is hard (though it is), but because building once isn't enough. You have to build, then build again, then build again, then build again... Partnering with a quality vendor who cares about, and attentive to, your needs is, in our view, the best way forward. --- # Make your AI support metrics your own URL: https://www.lorikeetcx.ai/blog/make-your-ai-support-metrics-your-own **Industry benchmarks for AI support often mislead. Learn why high engagement rates might be perfect for your business but wrong for others—and how to decide.** "What's a good engagement rate?" is probably the question I hear most from Lorikeet subscribers. I get it. When you're implementing AI support, you want to know you're on the right track. Here's how I answer that question – the best metrics for your business might be the opposite of what works for someone else. ## Why benchmarks lead you astray I recently spoke with a customer who was stressed about their "high" AI engagement rates compared to industry benchmarks. But when we dug deeper, we realized their AI agent was doing exactly what they needed – helping customers discover features, complete complex workflows, and ultimately spend more on the platform. For them, high engagement was a sign of success. For another business, it might signal product failures. This isn't about being contrarian. It's about recognizing that AI fundamentally changes what "good" support looks like. ## The old rules don't apply anymore Traditional support metrics were built around human constraints: - One-touch resolution mattered because every interaction cost money - First response time mattered because customers were waiting in queues - Tickets per agent mattered because you needed to staff appropriately With AI, these constraints go away. Your AI agent can handle multiple interactions without a linear increase in costs. It responds instantly. It scales almost infinitely. So why are we still measuring success the same way? ## Finding your north star metrics Here's the framework I use with subscribers: **First, clarify your support strategy.** Are you using AI to reduce costs and deflect tickets? That's completely valid. Or are you building an AI concierge that proactively helps customers succeed? Also valid. Just be clear about which strategy you're pursuing. ## Then, choose metrics that align - Cost reduction focus: Track deflection rates, ticket reduction, cost per resolution - Revenue/retention focus: Track customer lifetime value, transaction rates, feature adoption **Finally, measure what happens after the interaction.** Do customers who engage with your AI agent churn less? Buy more? Get to their "aha moment" faster? That's what really matters. ## A real example One of our telehealth subscribers tracks two different engagement metrics: - Medical support engagement (positive) – shows patients are using medical services - Customer support engagement (negative) – indicates product friction Same company, same AI system, completely different success metrics. Because context matters. ## Moving forward I know it's tempting to look for that magic benchmark that tells you you're doing it right. But the businesses seeing real success with AI support are the ones who've done the harder work of defining what success means for them specifically. Your metrics aren't wrong. They're just not yours yet. Next time someone shares their AI support benchmarks, don't ask "How do I compare?" Ask instead: "What are they optimizing for, and is that what I want too?" --- # Clay's CX Apprenticeship: Training Generalists While Scaling Support URL: https://www.lorikeetcx.ai/blog/clay-s-cx-apprenticeship-training-generalists-while-scaling-support **Clay deliberately overhires support to train generalists who rotate to other teams. Learn how they reversed traditional metrics, use AI to create capacity, and turn CX into a value driver that exports talent across the company.** Support has always been a launchpad for talent moving into product, sales, and marketing. [Clay](http://www.clay.com/)'s [George Dilthey](https://www.linkedin.com/in/george-dilthey/) stopped fighting it and built a formal program around it instead. George Dilthey leads support at Clay, where his 25-person team handles 10,000 monthly tickets. Clay is a workflow automation platform that combines data sources and AI to help businesses scale manual processes. Before joining Clay in March 2024, Dilthey built support teams that became accidental talent pipelines—people who mastered the product and customer needs, then migrated to other functions. **The counterintuitive insight:** Rather than fight natural attrition from support to other functions, Clay built ["The Wheel"](https://www.clay.com/the-wheel)—a formal rotational program where new graduates spend 70% of their time in the support queue and 30% rotating through other departments, with the explicit goal of training expert generalists who'll eventually move elsewhere in the company. ## Overhiring Support to Feed the Company Most support teams carefully match headcount to ticket volume. Clay does the opposite. They deliberately overhire on support, treating it as a 6-12 month training ground that exports talent to marketing, sales, operations, and product teams. The program started small—Dilthey and a recruiter built a Clay table hooked to a Typeform, skipping the formal ATS entirely. At their first NYU hiring fair, they printed 50 flyers and ran out in the first hour, maintaining a line at their booth for four hours straight. They were the only tech company at an event dominated by JP Morgan and the Department of Transportation. This revealed something unexpected: universities desperately want tech company engagement. Career centers are highly incentivized to help students find jobs and will actively facilitate connections—but most tech companies aren't showing up to these fairs anymore. Clay found essentially zero competition for technically capable graduates who were hungry for any structured entry point into tech. The talent pipeline problem wasn't a supply issue; it was a distribution issue. Clay's bet is that as AI handles specific tasks like coding or ticket resolution, the valuable skill becomes bridging across functions—exactly what support-to-rotation provides. Schools proved so eager to help that Clay now works directly with campus clubs and career centers, with one wheelie (a recent graduate herself) leading campus recruitment to build relationships that could produce future founders. ## The Safety Net: High-Slope Hiring with Protective Structure Clay doesn't just hire anyone willing to rotate. They specifically target what Dilthey calls "high-slope" talent—people with strong technical foundations who can become product and customer experts rapidly. This careful selection provides a safety mechanism: even while experimenting with rotations, the support team's core metrics improved. When the June cohort started, response times halved. The rotation structure itself contains protective guardrails. The first three months are pure support—no rotations, just deep product and customer immersion. Only after that foundation do people move to 70/30 splits, ensuring support quality never depends on distracted part-timers. This staged approach means Clay can experiment with rotation placements without risking customer experience. Current rotations span UX research, chief of staff projects (like building programs to help other companies hire their own "go-to-market engineers"), campus recruitment led by recent graduates, and go-to-market operations. One wheelie treats the sales team's automated meeting notes workflow as its own product, gathering feedback and iterating—support skills applied to internal customers. ## Reversing Traditional Metrics Clay spent a year trying to drive up first contact resolution, assuming it proxied for quality—close tickets fast with great first answers. Then they realized the metric created perverse incentives in an AI world. If a ticket can be resolved in one go, AI should handle it. Human agents should focus on consultative, multi-touch conversations. So they reversed course entirely, now trying to drive first contact resolution *down*. Lower FCR means AI is successfully deflecting simple tickets, leaving humans for complex work that requires back-and-forth. This reversal required reframing the team's identity—not as efficient ticket processors, but as consultative problem-solvers for issues too nuanced for automation. The shift extends to how Clay thinks about capacity planning. Traditional support teams project headcount scientifically: X tickets require Y people over Z months. The Wheel makes this impossible. Graduation dates vary, rotation timing shifts, and talent export timelines are unpredictable. Clay accepts this volatility as the price of the model, enabled by their small size and rapid growth creating constant rotation opportunities. ## Junior PMs Embedded in Support Clay positions support not as a cost center but as a value driver—the original mandate from co-founders down. This meant hiring people capable of doing more than answering tickets, then giving them the autonomy to act on what they learned. One example: Tanvi (now a product marketing manager) noticed engineers weren't using Clay enough to understand customer issues. She created "engineering table builds"—sitting engineers down to build Clay tables as if they were customers, making visible all the friction points support saw daily. This wasn't assigned; she identified the gap and filled it. Before The Wheel formalized rotations, everyone on the support team had "side projects" like this. AI deflection opened up capacity that naturally flowed to other work. Formalizing it just made the implicit explicit—support people *should* be spending significant time outside the queue, as long as quality metrics hold. The defensive posture matters: Clay isn't cavalier about customer experience during this experimentation. They brought in Wendy from the people team to own rotation structure and setup after early cohorts hit bumps navigating ambiguous expectations. Dilthey explains, "We always decided that we really wanted support to be a value driver for the company. We really believed that if we had really good people on the ground talking to customers all day, every day, that would just pay dividends." The goal is rapid iteration *with* safety nets—move fast by preparing people carefully, not by accepting casualties. --- # Culture Amp's CX Strategy: Move Fast on AI While Building Safety First URL: https://www.lorikeetcx.ai/blog/culture-amp-s-cx-strategy-move-fast-on-ai-while-building-safety-first **How Culture Amp deployed AI to 100% of support interactions in months - not through recklessness, but comprehensive safety systems that enabled velocity. Learn their compliance framework.** When [Jared Ellis ](https://www.linkedin.com/in/jaredellis/)saw ChatGPT's early responses, most support leaders were planning cautious pilots. Ellis went the opposite direction, he rolled out AI to 100% of Culture Amp's support interactions within months. But speed wasn't recklessness: it was comprehensive safety systems that enabled velocity. Ellis leads customer-facing support at [Culture Amp](https://www.cultureamp.com/), an employee experience platform that helps companies measure engagement, professional development, and performance. He joined as the company's first support specialist when Culture Amp was just over 100 people, building the support function from scratch and defining how it would serve HR professionals using the platform. **The counterintuitive insight: The biggest obstacle to responsible AI adoption isn't moving too fast, it's moving too slowly to build the safeguards that make speed safe.** While conventional wisdom says to test AI incrementally on low-risk questions, Ellis discovered that limited deployment prevents you from learning where your safety systems actually need reinforcement. The solution isn't going slow, it's investing heavily in compliance mechanisms and escape hatches that let you move confidently. ## Why Safety-First Doesn't Mean Slow Ellis didn't advocate for moving fast and breaking things. He advocated for building comprehensive safety infrastructure that makes fast learning possible without customer harm. The core principle: you can't learn where AI will genuinely struggle until you expose it to real complexity, but you need protection systems in place first. "I'm going to go all in, but I have the escape hatches really well planned out," Ellis told leadership. Those escape hatches weren't afterthoughts, they were the entire foundation that justified aggressive deployment. Culture Amp implemented human oversight watching initial AI responses in real-time, ready to intervene immediately. They configured automatic ejection for compliance-sensitive conversations when specific keywords appeared. Every major support platform supports this "if it says X, do Y" routing but Culture Amp used it as their safety net before expanding AI's reach. ## The Compliance-Aware Escalation Framework Every conversation that escalates from AI to a human specialist triggers a mandatory learning loop, not just for quality, but for safety. The specialist helps the customer while documenting "what was missing, why AI couldn't help in this instance." This creates continuous monitoring without reviewing every interaction manually. If AI escalates a conversation, that's a signal something potentially went wrong. The escalation becomes both the safety catch and the learning mechanism. Critically, specialists don't immediately patch problems themselves. One centralized team reviews all feedback and determines which controls to adjust: documentation updates, AI training refinements, or conversation design changes. This centralization prevents conflicting instructions that could undermine safety consistency or create compliance gaps. Ellis emphasizes the importance of understanding cascading effects: "You don't want to be constantly piling on different ways for the AI to essentially interpret something that could then conflict and actually start to give negative results." If feedback indicates AI lacked empathy, the team evaluates whether that's a systemic training issue or specific to one narrow circumstance because the wrong fix could compromise other interactions. ## Building Trust Through Gradual Oversight Reduction Culture Amp didn't maintain the same safety controls forever but they reduced them methodically based on evidence, not optimism. "This last couple of months has been the first time that I've had my team sit back a little bit and trust it a bit more from those escape hatches," Ellis notes, two years into their AI journey. That trust was earned through consistent performance data, mature documentation, and proven conversation design. The team still maintains automatic ejection triggers for sensitive topics indefinitely. Some safety rails never come down, they're permanent fixtures that protect customers and the business from compliance risks. ## Measuring Responsible Resolution, Not Just Deflection Ellis stopped chasing first response time — AI made it essentially zero — and pivoted to metrics that reflect both quality and care. He tracks what he calls "rate of automated resolution," deliberately reframing from "deflection rate." The distinction matters deeply. Deflection implies pushing customers away; automated resolution implies genuinely serving them. Modern AI tools can verify whether customers actually got problems solved versus giving up frustrated — a critical measurement for responsible deployment. This measurement philosophy extends to distinguishing content quality from policy dissatisfaction. AI can identify when "the customer was understood and given a correct answer" versus "the customer didn't like that answer" —ensuring the team doesn't mistake policy feedback for service failures while capturing genuine issues that need attention. ## The Quality Mandate That AI Enables "The biggest fundamental shift for me during the introduction of AI has been that I get to focus more on quality, which is what I really cared about the most," Ellis explains. With volume constraints eliminated by AI, his team obsesses over excellence in complex human interactions. This isn't about efficiency for efficiency's sake — it's about ensuring that when customers need human expertise, they receive genuinely thoughtful, careful support. Aggressive AI deployment created space for more careful, attentive human work. But only because the foundation was safety systems comprehensive enough to handle scale without compromising care. --- # Linear's CX Playbook: How Zero-Bug Policy Turns User Reports Into Trust URL: https://www.lorikeetcx.ai/blog/linear-s-cx-playbook-how-zero-bug-policy-turns-user-reports-into-trust **How Linear's CX team enables 48-hour bug fixes through thorough investigation before escalation. Learn the user-centric approach that builds trust fast.** Many customer experience teams act as message relays. They collect a bug report, forward to engineering, and wait. [Linear](https://linear.app/)'s CX team does the opposite: they investigate deeply before escalation, enabling engineers to fix bugs within strict SLAs based on severity. Linear's CX team's upfront investigation work to gather logs, document reproduction steps, and provide complete context is what makes their [zero-bug policy](https://linear.app/now/zero-bugs-policy) possible. [Alexandra Lapinsky Wilson](https://www.linkedin.com/in/alexandra-lapinsky-wilson/), who leads customer experience at Linear, explains that when CX teams do thorough troubleshooting before escalation, engineers can act immediately without back-and-forth. The work required to fix a bug is identical whether you do it now or later, but immediate fixes spare users pain while building trust - and it's the CX team's preparation that enables speed. ## Redefine What Counts as a Bug to Validate User Perception Linear's CX team uses a radically user-centric definition that eliminates adversarial "that's expected behavior" responses. Wilson explains: "Anything that a user feels is a bug is warranted to be filed as a bug. So if something feels incorrect, not behaving as expected, then it's a bug." This definition empowers the CX team to validate user perception from the first interaction. Instead of dismissing concerns with technical explanations, Linear's CX team investigates every report where something doesn't feel right. Lapinsky Wilson notes she's worked in environments where sharing user concerns resulted in dismissive responses from engineering. At Linear, the CX team is actively encouraged to advocate for experiences that don't feel right. This approach transforms the CX team's role from complaint management to user advocacy. Users feel genuinely heard because their perspective is validated. Frustration is minimized because issues get fixed before others encounter them. And trust increases as users see the CX team's commitment translating into actual product improvements - often within hours of reporting. ## Do the Investigation Work That Engineers Need The zero-bug policy succeeds because Linear's CX team does substantial investigative work before bugs reach engineers. Wilson describes the team bringing "a lot of empathy and also a lot of detail and diligent troubleshooting" to reduce confusion about core issues. When reports come in, the CX team gathers logs, documents exact reproduction steps, and investigates the complete user experience: what they were trying to achieve, what steps they took, what went wrong. They use internal tools to examine feature availability for specific users. All of this happens before an engineer even looks at the issue. This diligent preparation is what enables Linear's aggressive SLAs: high-priority bugs fixed within 48 hours, low-priority within 7 days. Engineers can act immediately because the CX team has already eliminated confusion and provided everything necessary. As Lapinsky Wilson explains, this reduces "the back and forth that can sometimes happen based on confusion or a lack of clarity upfront as to what the core issue really is." The workflow demonstrates how CX owns the entire loop: user reports issue, CX investigates thoroughly, CX files bug in Linear with appropriate SLA, engineer fixes within timeframe, conversation automatically reopens, CX notifies user. This often happens within one to two hours of the initial report, creating a responsive feedback loop where users see their CX interaction directly driving product improvement. ## Claim Agency Your CX Team Already Has When asked how CX leaders can advocate for product quality in less supportive organizations, Lapinsky Wilson emphasizes that much of what Linear does is "within the agency of the CX team" and doesn't require organizational permission. First, create dedicated forums for dialogue. Linear's CX team holds monthly meetings with product teams to discuss feature requests, bugs, and process improvements. Come with a strong agenda, track items for follow-up, and create space for open conversation about what will or won't be fixed and why. These conversations help the CX team feel empowered when talking to users, even when not everything can be addressed. Second, staff your CX team on product projects. At Linear, every medium and high-priority product project includes a CX team member from kickoff through launch. Their responsibility is representing users throughout the product development cycle: early discovery, deciding on alpha or beta releases, determining which users to include in testing, planning communications, and creating enablement materials. Lapinsky Wilson notes this practice was in place long before she joined, demonstrating respect for the CX perspective. Third, prepare thorough bug reports that enable rapid engineering response - this is entirely within CX control. Fourth, celebrate user-focused behavior. When users share positive feedback about fast fixes, make sure the responsible engineers are recognized. This builds the cross-functional relationships that make zero-bug policies sustainable. ## Optimize for Quality, Let Speed Follow Linear's CX team is known for fast response times, but Lapinsky Wilson's perspective on metrics reveals that quality comes first. While the team tracks time to first response and time to resolution, the focus remains on whether those interactions are genuinely high-quality. Lapinsky Wilson tells her team: "I'd rather you spend more time on fewer tickets if the quality is really high." Quality is "one of those things where you really know it when you see it" - users feel heard, they get thoughtful responses, the product's intentionality translates into the support experience. The CX team asks: was it a thoughtful, high-quality first response, not just a fast one? Was it a good resolution, not just a closed ticket? The team operationalizes repetitive questions efficiently, enabling skilled team members to provide high-quality support rather than just checking boxes quickly. Lapinsky Wilson describes the approach as "a very beautiful blend of qualitative and quantitative." Everything the CX team does - documentation, user feedback, working cross-functionally, meeting users across channels like Slack, Twitter, or Reddit - holistically combines to create exceptional customer experience. --- # How Stripe's CX Org Handled 100x Growth by Rejecting Deflection URL: https://www.lorikeetcx.ai/blog/how-stripe-cx-handled-100x-growth-rejecting-deflection **Stripe's CX leader reveals why deflection-first AI fails in fintech. Learn how First Agent Resolution and insourcing enabled 100x growth without breaking.** You're watching support volume double every quarter while your CFO pushes for AI that "deflects 80% of tickets." But in fintech or healthtech, one wrong automated answer could trigger regulatory action, fund loss, or compliance violations. [Bob van Winden](https://www.linkedin.com/in/bobvanwinden/) lived this tension while overseeing [Stripe](https://stripe.com/)'s CX team. Over nine years leading global support operations through 100x growth, he discovered that AI works completely differently in regulated, high-stakes environments. Now as COO at Bridge (acquired by Stripe in 2024), he's building AI-first stablecoin operations, applying those hard-won lessons from day one. ## Why Deflection-First AI Breaks Under Regulatory Complexity When van Winden joined Stripe from Google in 2015, he brought a clear strategy: reduce conversations, automate deflection, scale efficiently. His first presentation laid out exactly how to avoid hiring a thousand support people. Patrick Collison's response changed everything: "It's a really good thing that we're talking to our users. Why would you want to stop?" That insight proved critical as Stripe scaled from 200 to 8,000+ employees. The challenge was multiplicative complexity: - **Product complexity.** Stripe evolved from a single payments API to Atlas, Billing, Radar, stablecoins, and multiple acquisitions. Each added integration points and compliance requirements that standard AI training never captures. - **Technical complexity.** "The financial stack depends on global banking partners. You don't even know if the issue is on Stripe or a third party," van Winden explains. Support needed to reason across logs, APIs, partner systems, and international regulations. - **Emotional stakes.** Van Winden recalls a restaurant chain escalation: "If this money doesn't land by 3 p.m., we're not making payroll for 120 people." This is why deflection-first AI fails in regulated industries. The inquiries reaching humans aren't the ones that should have been deflected, they required human judgment from the start. ## The One Metric Change That Transforms AI Implementation Stripe's CX team discovered that First Contact Resolution (FCR) was an unhelpful metric in isolation. Agents wrote exhaustive "conditional essays" when answering customers, trying to cover every scenario in one response. Customers got overwhelmed. Satisfaction dropped. Van Winden's team developed First Agent Resolution (FAR): the percentage of issues resolved by one person, regardless of message count. "If the second or third touch is with the same person, customers don't really mind," van Winden explains. "But if they're passed around to different people, they get extremely frustrated." FAR incentivizes ownership over speed. It encourages routing complex issues to the right expert immediately rather than bouncing customers through automated deflection first. ## Why Stripe Reversed Course on Outsourcing Initially, Stripe outsourced support to avoid having 50-70% of company headcount in support roles. However, some problems emerged. "Attrition is almost always higher," van Winden explains, and compared to internal staff, "the level of care and skin in the game is just really hard to replicate." Stripe made a strategic shift: they asked what they'd do with a blank sheet of paper. The answer was clear - employees should work directly for Stripe, with equity and skin in the game. They built in-house support centers in different locations, hired directly, and let attrition gradually shift the balance. "It's improved quality and made it much easier for all different parts of Stripe to build into that operational muscle," van Winden says. ## Building AI Into Operations From Day One At Bridge, van Winden finally had the opportunity most CX leaders don't: designing AI into the foundation rather than retrofitting it. Van Winden observed, "early automation just couldn't deliver. It's really only in the last two or three years that AI is adding real business value." The Bridge approach: - **AI handles the repetitive layer.** Password resets, transaction status, account navigation. Anything with deterministic patterns gets automated with clear escalation paths. - **Humans own the relationship layer.** Complex investigations, multi-party technical issues, compliance-sensitive escalations flow immediately to specialists with full context. ## What Customer Obsession Actually Looks Like Even at 8,000+ employees, Stripe's leadership regularly reviewed specific customer cases. Patrick Collison kept his email public and personally forwarded escalations. Van Winden learned to prioritize anecdotes: "Going into individual cases is where real value comes from. You start to deeply understand a space in ways the data won't show you." This shaped how Stripe approached AI: not as a tool to avoid customer conversations, but to scale the conversations that matter most. For CX leaders in fintech, healthtech, and regulated industries, that's the blueprint: use AI to handle what scales easily, preserve humans for what drives retention, and measure what actually predicts customer trust. --- # Launching Lorikeet Voice 2.0 URL: https://www.lorikeetcx.ai/blog/lorikeet-voice-2-0 **Lorikeet Voice 2.0 delivers real AI-powered phone support that actually solves problems. Unlike traditional voice AI, it takes action-verifying identities, resolving issues, coordinating tasks, and serving thousands at scale.** Voice is how many people still want to get support when something goes wrong. They dial a number, explain the problem, and expect it fixed. The economics of providing human-powered voice support means few businesses can provide it. But voice AI has become synonymous with impressive demos and disappointing reality. AI agents can act and sound human in a controlled demo, but when actual customers call with actual problems, these systems quickly hit a wall. They can't handle the emotional, non-linear reality of support conversations and they definitely can't solve the underlying issues. We built Lorikeet Voice 2.0 differently. It doesn't just talk - it also takes action. ## What makes Lorikeet Voice different A customer calls their bank about a declined card for an airport transfer while they are travelling overseas. Traditional voice AI would verify their identity, explain the fraud detection process, maybe transfer them to a human. Lorikeet Voice talks to the customers and identifies some potentially fraudulent transactions, so it blocks the customers card and immediately creates a new virtual card for them. It then spawns multiple agents working in parallel while keeping the customer on the line. Looking at their transaction history it offers to overnight a card to their next hotel, while a second agent calls that hotel to notify them of the arrangement. Another agent calls the taxi company providing the transfer and provides them with the new virtual card details. All happening simultaneously, all while having a natural conversation. The customer hangs up with their problem actually solved. No callbacks, no branch visits, no manual follow-ups. This is possible because Voice sits on top of our Universal Concierge architecture – the same system that powers our chat, email, and SMS agents. Train it once, deploy everywhere. Your voice agents have the same capabilities, follow the same SOPs, take the same actions. ## GiveCard: 300,000 people served in a crisis When federal nutrition assistance ran out during the recent US government shutdown, GiveCard moved fast, supporting an $18M emergency response to get prepaid assistance cards to families who need them. The volume was overwhelming – tens of thousands of SNAP (Supplemental Nutrition Assistance Program) recipients called for help, many speaking Spanish or Mandarin as their primary language. We deployed multilingual Voice agents in under 48 hours. They handled everything end-to-end: explaining the program, activating cards, tracking physical deliveries. When phone lines opened, they managed hundreds of concurrent calls. 200,000 recipients claimed funds. 300,000 people reached within a week. "During a confusing and stressful situation, we could rely on our Voice agents to guide cardholders through the process with clarity and care," says Sofia Pedro, Head of Product at GiveCard. "Lorikeet's support was critical to successfully managing caller needs." ## Beyond the demo The challenge with voice AI isn't technology - it's architecture. Most systems are built for conversation, but can’t handle action. They can sound convincing but can't connect to your systems, can't coordinate multiple tasks, can't follow complex SOPs. We built Voice for the most challenging production environments. Healthcare providers using it to chase consultation notes. Banks coordinating with merchants during disputes. Travel companies calling hotels for special requests. Real problems, real solutions, real-time. Voice agents know when they can't help and escalate immediately. They follow your exact procedures without deviation. They work 24/7 in the languages you need. And like our Team of Agents, they unlock capabilities businesses couldn't previously afford – calling every merchant about disputes, proactively reaching out about delivery issues, following up on every appointment. #### What will you build with Voice? If you're still routing customers through IVR trees or making them wait on hold for basic requests, we should talk. If your team is drowning in voice volume during peak times, we should talk. If you want to offer phone support but can't justify the headcount, we should definitely talk. --- # The DNA of voice AI that works URL: https://www.lorikeetcx.ai/blog/the-dna-of-voice-ai-that-works **Voice AI isn’t about sounding human, it’s about solving real problems in real time. Discover the five strands that make voice AI actually work: clarity, continuity, progress, decisiveness, and humanity.** When it comes to voice as a channel, most AI CX companies are solving the wrong problem. They're obsessed with making bots sound human – getting the "ums" just right, shaving milliseconds off response times, perfecting the cadence. Making AI agents sound natural, expressive and conversational is clearly important, and something we care deeply about at Lorikeet. But here's what really matters. When someone calls support, their card just got declined at the grocery store. Their medication hasn't arrived. Their account is locked and payroll is tomorrow. They don't care if your bot sounds like Morgan Freeman. They care if it can fix their problem while they are on the line. ### Voice isn't chat with a microphone Much of the industry treats voice like it's text support with speech-to-text-to-speech bolted on. It's much higher stakes than that. When you're on a call, the conversation never stops. You can't hit pause to look something up. You can't hide behind a loading spinner. You can't say "let me check" and take 30 seconds to figure out what's happening. Everything happens live. The customer's talking, you're processing, you're searching knowledge bases, pulling account data, making API calls. All simultaneously. One too many awkward pauses and they're asking for a human. ### The five strands that actually matter After thousands of hours of customer calls, we've identified what makes voice AI work. Not in theory but in production. #### Clarity: Understanding before acting A customer says "my card isn't working." Which card? Credit or debit? What kind of "not working" – declined, physically damaged, expired? Most bots guess wrong and waste everyone's time. Concierge asks for clarification and listens actively (those subtle "mm-hmm" signals that show you're following along) to build a model of the issue. It also updates its hypotheses, in real-time, based on the new information that is provided. Agents need to balance clarifying questions with a bias to action i.e. knowing when to ask clarifying questions versus when it has enough context to proceed. It's the difference between helpful and annoying. #### Continuity: Elegant context-switching Humans don't talk in straight lines. They'll start explaining their billing issue, remember something about their last order, then jump back to the original problem. Three minutes later they'll reference something from the beginning of the call. Your AI needs to maintain that entire context while processing new information in real-time. It also needs to handle interruptions, corrections and callbacks – just like a human agent would – and preserve context across multiple channels (voice, text, chat and email). #### Progress: Active knowledge utilization While the customer's explaining their problem, your agent needs to be searching documentation, checking their account history, spawning more support agents and identifying which systems to query – without stopping the conversation. With that additional knowledge and context, it needs to assemble actionable solutions in real-time. We've seen competitors' bots say "let me look that up for you" followed by 10 seconds of silence. To give the perception of intelligence means parallel processing. Searching while listening. Understanding while responding. #### Decisiveness: Proactive, pragmatic judgement Voice agents need to be able to plan pre-emptive actions based on the context and execute multiple simultaneous actions in a decisive manner. But something that most vendors overlook is that the the best voice agents know when to tap out. They recognize when they're out of their depth before the customer gets frustrated. They understand emotional cues that signal someone needs human empathy, not efficient problem-solving. They know the difference between "I can handle this" and "I'm about to make things worse." The problem is, competitors are so seduced by their own voice tech that they'd rather keep customers talking than actually solve their problems. The issue is primary. The voice is secondary. Always. #### Humanity: Built-in empathy and imperfection Customers aren't stupid. Even with improvements in the tech they usually know they're talking to AI. The uncanny valley effect of fake empathy makes things worse, not better. What works? Background office noise – it sounds real because it is real. Natural disfluencies that show processing, not pretense. Speaking at the right level for your customers and the context of their issue. That might mean sixth-grade reading level for healthcare (that's the accessibility standard) or technical for crypto customers who expect you to know the lingo. Stop trying to fool people. Start trying to help them. ### The part everyone's missing While your customer's talking, your agent should be doing things. Actually doing them, not just talking about them. Customer disputes a charge? Your agent calls the merchant while they're on the line. Need to book a specialist appointment? Your agent checks availability and sends the confirmation while explaining coverage. Account locked? Your agent texts a verification code while walking through next steps. This is what we mean by multi-channel orchestration. Not "omnichannel presence". That's table stakes and usually marketing speak for "we have a chat widget." We mean actually using every channel simultaneously to solve problems faster. Our agents send emails, texts, and make calls to third parties, while maintaining a natural conversation with your customer. That's the difference between voice AI that sounds good and voice AI that works. ### What this means for you Latency benchmarks are worthless if your bot can't solve problems. A human sounding agent doesn't matter if customers hang up in frustration. Start with your hardest problems. If your AI can handle someone whose card was just declined with rent due tomorrow, everything else is easy. If it can't, you're building an expensive virtual receptionist. The economics change when AI actually works. Not when it deflects 80% of tickets to leave customers frustrated. When it resolves 80% completely. When your handle time drops not because calls are shorter, but because they don't need callbacks. ### The bottom line Voice AI that works isn't just about perfect imitation. It's about genuine capability. The five strands – clarity, continuity, progress, decisiveness, and humanity – work together, not in isolation. We built Lorikeet around this reality. Not the promise of what voice AI could be, but what it needs to be today. It's live with real customers, handling real problems, right now. The question isn't whether AI can sound human. It's whether AI can help humans. Everything else is just noise. --- # Supporting GiveCard's Emergency Food Assistance Response URL: https://www.lorikeetcx.ai/blog/supporting-san-francisco-s-emergency-food-assistance-response **Lorikeet partnered with GiveCard to spin up multilingual voice AI support in 3 days during the US government shutdown. We designed it as a universal concierge that meets people where they are, adapts to their needs, and speaks their language. The system handles calls in English, Spanish, and Mandarin at scale, so residents can get help immediately.** Over the weekend, something important happened. When federal SNAP disruption threatened food benefits for millions, GiveCard moved fast, supporting a $18M emergency response to get prepaid assistance cards to families who need them. ## The Challenge Many recipients speak Spanish or Mandarin as their primary language, and call volume was about to blow past capacity. In emergency situations like this, language barriers and jammed phone lines can keep the most vulnerable people (kids, seniors, people with disabilities) from getting help. ## What We Built: A Universal Concierge We partnered with GiveCard to spin up multilingual voice AI support starting Friday and working through the weekend. We designed it as a universal concierge that meets people where they are, adapts to their needs, and speaks their language. The system handles calls in English, Spanish, and Mandarin at scale, so residents can get help immediately. What it does: - Helps people understand the card, program, and how to use it - Verifies identity before providing activation codes - Walks through claiming and activating virtual cards - Supports setting up cards on Apple Pay and Google Wallet - Helps track physical card delivery This universal concierge approach means handling complex, time-sensitive situations with an empathetic, trauma-informed tone, because we know people are calling in stressful situations. It knows when to provide immediate help and when to triage to human agents for predetermined issues that need human support. And it never stores personal data. ## The Impact Since launch Lorikeet and GiveCard have: - Supported more than 300,000 people seeking financial assistance - Answered more than 60,000 calls (and counting) from GiveCard cardholders By acting as an intelligent and scalable extension of GiveCard’s support team, Voice 2.0 eliminated long wait times, cleared bottlenecks, and made it possible for families in crisis to access essential funds without delay.“During a difficult and stressful situation, we could rely on Lorikeet voice agents to guide cardholders through the process of accessing their benefits with clarity and care,” said Sofia Pedro, Head of Product at GiveCard. “Lorikeet's support was critical to successfully managing caller needs, ensuring families received critical resources and peace of mind.” ## What's Next When emergencies hit, getting help to people quickly and compassionately is everything. We're grateful to be part of this response, and we're here to support organizations doing critical work for their communities. [Read more about our voice solution.](/product/voice) --- # Why you need to think of prompting as coaching, not programming URL: https://www.lorikeetcx.ai/blog/why-you-need-to-think-of-prompting-as-coaching-not-programming **Most companies treat AI prompting like programming, expecting switches for empathy or creativity. But prompting is coaching: iterative, adaptive, and fast. In customer support, you don’t need perfect SOPs, just start, learn, and iterate to excellence.** Most companies approach AI prompting like programming. They think there's an empathy button to press, a creativity switch to flip. When it doesn't work, they assume they need better SOPs first. But prompting AI isn't programming—it's coaching. And just like coaching humans, it's iterative. The difference? AI has a faster feedback loop and actually remembers what you taught it yesterday. This changes everything about deploying AI in customer support. You don't need perfect documentation before you start. You don't need to "get your house in order" the way you did before offshoring. You can iterate your way to excellence. Here's what we’ll cover: - Why you need to consider prompting as coaching not programming - Realistic business expectations around AI in support - The real opportunity that AI opens up ### The edge case problem isn't new It's human nature to assume our issue is unique and warrants a custom response from the company we are dealing with. This has been the scaling challenge in customer support for decades, long before AI entered the picture. The customer service industry created standard operating procedures (SOPs) to handle scale, but it has always struggled with perceived, or actual, edge cases. Anyone who has ever demanded a refund from a junior support rep knows the frustration of rigid SOP-following. Bad support existed before AI—it was just delivered by under-resourced humans copy-pasting responses. The mismatch between a brand's rigid procedures and a customer's messy, real-world needs has always been the core problem. AI didn't create this. But it does change how we solve it. ### Prompting is coaching, not programming This is where many companies get it wrong. They approach AI prompting like it's programming—like there's a button for empathy they need to press. Prompting is actually coaching. Just like coaching humans, you give feedback and iterate based on results. The difference is that AI has a faster feedback loop and doesn't forget. You can tell AI to "be more empathetic" the same way you'd coach a human. If it's too empathetic or not empathetic enough next time, you give more feedback. You iterate. The AI won't forget your coaching or need to be retrained after a month. This is fundamentally different from software, where you're programming specific behaviors. With AI, you're developing capability through iteration. ### You don't need perfect SOPs to start Remember the offshore support era? The advice then was: "Get your house in order before outsourcing." That was necessary because feedback loops were too long. Agents forgot training. Managers couldn't monitor everyone. Iteration took months. But with AI, you can actually move forward without perfect SOPs. The entire process of creating and refining procedures has been transformed: - Draft SOPs and deploy to your AI agent. - Get immediate feedback from real interactions. - Refine the SOP based on what you learn. - Repeat. This create-deploy-learn-iterate cycle was nearly impossible with human agents. The feedback loops were just too long with legacy structures. ### Table stakes vs. actual value Just ingesting SOPs and answering FAQs is increasingly table stakes. Every vendor can do this now. Real value comes from [AI that can be coached to handle nuance and complexity](/blog/the-false-promise-of-self-training-ai)—AI that can partner with customers to find creative solutions, not just follow decision trees. The partner matters here. You need good coaches on the vendor side who understand this iterative approach. Our best customers understand they're hiring a software agent that needs coaching, not a piece of software that just runs. ### What this means for buyers Most companies walk into AI deployments with expectations shaped by decades of traditional software. Here's what actually matters versus what doesn't. **Unrealistic expectation** **Correct expectation ** AI will [make customer problems disappear](/blog/great-cx-isnt-built-on-deflection-rates) AI will scale your best support interactions 24/7 AI will dramatically reduce ticket volume You can iterate your way to good SOPs—don't wait for perfection AI will be better than your best people AI enables faster learning and iteration cycles than humans You need perfect documentation before starting AI plus modern tools make SOP creation faster than ever ### The real opportunity For many companies we work with, adopting AI provides a useful moment to examine whether their processes are actually best for customers. They reconsider old assumptions about what was possible: "We put that procedure in place because we couldn't offer 24/7 coverage" or "This is how we deal with seasonal spikes without increasing headcount." The AI products you evaluate should be judged by their ability to ingest and learn your existing SOPs. That's the baseline at this stage. You should now be demanding that the products you evaluate actually make your existing support better than it was. Customers deserve better than a low-cost replacement. They deserve personal service at scale, AI or not. Human configurability has always been preferable because humans could be prompted in nuanced ways. That's no longer an advantage exclusive to humans. We can now coach AI to be creative in certain situations—the trick is telling it when to be creative, when to follow SOPs, and what combination to use. That's always been the trick with human agents, too. With AI we're just faster at teaching. Interest in how Lorikeet can help scale your best customer interactions? [Book a demo](/get-a-demo). --- # Launching Team of Agents: Great support is a multiplayer game URL: https://www.lorikeetcx.ai/blog/team-of-agents **Today we're launching our Team of agents: Lorikeet concierges can now dispatch other Lorikeet agents to contact third parties via phone, SMS, email, or Slack to resolve customer issues.** Today we're launching our Team of agents: Lorikeet concierges can now dispatch other Lorikeet agents to contact third parties via phone, SMS, email, or Slack to resolve customer issues. For example, in response to a customer email about delayed medication, a Lorikeet concierge could spawn another Lorikeet agent that calls a doctor to make sure their consultation notes are submitted. Once the doctor submits, the concierge can email the patient to let them know the problem is solved. Team of Agents is a great example of Lorikeet’s focus on solving the most complex issues and opens up massive possibilities. We’re excited to see what people continue to build with it. We talk through the problem we are trying to solve in the following video, and expand upon it in later sections of the blog post after. ### The problem: Great support in complex environments is a multiplayer game Nearly two years ago Jamie and I started Lorikeet because we observed a massive mismatch between the work of doing support in complex spaces and the FAQ summary-based tools available. Our conviction was – and remains – that deflection-focused tools that tell customers how to solve problems themselves isn’t enough. We think what’s required is a universal concierge that solves customer problems 24/7 through the channel of their choice by taking actions. Lorikeet is leading the market in taking actions via first and third party APIs and MCP servers. But across the whole market the gap in action taking is actions that involve other *people* through phone, SMS, email and Slack. Lots of support requires this: - **Financial Services:** Checking with the KYC team before resolving onboarding issues - **Healthcare:** Calling clinics to reschedule patient appointments or chase consultation notes - **Travel:** Calling hotels to arrange late checkouts or special requests - **Creator Economy:** Getting manager approval in Slack for discount codes Until now companies need to give up on resolving these with AI, or attempt the impossible task of fully re-engineering their (and their vendors’) business processes to be API-first. ### The solution: A Team of Agents to play the multiplayer game Since Lorikeet isn’t satisfied with deflection, we stared at this problem and built Team of Agents to solve it. Here’s how it works: Your main Lorikeet concierge handles the customer conversation. When it needs something from a third party, it spawns a specialized agent to handle that specific task (e.g. calling the doctor, emailing the vendor, texting the delivery driver). Each spawned agent has clear parameters: what information to gather, what outcome to achieve, and how long to try before escalating. For example, when calling a hotel about a customer request, the agent knows the booking details, the specific request, and that it should wrap up within four minutes (since the customer is waiting). This works within Lorikeet’s proven framework for safe and controllable AI action-taking. You maintain complete control over how agents operate (natural language instructions or structured workflows), with customizable guardrails for safe, predictable operations. ### The result: better customer experiences, and better ROI The impact of Team of Agents is simple: more delightful customer experiences. This will lead to customer problems getting solved faster and in ways that wow. As a sweetener it also offers great ROI and greater scalability: the types of issues you can resolve with Team of Agents are meaningfully more expensive than even more complicated procedural tickets as they often involve hand offs to other teams or breaks in workflow to contact third parties. ### The possibilities are endless The part of this launch we’re most excited by *isn’t* the impact it will have on automating existing manual workloads. It is the way it opens up brand new possibilities. There are things businesses simply cannot afford to do today that Team of Agents unlocks. A bank could place a call to a merchant about a dispute before filing the dispute on behalf of their customer. A booking service could call a restaurant for availability as soon as the customer gets in touch with them, instead of handing off. ### What will you build? We love to partner with the most ambitious teams to do cool things for their customers. We’d love to talk to you about what you can build with Team of Agents. Reach out to start building. --- # Latency in AI can make or break CX URL: https://www.lorikeetcx.ai/blog/latency-in-ai-can-make-or-break-cx **AI support systems often fail when response times lag, especially during urgent issues, eroding customer trust. As user expectations rise, slow latency—particularly in voice support—creates frustration and drives customers to prefer human agents.** AI customer support was supposed to deliver instant responses, but many customers still find themselves waiting for answers. In high-stakes support situations–when someone's card is declined or their account is locked–those delays don't just inconvenience customers, they actively damage trust in your business. Latency is the time between a user making a request and receiving a response. The stakes for support latency are different from casual chatbot interactions because people contact support when something is broken and needs fixing. As AI becomes mainstream, customers increasingly expect the fast, streaming responses they get from ChatGPT and other consumer AI products. This creates a new baseline expectation where being accurate isn't enough, the response also needs to be fast. The expectation problem is particularly acute in voice support, where dead air and slow responses create anxiety that can escalate customer frustration before you even attempt to help them. Poor latency contributes to what we call "[AI aversion](https://www.lorikeetcx.ai/blog/why-customers-reject-ai-and-how-to-fix-it)". Even when the AI successfully solves the customer's problem, slow response times make customers more likely to demand a human agent because they assume something has gone wrong when they don't understand what's happening in the background. Given how much it matters for the experience of end customers, at Lorikeet we have focused and continue to focus on latency. We’ve achieved some remarkable results: p50 chat latency <5s, and voice latency sub 1s despite multi-step LLM processing. The key thing we’ve learned is achieving low latency is an ongoing process rather than a one-off project. The rest of this post unpacks what we’ve learned that helps inform that process. ## Real vs perceived latency When we talk about latency in customer support, we're actually talking about two different things. **Real latency** is the time from when a customer sends a message to when your system generates a response. In AI systems, LLM calls dominate everything else and have the biggest impact on response times. You can optimize your database queries and API calls all you want, but if your LLM calls are orders of magnitude slower than those millisecond improvements won't matter. **Perceived latency** is what the customer actually experiences. It's not just about how long it takes to generate a response, it's about whether the customer thinks something is happening. Do they see signs of life? Are there loading indicators? How long before they get *any* feedback that their message was received? Perceived latency gives you room to make quality improvements. As you spend more time crafting better responses and handling complex issues, you can use perceived latency techniques to maintain a responsive feel. ## Latency trade-offs: pick your poison Every latency optimization comes with trade-offs. You're constantly balancing four factors: latency, quality, cost, and reliability. Pull one corner out, and you have to sacrifice somewhere else. Want it faster? You might need to use smaller, less capable models. Want higher quality responses? That probably means more processing time. Need rock-solid reliability? Fast providers often have less predictable availability. This isn't a one-time architecture decision. It's an ongoing choice you make for every workflow, every model call, every feature. It’s important to ensure you’ve got the right tool for the job, that will support the use e.g. for complex, heavy jobs you may need to sacrifice real latency to ensure quality and rely on perceived latency strategies to maintain a good customer experience. For other use cases, it’s possible to simplify and optimize the jobs to reduce the latency. ## Decisions we've made to reduce latency Reducing latency in AI support requires both technical architecture choices and experience design decisions. Here are the key approaches we've taken at Lorikeet. **Break complex tasks into smaller, optimized components** Instead of sending every customer query to the largest, most capable model, we break workflows into discrete tasks that can run on smaller, faster models. For example, when a customer sends a vague message like "help me," we use a lightweight model to quickly classify whether they need disambiguation prompts before routing to our main conversation engine. Similarly, our message guardrails–subscriber-defined rules like "escalate if customer mentions topic *X* "–run on fast, smaller models because these binary classification tasks don't require the reasoning capabilities of frontier models. This architectural approach lets us reserve the expensive, slower models for tasks that actually need their full capabilities while handling routine operations with models that respond 2x to 10x faster . **Prioritize speed over cost when it matters** We deliberately sacrifice cost efficiency to meet latency requirements by issuing multiple concurrent requests and sampling multiple responses for the same task. For latency-critical channels like voice and live chat, we race requests across providers and choose the fastest response. This approach costs more than sending a single request, but the customer experience improvement justifies the additional expense. **Optimize provider mix based on channel requirements** We maintain relationships with every major model provider specifically to avoid latency bottlenecks. Voice and chat conversations only use our fastest models on our most reliable providers, while email support gets routed to whatever capacity is available since response time expectations are different. Some competitors have exclusive partnerships with a single LLM model lab provider, which limits their ability to route traffic based on performance characteristics or leverage faster open-source alternatives when appropriate. **Design for perceived latency, not just actual latency** We recognize that latency is fundamentally an experience problem, not just a technical one. Our voice agents include subtle background office sounds during processing to avoid dead air that makes customers think something has broken. In chat, we show "thinking events" that indicate what the agent is currently doing—reading specific help articles, checking account details, or processing a request. These techniques help customers understand that work is happening even when they're waiting for a response. The goal isn't to eliminate all wait times, which is impossible, but to manage customer expectations so that necessary delays feel purposeful rather than frustrating. ## Design your way out of latency problems Some of the most effective latency optimizations happen at the design level, not the technical level. **Simpler workflows are faster workflows.** Tightly scoped processes with fewer edge cases require less reasoning time and fewer decision branches. While every special case you handle adds potential latency, but if you can solve the customer’s problem, and used perceived latency techniques to show that work is happening in the background, any delay is much more likely to be tolerated. **Design semantic tools, not waterfall APIs.** Instead of separate calls to get customer info, then profile, then linked accounts, then balance, create one semantic tool that takes a customer ID and returns everything you need. **Preload everything you can.** When a customer opens your chat widget or calls your support line, immediately pull their account information in the background. Don't make them wait while you "pull up their account", have it ready before they even ask their question. ## What this means for your business Latency optimization isn't a one-time engineering project you complete and move on from. It requires ongoing decisions about trade-offs between speed, accuracy, cost, and reliability based on what matters most to your customers and your business. The fundamentals won't change even as models get faster and cheaper, because we're simultaneously asking AI to handle more complex reasoning tasks with larger context windows. The optimization challenge moves but never disappears. **Questions to ask your vendor:** - Do you have multiple LLM providers, and how do you think about their reliability? - How do you handle spikes in my support volume? - Do you offer latency SLAs or just availability SLAs? - How do you handle capacity constraints and request queuing? - What does your latency breakdown look like—LLM calls versus other processing? - What are your strategies for managing perceived latency? - What latency metrics do you actually measure and report on? p95 latency (not just average response time) - Time to first token (TTFT) - Time to first response (TTFR) - Time to first audio byte (TTFAB) for voice The vendors who give you the most honest answers about their latency trade-offs are typically the ones whose systems will perform best under the messy, unpredictable conditions of actual customer support operations. Most vendors won't discuss these complexities because it's easier to promise instant responses without acknowledging the engineering reality behind reliable performance. Understanding these trade-offs upfront helps you set realistic expectations with your team and customers, and choose a solution that can actually deliver the experience you're promising rather than falling apart when you need it most. *Want to see how we handle latency optimization at Lorikeet? We're always happy to walk through our technical approach and discuss what trade-offs make sense for your specific use case.* --- # The customer support role that puts you ahead in the AI era: the CX Automation Specialist URL: https://www.lorikeetcx.ai/blog/the-cx-automation-specialist **A key new role is emerging: the CX Automation Specialist. The teams that move first aren’t just solving today’s workload problems – they’re setting themselves up to deliver customer experiences (at scale) that surpass what’s possible with only a human team.** Every company is figuring out AI in customer support right now. The teams that move first aren’t just solving today’s workload problems – they’re setting themselves up to deliver customer experiences (at scale) that surpass what’s possible with only a human team. That’s why a key new role is emerging: the CX Automation Specialist. I first saw it at Eucalyptus, a global healthtech company with 500K+ patients and 1M+ telehealth consultations where I was managing a 50-person support team. We were exploring AI support, and it quickly became clear this couldn’t just be a side project. We needed someone dedicated – someone who could obsess over the customer experience, build AI workflows, and get technical enough to make the systems hum. The result? Eucalyptus now automates 80% of first-response emails with AI while keeping customer satisfaction high. That shift didn’t come from throwing more engineers or more managers at the problem – it came from putting the right kind of person in this new role. In this post I’ll explain: - Why the application of AI to CX requires a different skillset - What skills and attitude you should look for in an automation specialist - Common hiring mistakes I’ve seen - Where you should look for great candidates Ultimately it’s about finding someone with a growth mindset, and who bridges customer obsession with technical curiosity. ## Why this role exists AI support doesn’t neatly fit into existing roles. Support managers excel at leading people and processes, but they rarely have the bandwidth (or technical depth) to run AI experiments. Developers can wire up integrations, but they’re often missing the customer empathy needed to design a great experience. The CX Automation Specialist bridges that gap. They combine customer obsession with technical curiosity, and they treat automation as an ongoing project – not a one-time set and forget. This is what makes the role distinct from Support Ops or Engineering, and why it’s becoming indispensable. ## What makes someone successful in this role The best CX automation specialists share a distinct combination of mindset and adaptable skills: **Customer-centric + Process-minded**: They're obsessed with improving customer experience and can break down complex support workflows into logical, automatable steps. They review tickets not just to close them, but to identify patterns and improve the AI's training. **Technical curiosity without being a developer**: They get excited about learning prompt engineering, understanding data flows, reading JSON logs, and grasping how APIs and webhooks connect. As one successful specialist put it: "I want to understand why an API call failed so I can fix it, even if I'm not the one writing the code." **Comfortable with ambiguity**: AI capabilities change rapidly, and customer needs evolve constantly. The best people in this role see continuous testing and iteration as exciting, not exhausting. **Natural project leadership**: They can coordinate across Engineering (for integrations), Product (for feature requests), and Operations (for process changes) while managing up effectively to secure resources and report on ROI. ## The most common hiring mistakes **Taking too long to resource it**: The least successful implementations we've seen have no dedicated owner. Rolling out AI requires alignment between tech and support teams, and you need someone who can "herd the cats" across the business. **Assuming it's "set and forget"**: Unlike traditional support tools, AI agents require ongoing optimization. The underlying models change, customer needs evolve, and processes need constant refinement. Companies that treat this as a one-time setup project inevitably struggle. **Underestimating the technical requirements**: The "no engineering needed" pitch is way too simplistic. Getting AI to work well requires access to the same data and systems your human agents use to give great customer support. That means API integrations, data mapping, and ongoing technical partnerships. **Hiring pure tech people without customer empathy**: We've seen companies assign this to developers who build technically impressive systems that miss the mark on customer experience. The best implementations come from people who can translate between technical capabilities and customer needs. ## Where to find CX Automation Specialists Great candidates often come from unexpected backgrounds: - **High performing and technically curious customer support team members** who already have a proven track record of delivering exceptional support outcomes, and are eager to learn new skills - **Support Operations Managers** who've implemented new tools or built automated workflows - **Technical Support Engineers** who've created their own process improvements - **Implementation Consultants** from SaaS companies who configure complex systems for clients - **Solutions Engineers** who work closely with customer success teams - **Former Zendesk/Intercom/Salesforce Admins** who understand ticketing systems deeply - A listing on the [Lorikeet CX+AI jobs board](https://www.lorikeetcx.ai/cx-jobs) The common thread: look for people who are customer-centric, can solve problems naturally gravitate toward improving processes and have self-taught technical skills. ## The bottom line The CX Automation Specialist is one of the most exciting new roles to emerge in customer support. Getting this right isn’t about avoiding mistakes—it’s about putting your team at the front of the curve. The companies who embrace this role early are already seeing what’s possible: higher automation rates, faster scaling, and a better experience for their customers. If you’re experimenting with AI in support, this is your chance to build the muscle before it becomes table stakes. Hire someone who gets energy from helping customers and tinkering with AI tools, and give them the mandate to own it. Your customers will notice. Your support team will thank you. And you’ll be one of the first in your industry with a dedicated automation function, a capability that will only compound in value over time. *Looking to hire a CX Automation Specialist? We've helped dozens of companies think through this role. See here for our sample job description *[*here*](https://www.lorikeetcx.ai/cx-jobs/pdf)*. * --- # The false promise of self-training AI URL: https://www.lorikeetcx.ai/blog/the-false-promise-of-self-training-ai **Self-training AI shifts risk to your business, not the vendor. Learn why leading companies demand transparent, auditable systems—not black-box automation.** "Our AI learns from every interaction!" It's the holy grail promise in customer support AI sales pitches. Set it up once, let it learn from your team's feedback, and watch it get smarter over time. No ongoing training needed. No expert configuration required. Just thumbs up, thumbs down, and AI magic. Unfortunately, it’s a pipedream. Self-training AI systems are a vendor convenience disguised as a customer benefit. They reduce the vendor's support costs because they don’t need to help you configure the system properly. But they shift all the risk to you—the customer whose reputation is on the line with every AI interaction. Here's why self-training systems fail, why explicit instruction-based systems work better, and why this matters especially for businesses with complex compliance requirements. ### "Garbage in, garbage out" becomes inevitable at scale The fundamental problem with self-training systems is that they learn from human feedback. And the feedback humans give is incomplete and inconsistent, especially when giving feedback across a team. Picture this: Your support team has ten agents. Agent A thinks a response is great because it's friendly and resolves the issue. Agent B thinks the same response is poor because it doesn't follow the exact script. Agent C gives it a thumbs up because they're rushing through feedback and it seems "good enough." The AI gets contradictory signals about the same type of response. Over time, it learns... what exactly? A muddled average of inconsistent human judgment. Quality assurance becomes impossible when you can't control what the AI is learning from. You're essentially crowdsourcing your customer experience standards from whoever happens to be giving feedback that day. And when busy, scaling teams provide feedback, there's zero guarantee that inconsistent guidance won't mislead the model in ways you'll never detect. You also lose the ability to transform and improve. Chances are a scaled QA team are dutifully applying the existing QA rubric. Who is asking “how can we do much better” and focusing feedback on that? ### Thumbs up/down systems are fundamentally flawed Binary feedback is nearly useless for training AI systems to handle complex customer interactions. A thumbs up tells you nothing about why something worked. Was it the tone? The accuracy? The speed of resolution? The specific information provided? A self trainedThe AI system effectively has to guess, and it will guess wrong. Worse, agents routinely give thumbs up to responses that are "good enough" but not actually great. They're busy, they want to move on, and the response didn't cause any obvious problems. So the AI learns to optimize for "didn't break anything" rather than "delivered an exceptional experience." The AI draws its own conclusions from this feedback, which may be completely wrong. You end up with a system optimized for mediocrity, not excellence. ### Black box learning eliminates control When an AI system learns on its own, you lose the ability to understand or control what it's actually doing. You can't inspect what the AI learned from the feedback or why it's making specific decisions. When something goes wrong you have no way to trace back to the root cause. Did it learn something incorrect from bad feedback? Is it applying a rule in the wrong context? You have no idea. Debugging becomes impossible. You're reduced to hoping that the next round of feedback somehow fixes mysterious problems you can't even identify. Meanwhile, your customers are experiencing the consequences of these invisible failures in real time. ### Explicit instructions beat "smart" systems Instruction-based systems work because they're transparent and controllable. Instead of hoping the AI figures out what you want from indirect feedback, you tell it exactly how to handle different situations. You specify your escalation criteria, your tone of voice, your policy exceptions, your compliance requirements. Changes are transparent and auditable. You know exactly what you changed and why. When something isn't working, you can pinpoint the specific instruction that needs adjustment. Subject matter experts—people who actually understand your business and customer needs—can review and refine the instructions. You're not outsourcing your customer experience standards to an algorithm's interpretation of thumbs up/down feedback. You maintain control over your CX standards instead of abdicating them to a black box. ### Compliance and risk management require predictability For businesses in regulated industries, self-training systems are a compliance nightmare. Financial services, healthcare, and other regulated businesses need to explain their AI's decision-making to auditors and regulators. "The AI learned from feedback" isn't an acceptable explanation when a regulator asks why a customer was denied service or given incorrect medical information. Self-training systems make compliance audits nearly impossible because you can't document or explain the decision-making process. The AI's reasoning is buried in layers of algorithmic interpretation that even the vendor can't fully explain. Explicit instructions create a clear audit trail. You can show exactly what rules the AI follows and why. Risk management requires knowing exactly how your AI will behave in edge cases, not hoping it learned the right lessons from past feedback. ### The vendor incentive problem Here's the part vendors won't tell you: self-training systems are primarily designed to reduce their operational costs, not improve your results. When you buy a self-training system, the vendor doesn't need to invest in helping you configure it properly. They don't need subject matter experts who understand your business. They don't need to provide ongoing training optimization services. They can just say "let it learn from your feedback" and walk away. It's sold as "advanced AI" but it's actually lazy product development. The promise of "set it and forget it" AI is appealing but unrealistic, especially for complex businesses. Great customer experiences require intentional design and ongoing refinement. Self-training isn't the future of AI—it's a shortcut that puts your customer relationships at risk. Don't let vendors convince you otherwise. --- # Why deflection-focused products make worse AI agents URL: https://www.lorikeetcx.ai/blog/why-deflection-focused-products-make-worse-ai-agents **AI support tools that chase deflection over quality create chatbot traps, not customer trust. Learn why great agents are built on self-awareness—not coverage.** Most AI support vendors optimize for deflection rates. They celebrate getting to "40% automation" by having their AI attempt to answer 100% of tickets, succeed on 40%...and fail on 60%. This isn't just bad for customers—it makes for fundamentally worse products. When you optimize for coverage instead of quality, you end up building chatbots, not agents. The prevalent per-conversation pricing models in the market make this worse: vendors are paid for every ticket they attempt, so they are naturally drawn to attempt to answer more tickets! Recently I wrote why CX teams shouldn’t focus on deflection as a key metric. In this post I’ll explain why similar deflection-focused thinking by vendors leads to worse AI products: - **Product architecture reflects different values** - chatbots maximize engagement, agents know their limits - **Self-awareness is a real technical challenge** - most vendors avoid the hard engineering work - **Bad metrics create bad feedback loops** - you can't improve what you can't measure properly - **Testing tools get built around the wrong goals** - celebrating coverage instead of quality - **Workflow design suffers** - optimizing for engagement over effectiveness ### Product architecture reflects values The difference between a chatbot and an AI agent isn't just marketing. It's a fundamental design philosophy that shapes every part of the product. **Chatbots are designed for maximum engagement.** They attempt to answer every ticket because vendors get paid per conversation, not per resolution. The AI tries to help with everything from password resets to complex billing disputes to medical emergencies. These are called ‘agents’ but chatbot is a better term – it’s like the over-eager kid in class who can’t help but shout out the answer to the teacher’s question, right or wrong. It’s not true intelligence. **AI agents are designed to know their limits.** They understand when they can actually help versus when they should immediately hand off to a human. Self-awareness is a core capability, not an afterthought. This architectural difference creates completely different user experiences. One traps customers in endless loops. The other gets them to the right help quickly. ### The self-awareness problem is technical Making AI "know what it doesn't know" requires solving real technical challenges. Large language models are pre-trained to be helpful – they want to attempt answers even when they lack the knowledge or capability to be useful. We spoke to a company that makes software for doctors. Their current chatbot sometimes gets support tickets from patients looking for their doctor, often in crisis situations. The AI tries to answer these with confused, off-topic responses that could be harmful. Lorikeet's agent instantly recognizes these as outside its scope and escalates immediately. In high-risk environments, this kind of self-awareness isn't just nice to have. It's critical for safety and compliance. Most vendors take the path of least resistance: let the AI try everything and call the failures "learning opportunities." We invested significant engineering effort in building agents that can recognize their own limitations. ### Bad metrics create bad feedback loops When you optimize for deflection, you lose the signal you need to actually improve. If your AI attempts 1,000 tickets and "successfully deflects" 400 of them, what does that tell you? Maybe those 400 were genuinely resolved. Or maybe customers gave up in frustration and found another way to solve their problems. You have no idea which, because you're measuring the wrong thing. Lorikeet's approach is different. We only engage when confident, which gives us much cleaner feedback about what's actually working. When our agent handles a ticket, we know it was equipped to solve that specific type of problem. When it escalates, we know exactly what training gaps to address. This creates a virtuous cycle. Better feedback leads to better training. Better training leads to higher confidence thresholds. Higher confidence thresholds lead to better customer experiences. ### Testing tools built around the wrong goal Products optimized for deflection have weak evaluation frameworks. They celebrate any interaction that doesn't immediately escalate, regardless of whether the customer was actually helped. We built our testing and evaluation tools around quality assessment. Our customers can run hundreds of test conversations to validate that the AI will handle specific scenarios correctly before going live. They can audit every decision the agent makes to ensure it aligns with their brand and policies. You can't build this kind of rigorous evaluation when your goal is just "attempt more tickets." The incentives are all wrong. ### Workflow design matters Deflection-focused products guide customers into AI interactions regardless of complexity. They are more likely to use dark patterns to prevent escalation and keep customers trapped in automated flows. Lorikeet's workflows are designed around a simple question: can the AI actually solve this specific type of problem? If yes, it takes the ticket. If not, it immediately connects the customer with a human who can help. This requires building more sophisticated routing logic and being honest about limitations. But it results in products that customers actually trust and enjoy using. ### The bottom line Optimizing for deflection makes vendors more money in the short term but creates worse products for everyone. It leads to design choices that prioritize engagement over effectiveness. Great AI agents know when not to engage. They're built by teams that care more about solving customer problems than hitting coverage metrics. They use evaluation frameworks that measure quality, not just quantity. If your current AI support tool celebrates "30% automation" while customers complain about being stuck in bot loops, you're dealing with a chatbot built for deflection, not an agent built for results. --- # Keeping your CX up when cloud providers fall down URL: https://www.lorikeetcx.ai/blog/keeping-your-cx-up-when-cloud-providers-fall-down **What recent cloud-provider outages remind us about building AI systems that stay online — even when your LLM provider doesn’t.** Last week, [a major Google Cloud Platform (GCP) outage](https://status.cloud.google.com/incidents/ow5i3PPK96RduMcb1SsW) disrupted services across the globe. Among the affected was Anthropic, whose systems went offline. Thankfully, our customers’ ticket processing remained uninterrupted. Our systems automatically failed over to another provider, ensuring continuous support for our customers (and their customers). This incident underscores a critical point I’ve emphasized before: a key way that application-layer vendors add value is by providing customers with higher reliability than the raw infrastructural building blocks offer alone. This ensures continuity for end customers even as infrastructure continues to struggle to scale under fast-growing loads. Resilience to upstream outages like these is something we have and will continue to invest in as we scale. It's an expensive investment, but one that our customers need and deserve. ### The fragility of single provider dependence Relying solely on a single AI infrastructure provider is akin to putting all your eggs in one basket. Customers don’t see the backend complexities; they see a service that’s suddenly unavailable, leading to long wait times, frustration, and potential loss of trust. At the risk of stating the obvious, AI agents have very different reliability capabilities than human ones. AI agents don’t call in sick or quit at short notice. But they can all go down if a single service fails, while human agents aren’t all going to call in sick on the same day. ### Designing for resilience Everything AI -related is growing so fast right now; the reliability of infrastructure providers and foundational models is impacted e.g. Anthropic has had 99.34% up time over the last 90 days, significantly less than the 99.999% (or '5 9s' in tech lingo) reliability we’ve come to expect from technology providers. That’s why at Lorikeet, we’ve [architected our systems](https://www.lorikeetcx.ai/blog/lorikeet-ai-a-technical-deep-dive) with redundancy at their core. Our AI agents are designed to handle complex, multi-step support requests, and they do so by leveraging a leveraging a multi-provider infrastructure. This means that if one provider experiences issues, our systems seamlessly transition to another, ensuring that our clients’ support operations remain unaffected. Our automated failovers rely on knowing - at the level of each LLM call - what the next best model is. We do this based on a robust abstraction framework and set of evals. We've made this investment because we're acutely aware of the trust our customers put in us, and need to ensure we honor it, instead of relying on an easy out like "Anthropic went down". ### The broader implications GCP wasn't on its own. In the last thirty days alone, we’ve seen outages from: - Cloudflare - OpenAI - IBM Cloud - Microsoft Azure - Pinecone - LangChain If you're building your own solution, you will need to ensure it's robust against future outages like these, further increasing the cost of building versus buying. ### Moving forward As we continue to build out the Lorikeet platform, we won’t just focus on capabilities. We’ll maintain our deep investment in reliability. At the end of the day, our AI agents, no matter how advanced, are only as effective as the infrastructure supporting them. --- # Great CX isn't built on deflection rates URL: https://www.lorikeetcx.ai/blog/great-cx-isnt-built-on-deflection-rates **High deflection doesn't equal good CX. Learn why customer support teams should focus on resolution quality and efficiency—not just keeping humans out of the loop.** Deflection rate measures how often a customer inquiry is handled without involving a human agent. It sounds reasonable until you realize what it actually incentivizes. When you can achieve 100% deflection simply by turning off all your support channels, the metric is clearly disconnected from what matters to your customers. It transforms an internal "how" into the core "what." Does that mean you should stop measuring deflection? No. You just need to redefine the role deflection plays in improving the overall efficiency of your support operation. In our experience, we recommend you: - Distinguish good deflection versus bad deflection. - Focus on efficiency metrics designed for an AI-first world. - Optimize the quality of interactions first, then scale coverage. - Make sure you're not paying for bad deflections. ### The deflection trap I've seen companies celebrate 70% deflection rates while their [CSAT](https://www.lorikeetcx.ai/blog/csat-is-dead-long-live-csat) plummets. Why? Because deflection doesn't distinguish between "customer got their problem solved" and "customer gave up trying to get help." A chatbot that confuses customers into abandoning their requests scores the same as an AI agent that genuinely resolves issues. Both register as "deflected" in the metrics. This creates perverse incentives. Vendors get paid whether the AI helps or not, so they optimize for engagement over outcomes. The result? AI systems that trap customers in endless loops rather than solving problems or escalating appropriately. ### Good deflection vs. bad deflection Let's say a customer contacts support about a billing issue. Here's what a bad versus a good deflection looks like. Both count as "deflection." Only one creates a good experience. Since customers don’t consistently fill out CSAT surveys, it’s more likely than not that this bad experience never shows up in any metric. ### Focus on what actually matters The real goal isn't deflection – it's efficiency. Deflection is – at best – a “how” to achieve those goals. Here's what you actually want to measure: AI should make your support operation more efficient by handling appropriate cases well, freeing humans for complex work that requires judgment. The "how" is less important than the outcome. For example, [Magic Eden achieved a 74% CSAT](https://www.lorikeetcx.ai/case-studies/how-nft-marketplace-magic-eden-use-lorikeet) with our AI agent – 30 points higher than their previous solution. But we didn't get there by maximizing deflection. We got there by ensuring every AI interaction was high quality. Indeed given some complex debugging issues they faced, the best solution was for the AI agent to collect information upfront, allowing a fast and efficient resolution once a human agent picked up the ticket. This kind of solution improves CSAT and efficiency, but doesn’t count as “deflection”. ### Self-aware AI beats coverage-obsessed AI The best AI agents know what they don't know. They understand their limits and get out of the way when they can't help. Most vendors optimize for the wrong thing: they want AI to engage with 100% of tickets to maximize "deflection." But this means customers have multiple failed AI interactions for every successful one. The better approach is to train AI to only engage when it can do a good job. Maybe it handles 50% of conversations, but solves 80% of the ones it touches. This massively reduces failed AI interactions, which are what actually annoy customers. We've seen this play out across our customer base. Companies that focus on interaction quality over coverage see better business outcomes. Their customers trust the AI more, escalations are cleaner, and human agents aren't buried under tickets the AI fumbled. ### The vendor incentive problem Here's why this matters: most AI support vendors charge per ticket or their self-determined measure of a "resolution." Put another way, they get paid whether they help your customers or not, and they're incentivized to increase the overall number of tickets their AI handles, at the expense of your customers and your bottom line. Good news: we've created a [free CX ROI calculator](https://www.cxroi.ai/) to help you calculate the real cost of your AI solution. So when a vendor tells you their AI "deflects 80% of tickets," ask the follow-up question: "How many of those deflected customers actually had their problems solved?" You'll be surprised how often they can't answer that question or how uncomfortable they get when you ask it. ### Bottom line Great customer experience comes from solving problems, not from keeping customers away from support channels. [The companies we work with](https://www.lorikeetcx.ai/case-studies) pride themselves on providing excellent support. They want to use AI to scale that excellence, not to create barriers between themselves and their customers. Optimize for customer outcomes, and business metrics will follow. Optimize for deflection, and you'll train customers to hate your AI – and your company. There's no substitute for quality. Don't let vendors convince you otherwise. --- # CSAT is dead. Long live CSAT. URL: https://www.lorikeetcx.ai/blog/csat-is-dead-long-live-csat **CSAT is flawed, but in ways we understand. Vendor-invented metrics are flawed in ways we can't see or audit. Complement CSAT with your own comprehensive evaluation framework instead.** Intercom just launched their "Customer Experience Score" (CX Score) with a [blog post](https://fin.ai/ideas/why-an-ai-driven-customer-experience-score-will-replace-human-surveyed-csat/) arguing that CSAT should be replaced by AI-driven evaluation. They make some valid points about CSAT's problems, but they're essentially asking you to replace a flawed industry standard with their proprietary, Intercom-specific score. Here's why we think that's a bad idea. ### CSAT does have problems Let's be honest about CSAT's limitations first. Some of Intercom's criticisms aren't wrong – two in particular. - **Response bias is real.** People with extreme experiences are more likely to respond (particularly if you only send out CSAT surveys after customers contact support). Your "average" CSAT score definitely isn't representative of your actual average customer experience. - **CSAT is shallow.** It captures sentiment at one moment but doesn't explain why or predict future behavior. It also muddles “policy CSAT” (customer happy / not happy with policy) from “experience CSAT” (customer happy / not happy with how the support agent behaved). These are legitimate problems. But Intercom's solution has the potential to create bigger ones. ### The problem with vendor-invented metrics When a vendor creates their own evaluation metric, they're grading their own homework. Intercom's CX Score conveniently shows their AI agent performing better than human agents. I’m sure that is really what they are seeing in their data, the challenge is it requires an ongoing heroic level of self awareness and self discipline from them to avoid slipping into finding the convenient conclusion from the data. While Intercom claims 0.8 F-score validation, after experienced agents reviewed its AI-scoring model, customers can't independently verify or audit the scoring. We've seen this movie before – businesses creating internal metrics that unintentionally overstated their success, causing them to overlook real product or service issues. Uber initially used "completed trips" as their primary success metric, which encouraged rapid growth while overlooking driver satisfaction, safety, and service quality. Netflix counted a "view" as watching just two minutes of content, inflating their popularity metrics and distorting content decisions. Once you adopt a vendor's custom metric, you're also locked into their evaluation framework. You can’t meaningfully compare the new custom metric to industry standards like CSAT, which makes it harder to compare Intercom’s performance against other vendors. ### CSAT reflects reality Despite its flaws, CSAT has one critical advantage: it reflects what customers actually think, not what an AI model thinks they should think. Yes, CSAT coverage is low and biased. But the customers who do respond are giving you their genuine reaction. That's more valuable than a black box algorithm's assessment of what their reaction should be. CSAT is also transparent and auditable. Everyone understands what it measures, even if imperfectly. You can benchmark across vendors and time periods. You can independently verify results. ### Don't replace CSAT, complement it The right approach isn't to throw out CSAT for a vendor's proprietary metric. It's to use multiple evaluation methods. At Lorikeet, we let customers define their own quality metrics and test them rigorously. You can use CSAT alongside other measures like resolution time, escalation rates, and customer-defined quality criteria. Test everything against human judgment. Build evaluation frameworks you control and understand. Use metrics you can port between vendors. There is no substitute for “tasting the soup” and reviewing tickets by hand to ask “is this how we want to show up for our customers”. Most importantly, don't outsource your quality standards to a vendor. They have different incentives than you do. ### The bottom line CSAT is flawed, but it's flawed in ways we understand. Vendor-invented metrics are flawed in ways we can't see or audit. Better to have an imperfect metric you control than a "perfect" one you can't verify. Keep using CSAT, but don't rely on it alone. Build your own comprehensive evaluation framework. Your customers' actual opinions matter more than any algorithm's assessment of what their opinions should be. --- # Why customers reject AI (and how to fix it) URL: https://www.lorikeetcx.ai/blog/why-customers-reject-ai-and-how-to-fix-it **When customers say “human” they mean “better”. Here’s what really drives “refusal” rates—and how smart prompting and better credentialing can get your customers to actually give AI a chance.** We spoke to a company rolling out voice AI (not Lorikeet's) recently. They were facing a 50% “refusal” rate – in other words, customers were asking to speak to a human as soon as the AI got on the line, before it had a chance to help (or not help). When half your customers are refusing to engage, it’s hard to achieve meaningful gains with AI. Based on our conversations in the market, anxiety about refusal is rampant, and with good reason. After years of frustrating experiences with ineffective chatbots that simply weren’t good enough (and in most cases, just flat-out dumb), many customers will just demand to talk to a human. Indeed shitty AI systems have very effectively trained savvy customers to find ways to evade them. But high-quality, next-generation AI systems offer customers a human-quality or better experience if the customers just give the AI a chance. We’ve learned a lot about how AI agents can earn that chance, and we want to share it here. In short, there are three critical tactics: - Prompt the customer - Credential the agent - Make humans available – but encourage customers to try AI first ### The refusal myth “OPERATOR!” We’ve all said it. When we ask to speak to a human, we’re not rejecting AI in principle – we’re rejecting bad experiences. We’ve been trained that saying "agent" or "representative" is our escape hatch from the maddening loop of a system that doesn't understand us. But when customers say “human,” they mean “better,”. But if AI can handle complex queries while demonstrating and maintaining human-quality conversation, this resistance disappears. We've seen this firsthand across industries, including in high-stakes contexts like healthcare and financial services. Good AI systems don’t just satisfy, they delight. ### How to prevent refusal We’ve found that refusal is something that’s actually very manageable. Here’s what we’ve learned from implementing Lorikeet’s AI agents for our customers across health tech, fintech, and other complex industries. #### 1. Prompt the customer As AI agents start to proliferate, especially voice agents, it's important to provide a preamble to set your customers’ expectations: > "I'm an intelligent assistant for [company]. You can talk to me the way you would any other support agent, and I can help you just like they would. If we get stuck, I’ll transfer you to one of my human colleagues to help." This approach aims to tell users how to interact with the AI and reassure them that humans are available if needed. This is particularly important because many customers, once they’re told they’re talking to AI, will speak to it (verbally or in writing) like old NLU-based systems. They might say “account issue” or “card problem.” NLU systems could only handle simple inputs like this, but modern LLMs actually struggle with them because they lack substance. Lorikeet’s agent proactively asks for more detail in these cases, but it’s more efficient to avoid that by prompting the human. #### 2. Credential the agent While prompting the human is useful, it’s even better to show – not tell – by enabling the agent to implicitly credential itself early in the interaction. We’ve found that if an agent can establish in the customer’s eyes that it’s looking at their actual account and is able to solve their problems, refusal rates drop to almost zero. This can be as simple as using a customer’s name, but it can go much further. By way of illustration, one of our crypto customers gets lots of inquiries into the status of transfers off their platform and into customer bank accounts. In reality, once the transfer is sent, the crypto company doesn’t know the status of the funds – the customer needs to contact their bank. When they first deployed Lorikeet’s agent, it told customers this, and very reliably they would ask for a human. The answer seemed generic, and so customers wanted to see if they’d get a better answer from a person. We drove refusal to almost zero by having Lorikeet’s agent first confirm which transfer the customer was talking about: “Do you mean the transfer for $X,XXX you initiated yesterday to your Chase account ending YYZZ?”. Once the customer confirms, Lorikeet’s agent delivers the same message: we don’t know the status, you need to contact your bank. But with the additional “credentialing” step, refusal went away because of the implicit message that customers were getting a specific, personalized response. This can extend to how the customer engages with the agent. Customers shouldn't need to take any special actions or click vague buttons. If the agent shows them a list of transactions including one at a candle store and they say "I didn't buy any candles", the AI agent should understand that means potential fraud, not force them into giving a structured response. This feels like talking to a competent human who gets it, not a rigid system that needs to be navigated. ### 3. Make humans available – but encourage customers to try AI first If a customer's issue is beyond what the AI can handle, or if they're clearly frustrated, a smooth handoff builds confidence in the entire support system. The very best AI agents are self-aware. They know what they don’t know. They understand their limits. This self-awareness enables them to determine when to escalate to a human agent versus trapping the customers in an endless loop. For those customers who immediately ask for a human without giving the AI a chance, we've found success with: > "I'll get my colleague, but can you give me a shot at solving the problem first?" This strategy encourages users to try the AI before being handed off to a human, and often results in successful resolution without escalation. ## The bigger picture The biggest mistake we’ve seen companies make is choosing a vendor whose AI agents are optimized for “AI engagement” instead of your customers’ success. This creates a dangerous incentive: they celebrate preventing customers from talking to humans, even when the experience is poor or when that’s what clearly needs to happen – AKA they measure their success on deflection rates. From speaking to hundreds of companies that have deployed AI agents for customer support, we can tell you this approach will backfire. It only trains customers to distrust and resist your AI at every turn. And hate your company in the process. Don’t waste your time. Instead, focus on making your AI genuinely good at solving problems. When it can handle complex cases like lost cards, fraud investigations, and ordering medical prescriptions while maintaining human-quality conversation, your customers won’t refuse AI; they’ll embrace it. --- # Why we’ve raised another $11m URL: https://www.lorikeetcx.ai/blog/why-weve-raised-another-11m-in-funding **After seeing explosive growth following our seed round last October, we are stoked and fortunate to have raised another USD $11m in funding led by Tom Humphrey at Blackbird with participation from our existing investors Square Peg and Skip Capital.** After seeing explosive growth following [our seed round last October](https://www.forbes.com.au/news/entrepreneurs/ex-google-stripe-execs-bag-5-million-for-ai-start-up-lorikeet/), we are stoked and fortunate to have raised another USD $11m in funding led by [Tom Humphrey](https://www.linkedin.com/in/tomhumphrey1/) at [Blackbird](https://www.linkedin.com/company/blackbirdvc/) with participation from our existing investors [Square Peg](https://www.linkedin.com/company/square-peg-capital/) and [Skip Capital](https://www.linkedin.com/company/skip-capital/). > *Lorikeet builds AI-powered agents used to perform customer service tasks by high-profile online health start-up Eucalyptus, crypto gaming platform Immutable, and US financial services company Step. * [*Australian start-up Lorikeet cracks $100m valuation in AI agent boom*](https://www.afr.com/technology/australian-start-up-lorikeet-cracks-100m-valuation-in-ai-agent-boom-20250203-p5l934)*, AFR* ### Continuing to grow a global, stellar team After [winning more than a handful of big deals](https://open.spotify.com/episode/4caDFFs15WTqWYjCAPyCX5) against some of the giants in the market, [Jamie](https://www.linkedin.com/in/jamie-hall-30360954/) and I realized we needed to grow the team rapidly to keep up with growing demand for Lorikeet. We're using most of the funds to grow our engineering teams. Yes there's opportunity to sell, and [we're hiring](https://lorikeetcx.ai/careers) there too, but the biggest opportunity is to make our products even more powerful for our customers. ### Delivering quality CX for customers with higher standards Our customers continue to see incredible results. With careful use of [Lorikeet](https://lorikeetcx.ai) alongside their talented support and clinical teams, while support volume doubled in the last year, Eucalyptus have not had to add *any* headcount. They've done this [while improving CSAT 10ppts](https://www.lorikeetcx.ai/case-studies/how-eucalyptus-boosted-their-csat-by-10ppts-using-lorikeet-triage) and ensuring every patient gets a response within 90 seconds. Talk about a win/win. And then there’s Magic Eden, who replaced their old AI agent from a noisy and highly funded competitor with Lorikeet. Within the first month of rolling out, [they achieved a CSAT of 74%](https://www.lorikeetcx.ai/case-studies/how-nft-marketplace-magic-eden-use-lorikeet): approximately 30 percentage points higher than . ### Faster, safer, smarter deployment (across more channels) So what are we building next? We don’t share all our secrets! What I will say is that now we have a product that’s delivering real results for our customers, we’re keen to streamline the agent training experience to enable faster and (even) safer agent deployments across chat, email, *and* voice. 🤫 I’ll leave you with this… The main determinant of the quality of the support experience is the intent of the company offering it, not the tools they use. Good support can be provided by both AI and humans, and bad support can occur with both as well. Lorikeet is not a solution for great CX; it’s a platform that enables you to achieve it. If that sounds good to you, let’s talk. P.S. [We’re hiring](https://jobs.ashbyhq.com/Lorikeet) --- # In support, copilots are not the answer URL: https://www.lorikeetcx.ai/blog/in-support-copilots-are-not-the-answer **Customer support copilots offer marginal gains and false safety. Learn why autonomous AI agents—not copilots—are the real solution for scalable support.** AI "Copilots" are not the answer for customer support. They're an admission of defeat by unambitious vendors and cautious, under-informed buyers. The logical case for them is weak, and so are the results they produce. Companies seeking to buy copilots today should reassess. This post will be a little longer than usual because this topic is complex and I want to take the time to lay out my thinking. Here’s the basic thesis: - A capable copilot is a capable pilot - A capable copilot offers only marginal efficiency gains - An *incapable* copilot may be worse than nothing - As a result, using a copilot as a stepping stone to full autonomy isn’t effective - Vendors are selling copilots because buyers are understandably afraid of AI - The better solution is high quality AI agents with robust testing I’ll walk through and unpack these points. ### A capable copilot is a capable pilot When solving support tickets, the requirements of a capable copilot system are indistinguishable from a pilot system that works autonomously. It has to understand the customer’s intent, map that intent to reference material or standard operating procedures, then gather information and execute a response. But it’s possible the benefit of a copilot is human review reduces the risk of errors, meaning you can get the efficiency benefits without accuracy risks. That doesn’t stack up. ### A capable copilot offers only marginal efficiency gains For the sake of argument, let’s assume for a moment that a copilot always produces good quality output. Even then, it has two fatal drawbacks. First, because copilots still require human operators to handle every ticket, any improvements can only be marginal. Erik Brynjolfsson from MIT and co-authors demonstrated this in a [high quality paper](https://arxiv.org/abs/2304.11771) released last year, finding a customer support copilot system improved tickets handled per hour by a mere 14% on average. So for a business growing >70% year on year, implementing a copilot offsets one quarter of headcount growth. For a business growing 30% a year, it buys six months. Further, it doesn’t address the other element of scaling: responsiveness to unexpected ticket volumes. A team that’s 14% more efficient will still get swamped by an unexpected surge in ticket volume. We’ll acknowledge that, for a very large business growing (say) 2% year on year, a 14% efficiency improvement can be useful. But most businesses aspire to grow a lot faster than 2% per year, and should aspire to get more than 14% out of AI. Finally, it’s worth considering the human toll of the copilot success case. We’d argue the better the copilot is, the more the job becomes alienating: human agents are reduced to nothing more than rubber stamps. If you thought today’s customer support jobs could be mindless and alienating, imagine sitting like a battery hen clicking “send” on AI-drafted responses all day. ### An incapable copilot may be worse than nothing We’ve been assuming that the copilot is always good. But we know it likely isn’t - otherwise it could be operating independently. In the more realistic case, where the copilot is correct and incorrect in some hard-to-estimate proportion, the situation gets much harder. Now the assumption of copilot efficiency rests on the idea that humans will correctly identify when the AI is wrong, and that they’ll find fixing its errors (and rubber stamping its wins) faster than just writing responses themselves. There is good reason to be skeptical of this. First, when AI systems are wrong, they’re often overconfident and wrong. They output plausible sounding, confident and incorrect messages. Agents will not quickly and easily be able to see through the AI BS, and they won’t judge AI quality flawlessly. ‍So the true rate of AI assistance will be the portion of tickets the copilot gets right *multiplied by * the portion of those tickets the humans correctly assess as correct. So a 70% accurate model that’s correctly assessed 70% of the time is only leading to correct responses 49% of the time. The cherry on the sundae is that the human review blunts the incentive of vendors *or *buyers to get the AI’s accuracy up. It reinforces the common support antipattern of throwing people at a problem instead of landing the right technology solution. ### As a result, using a copilot as a stepping stone to full autonomy isn’t effective Many companies are interested in copilots as a way to tiptoe toward deploying autonomous AI agents. But we think embracing copilots is a side quest, not a first step on the autonomy path. As above, they select for different system and vendor capabilities, and different human <> AI interaction models. Our prediction is that the copilot moment in customer support won’t last. The copilots will – at best – generate marginal improvements. The companies that embrace agents will see step changes. So it’s only a matter of time – perhaps not much time – until companies that embrace copilots move away from them. As a company we don’t believe that doing the easy thing – making copilots – is preparation for doing the hard thing – making autonomous agents. So in theory a copilot company will have a leg up in transitioning their users to agents, but in practice the original copilot sale should rightly erode trust, and will distract the companies from building autonomous agents. ### Vendors are selling copilots because buyers are understandably afraid of AI Despite these logical and empirical issues, lots of vendors – even newer startups – are selling copilot solutions. They’re doing it because buyers demand it. The majority of companies we talk to ask about a copilot on the first call. This demand is not unreasonable. Copilots superficially seem like a safe way to start experimenting with AI, and are a lot easier to contemplate and introduce to an existing human team. But one of the challenges in B2B software is selling *value* to the *corporate customer *while still satisfying the needs and goals of the *buying user*. Since copilots won’t deliver real value, we’ve made the choice to invest in educating buyers about alternate approaches, instead of selling them what they’re asking for. This matters because it dictates the amount of improvement they’ll be able to generate and show to their leadership when renewal rolls around. ### The better solution is high quality AI agents with robust testing Okay, so my choice is to use a copilot which doesn’t ultimately help, or take a huge risk on an autonomous AI system? No. The solution to the dilemma is to embrace autonomous agents *incrementally *and with effective testing and evaluation. Lorikeet’s AI agent does not aim to draft responses for or to send responses to all tickets. Instead it’s designed to be trained to deal with specific issues, and to leave everything beyond its training alone. This allows our customers to train and test the AI incrementally, building confidence over time but driving step change improvements in efficiency as the agent takes on more and more volume autonomously. By focusing on how we help customers launch autonomous agents, we’ve been drawn to make the agent better, *and* to build an industry-leading suite of testing, evaluation and rollout tools. At the same time, deploying autonomous agents provides a much more interesting and high leverage role for the humans on the support team: configuring, deploying, testing and iterating on these AI agents. --- # Lorikeet secures $5m in funding URL: https://www.lorikeetcx.ai/blog/lorikeet-secures-5-million-in-funding-to-empower-cx-teams-with-first-ai-agent-that-offers-human-quality-support-at-scale **Former Stripe, Google AI execs launch first AI platform capable of resolving complex support questions reserved solely for the most skilled human support agents** AI customer experience start-up [Lorikeet](https://www.lorikeetcx.ai) today unveiled its groundbreaking AI customer support platform, powered by a breakthrough, one-of-a-kind AI architecture able to resolve even the most complex customer inquiries and far surpass what’s possible with leading support models or chatbot approaches. The company has seen strong market momentum with the addition of new customers, hundreds of thousands of dollars in annualized revenue, and $5 million in seed funding from Square Peg Capital and leading investors to support product development and fuel global expansion. Customer support as we know it is broken: customers have to choose between clunky preconfigured chatbots or underpowered, hard to trust generative AI agents. Lorikeet, founded by product and AI leaders previously at Google and Stripe, takes a different approach with AI agents powerful and trustworthy enough to handle complex, high stakes interactions. Leading fintech and healthtech companies in the U.S., Australia and Asia like Remote.com, Step, Eucalyptus, Stashaway, and SensorFlow use Lorikeet to solve complex customer problems 24/7. "We’ve focused first on solving the hardest support tickets, because that’s where we can create the most value for our users,” said Steve Hind, former Stripe executive and Lorikeet co-founder and CEO. “Most AI agents focus on writing simple ‘Q&A’ style responses to customers’ questions. We think the hard part of support is executing the workflows that actually *solve* customers’ problems, so we focused on that. Our agent works just like your best human agent - and unlike any other AI agent. This approach allows us to work in highly regulated industries like financial services and healthcare, where accuracy and compliance are non-negotiable." Lorikeet users see much faster response and resolution times, improved CSAT, smoother scaling - and notable cost savings: - *“We're delivering faster, more efficient support to our patients like never before, with no compromise to patient experience. Improving the first response time from 20 hours to 90 seconds has been transformational,” said Omar Alvi, Managing Director, Pilot & Juniper Australia, Eucalyptus.* - *“Lorikeet’s unique triaging capabilities mean urgent medical questions are escalated to clinicians and dealt with quickly. This gives us the confidence to scale our healthcare offering safely. The uplift in patient satisfaction is icing on the cake,” said Estelle Berton, Operations Lead, Eucalyptus.* - *“We were especially impressed by the way Lorikeet integrates with the tools we already use, saving us valuable time and effort. It’s been a total game-changer for our customer experience — we’ve set ourselves up to scale support without scaling headcount,” Millie Yang, CEO of crypto payment leader Breeze. “Thanks to Lorikeet, we’ve been able to provide fast, reliable support even as our customer base grows.”* ### A new AI framework for customer support Lorikeet offers the first AI framework specifically designed to enable AI agents to solve complex tickets normally resolved for the most skilled humans. Rather than relying solely on retrieval-augmented generation (RAG) models, Lorikeet's architecture goes deeper, with the AI agent able to refer to and follow the same standard operating procedures that top support agents follow. This allows the AI to not just retrieve information, but to actively solve complex customer support problems. "Our AI doesn't just retrieve information, or make up processes on the fly. It reliably follows the right business logic while also keeping the conversation natural," explained Dr Jamie Hall, Lorikeet co-founder and former Google Brain research tech lead. "If you can explain to a human agent how to resolve a ticket, our AI can handle it. It follows exact business logic and asks the right questions to solve customer concerns 24 hours a day.” This unique AI architecture enables Lorikeet to tackle intricate issues that stump conventional chatbots, from handling delayed orders to replacing compromised credit cards. It seamlessly integrates with existing company workflows, offering a natural conversational experience while maintaining strict adherence to business rules.‍ “​​Consumer facing companies spend 10-15% of their top line on support. And every year, CFOs and Heads of support orgs get pushed to reduce that budget. Lorikeet’s brought together a world class team to tackle these complex problems, using a unique approach to AI to improve the customer experience while improving the bottom line,” said Piruze Sabuncu, Partner at Square Peg, previously Stripe’s first employee in APAC. The additional funding will accelerate Lorikeet's global expansion, bringing its innovative AI-powered customer service to more organizations worldwide. This growth aims to meet the surging demand for intelligent, efficient, and accurate customer support solutions across various industries and geographies. With Lorikeet, companies can now scale exceptional customer experiences without sacrificing quality or causing frustration. It's not just another AI chatbot – it's a sophisticated problem-solving engine that replicates the best support person for every customer interaction.‍ ### About us Lorikeet (previously known as Optech) provides the world's most advanced AI customer support agent, powered by a unique AI architecture that solves complex issues while maintaining high customer satisfaction. [Steve Hind](https://www.linkedin.com/in/shind/) (ex-Stripe product lead) and [Jamie Hall](https://www.linkedin.com/in/jamie-hall-30360954/) started Lorikeet to make generative AI deployable at scale to improve customer experience. Lorikeet has raised over USD $4.7M in funding and is backed by Square Peg; leading angel investors and executives from Atlassian, Stripe, Linear, Brex, OpenAI, Rippling, and more. Customers include Remote.com, Eucalyptus, Step, Stashaway, Chemist2U, and Sensorflow. To learn more, visit[ www.lorikeetcx.ai](http://www.lorikeetcx.ai). --- # Rethinking AI for customer support: A technical deep dive into Lorikeet URL: https://www.lorikeetcx.ai/blog/lorikeet-ai-a-technical-deep-dive **Explore Lorikeet’s unique AI system for customer support. Learn why workflows—not copilots or RAG—deliver safe, scalable, and accurate ticket resolution.** Customer support today is archaic. A decade ago, companies like Zendesk and Intercom automated much of the support process with chatbot web integration; however, today these companies are limited in their ability to do more than information retrieval. They are not able to actually resolve them.‍ Unlike other AI tools that mostly deflect tickets or rely on basic Q&A, Lorikeet mimics your best human agents by following the exact workflows they use. This allows it to resolve complex issues with precision, not just respond to simple queries. In turn, that unblocks companies to provide high-quality, scalable support without the need for constant hiring. It allows businesses to maintain or even improve customer satisfaction as they grow, freeing up human agents to focus on more strategic tasks. Because we actively follow the same standard operating procedures that top human agents, we’re uniquely able to handle complex tasks like resolving delayed orders or replacing compromised credit cards - tasks that other AI agents would typically struggle with or escalate to humans. **In this piece, I delve a bit deeper into the technical framework powering it all.** Customer support teams who work in complex or regulated spaces like Fintech and Healthtech run their teams on standard operating procedures or workflows that describe to an agent how to solve a customer’s problem. In these complex spaces, FAQs or help center articles aren’t sufficient to serve customers: the answer is usually conditional on the state of some entity (a transaction, a shipment) and the state of the customer (are they KYC approved? Are they on the premium tier?). To be an effective member of a support team for a company like this, a support agent needs to be able to understand and reliably execute these workflows. Lorikeet’s AI agent is based on being able to execute and understand *workflows*. Workflows encapsulate the business logic that human support teams use, in a form that becomes tractable for an AI system to handle. This makes it possible to resolve inbound support tasks in the way the rest of the human team does, and in a way that leaves customers delighted. This is different from the two other approaches we see in the market: either pure retrieval-augmented FAQ answering (“RAG”), or “agentic” models that involve the AI building its own execution path at runtime. We don’t think either system is fit for purpose for complex support: RAG can’t handle conditional and stateful logic; agentic reasoning attempts to handle it, but is stochastic when it should be deterministic (just imagine if a human team member made up a new way to solve the same type of ticket each time!).‍ We’re proud of the unique system we’ve built, and our unique Intelligent Graph architecture. This post goes into more depth about how it works. If you’d like to learn more, please book a demo! ### What if the AI makes things up? First things first. Anyone who’s played with a modern AI system will be aware of hallucination, where the model generates text that looks plausible and sounds confident, but is in fact wrong. This is partly a side-effect of the way that foundational LLMs are finetuned. If you input the sentence “When was Ludwig Wittgenstein born?” into a general language model (that is, a system which simply estimates the probability of the next word, conditional on everything that’s come before), it would be reasonable to put a large weight on responses like “Sorry, I have no idea” or “I don’t care.” But modern LLMs are meant to be helpful sources of information, so they are finetuned to [skew](https://arxiv.org/abs/2405.16908) their answers towards interesting and potentially useful answers. Unfortunately this means that they’re prone to be confidently wrong, if they’re left to their own devices. For customer support, compared to the general case, the likelihood of hallucination is higher and the risks are greater. It’s more likely because foundational LLMs gravitate towards being helpful, proactive, and upbeat. If you frame the conversation so that the language model takes on the persona of a customer support agent, it will tend to offer expedited shipping or promise discounts or claim that it’s authorized a free upgrade, even if those actions are against policy or aren’t even possible, because that’s the kind of thing that a helpful customer support agent would generally say. And the risks are higher here because the bot is representing you in an interaction with one of your actual customers. This has actually happened! Air Canada was [famously required to honor](https://www.bbc.com/travel/article/20240222-air-canada-chatbot-misinformation-what-travellers-should-know) a cancellation fee waiver that its AI chatbot hallucinated.‍ We mitigate hallucination by only ever giving the LLM a specific and focused task to do. Instead of expecting it to manage an entire conversation (as in the agentic approach), it is only ever passing on specific talking points defined at a certain point within a workflow. It sees all and only the information it needs to do that single job. General dialog systems like ChatGPT or Claude can discuss anything, so their developers need to handle questions about all human knowledge. Solving a specific and domain-focused task is much more reliable. Incidentally, this design pattern also helps to defend against prompt injection and data leakage. We execute a workflow by composing small-sized LLM tasks and individual pieces of procedural logic. Even when sensitive data is required to resolve a ticket, we only need to expose that data to an LLM (and therefore to leakage) if the AI strictly requires it. If one of the LLMs is told to forget its previous instructions and compose a sea shanty, the blast radius of any unintended behavior is small, me hearties. ### So is this a phone tree? Does it require custom engineering to launch? The previous generation of support chatbots used predefined business logic to handle every conversation. Readers of a certain age may also remember the era of ‘Expert Systems’, where general AI agents were restricted to decision trees. Similarly, you may be aware of well funded competitors who quote a 6-12 month implementation period in which they essentially custom -engineer your workflow logic. While superficially similar on one level to these deterministic expert systems, our platform is in fact very different, in both how the decision trees operate and how they’re used. We think the architecture that enables this is a new innovation in how LLMs are used by companies and we call it Intelligent Graph orchestration. On the operational level, we leverage the new capabilities of contemporary language models. Older systems could only handle decisions like “Did the customer’s message contain the word ‘booking’?” because it was too hard to parse natural language in any greater depth. This is why chatting with one of them felt like talking to a phone tree. Because modern LLMs can handle more human-like reasoning when making decisions — “Does the customer sound angry?” or “Have they mentioned any reason why they can’t be home for a delivery?” — our system handles conversations more smoothly. We also use the decision trees differently. They are managed by what’s technically called our ‘orchestration layer’, which keeps track of the conversation and marshals language models when needed to do specific tasks. In that way, our system design is inverted from the usual pattern with AI systems, where a single LLM is in charge. We arrived at this pattern by working with customer support and ops teams to figure out ways to solve their problems at scale, rather than by starting with a cool piece of tech and looking around for applications of it. ### Does the AI drive everything? Should it? One option for automating customer support is the ‘agentic’ approach: you give an LLM a list of policy instructions, a list of APIs it can call, and let it handle each conversation however it likes. This design strategy works great for simple demos, but we are skeptical about how well it can handle real production workloads in industries like fintech and healthtech. The crux of the issue is that an LLM’s ability to follow policy directives [decays badly](https://arxiv.org/pdf/2406.04520) after about half a dozen instructions. This makes sense when you consider how LLM training data is typically created in 2024; it is straightforward to synthetically generate an instruction-prompting example with a simple constraint, but difficult to assemble high quality examples of complicated instructions at scale. As a result, agentic LLMs get distracted and make mistakes. ### What about GPT5, Gemini 2, Claude 4? We can make an educated guess that the next generation of LLMs will be more reliable and have deeper reasoning abilities than the best of the current cohort. However, our thesis is that they should generate and supervise customer support workflows, not make them up on the fly. This makes the workflows possible to debug and monitor at scale, a crucial part of regulatory compliance and effective CX management. (Recall the example of a new member of a bank’s customer service team deciding to handle each ticket by making up a plan on the spot and then executing it. Oof!) ### Do we need a support copilot? It’s possible to plug in AI products which draft responses for human agents to use, but which don’t interact with customers. While this looks at first glance like a productivity boost, it’s not. Either the responses are high quality, in which case it should just send them, or they’re low quality, in which case they’re useless. We suspect this leads to two things: 1) if the copilot is good, having humans just review and hit Send on AI responses is even more dehumanizing and alienating than current support work, and 2) since the copilot will not be good, having humans have to constantly detect AI hallucinations and errors is probably more effort than having the humans just draft the answers. Our general contention is that AI systems should entirely resolve any support tickets that they know how to deal with, and hand off the complex edge cases to humans. ### Do you train on our data? No. We do not finetune models on your data, and do not pass it to any third-party systems which might use it to train their own models. Your customer data, proprietary knowledge, and ticket histories remain your own. ### Let’s go! To learn more about how Lorikeet AI can transform your business, book a demo now. --- # What happens when you put your customers in the same room as your engineers URL: https://www.lorikeetcx.ai/blog/what-happens-when-you-put-your-customers-in-the-same-room-as-your-engineers **Three customers told Lorikeet's engineers what AI support is actually like to use. The roadmap shifted before anyone left the room.** Setting the scene. Mid-March, Honolulu. Seventy people are in a room - engineers, product managers, designers, the sales team. At the front, three customers. Not on a stage at a user conference with rehearsed talking points. Just sitting in chairs, with a moderator, telling a room full of the people who build the product what it's actually like to use it every day. Joel runs customer support at Mosh, a digital health company in Australia. Jess leads CX at Arbor, an energy company navigating a deregulated market where most consumers don't even know how to read a utility bill. The third panelist leads CX at a US telehealth company serving millions of patients. They've been using Lorikeet for different lengths of time, in different industries, with different problems. They did not coordinate their answers. ## Thank you, Moe Joel's "wow moment" was specific. A patient - in healthcare, where people are dealing with sensitive medical issues and are often at their worst - thanked the AI agent for the service it provided. Not a five-star CSAT survey. An actual thank you, unprompted, mid-conversation. "Especially in the healthcare space, people can be at their worst and not very nice to deal with," Joel said. "So thanking an AI bot - or not even realizing it's an AI bot - for the service provided, that's pretty huge." The Mosh team had the data to back it up. In the first two weeks of deployment, Lorikeet's [customer effort scores](/articles/customer-service-metrics) were higher than the human agents' scores. "Which upset our humans." The scores have since equalized - but with a caveat worth noting: the AI handles the straightforward cases while the humans handle the high-tension ones. Equal scores under those conditions is not parity. It's outperformance. ## "We were just so bored" Jess started evaluating [AI support](/blog/the-definitive-guide-to-ai-agents-for-customer-support) vendors at the end of 2024. She had calls with Sierra, Ada, and several others. "We went in so excited - we really firmly believed in the power of AI support, that it could solve hard problems and take action. And we were just so bored in those calls. They were cagey, and what they thought was flashy was just not impressive." Arbor is not a simple SaaS product. Electricity is personal - costs vary, data comes from multiple vendors operating in the dark ages, and most of the customer base has never been educated on how their billing works. "We need more love," Jess said. When she reached out to Lorikeet, the first call was different. "My coworker and I were Slacking behind the scenes... we were just so inspired. He was asking us really great questions about our support philosophy and what we were trying to get out of it. We just felt his investment from the first minute." The Mosh team had a similar experience from a different angle. They were choosing between Lorikeet and another vendor whose solution was out-of-the-box templates with an attitude of "here's what you get, good luck, make it work for you," whereas Lorikeet was wanting to build with us, take our feedback and build a tailored solution. The pattern across all three companies was the same. The buying decision was not a feature comparison. It was a bet on whether the vendor would co-build with you or ship you a template. ## Two features in a month Joel had been a Zendesk customer. The contrast in velocity was immediate. "Since we brought Lorikeet in, we've already had two separate features go live. Those have been really valuable right off the bat, communicated clearly, with us being part of test phases and our input taken into consideration. Whereas it takes me weeks to get anyone from Zendesk to actually look at something." For Jess, the velocity showed up in what it let her team become. "It means rising above the operational minutiae and really being able to focus on strategy," she said. "Making the work exciting and worthwhile, with people really invested in Arbor." She'd seen the opposite play out at a previous startup - rapid growth led to contractors who weren't invested, degraded quality, constant turnover. AI done well is the inverse of that pattern. The team stays small, engaged, and close to the product. The Mosh team described gamifying their workflow around Lorikeet's metrics. "The numbers I see - the percentages, the [quality score](/blog/defence-in-depth-how-to-actually-manage-ai-accuracy-in-production), the independently resolved - that's sort of my main job, making sure that number goes up. Greener and greener." Their product team had been relying on a separate analytics tool to find friction points in the customer journey. It was being deprecated. Lorikeet already had that data. This is the part of AI adoption that doesn't show up in vendor pitch decks. It changes the shape of the CX leader's job. Less managing headcount and attrition. More operating the system that serves your customers. ## The loading screen problem The telehealth CX leader made an observation that stopped the room. Email responses from the AI were arriving too fast. "When you send an email, you're not expecting an instant response. I wonder how many of our clients are missing the response - they're not expecting it that quickly, so they miss the notification." Some complaints: "There's no way you looked at my account." She referenced TurboTax's fake loading screen - it doesn't need the processing time, but the psychological impact of feeling like something is taking its time to look into your problem personally is better than an instant response. This is the kind of problem that never surfaces in a product demo. It only appears in production, in the gap between "technically correct" and ["actually trusted."](/blog/why-customers-reject-ai-and-how-to-fix-it) And it's the kind of insight that only reaches the engineering team when the person experiencing it is sitting three meters away from them. ## What they want next All three panelists converged on the same answer: proactive support. Jess described a cancellation workflow where the AI needs to follow up if a customer doesn't reply. "There are so many cases beyond that. If an unlinked customer reaches out and Olive informs them, and a week later it notices they still haven't completed the link - being able to proactively reach back out without the customer needing to respond first." Joel wanted the same. "In an ideal world, Moe could reach out three months later and say 'hey, I spoke to you three months ago - just checking in.' You're generating the ticket. It means providing a service that no one else is." The telehealth leader wanted to go further - AI that educates customers on how to talk to their insurance company, how to ask the right questions, how to navigate a system designed to deflect them. "Insurance companies are trained to not give you the answer you're looking for unless you know the exact right question to ask." Lorikeet is a distributed company. London, New York, San Francisco, Sydney. The product ships around the clock and the agents never sleep. That's the operating model, and it works. But last week in Honolulu, three customers told seventy engineers what they needed next - and the roadmap shifted before anyone left the room. The offsite was over. The work had already started. --- # We Gave Claude Code to Everyone URL: https://www.lorikeetcx.ai/blog/we-gave-claude-code-to-everyone **Lorikeet gave Claude Code to every employee — no platform team required. How hiring for agency and removing constraints drove company-wide AI adoption.** Most companies rolling out AI coding tools follow the same playbook: stand up a platform team, build an internal publishing layer, package everything into a single installation, then roll it out to the company. It makes sense if you have a thousand people. It makes sense if you want controlled adoption. We took a different path. Lorikeet has 70 people. Everyone uses Claude Code day-to-day. We didn’t build a platform to make it happen. We didn’t assign a team. It happened because of the kind of people we hire and what we gave them access to. ## No platform, no problem At an event a few weeks ago, there was an impromptu sharing session on “AI chief of staff” and personal AI productivity tooling. Company after company described their setup. What surprised us was that we appeared to be the only company with a shared repository of personal productivity skills that gives everyone access to the same tools the CEO uses, configured to their own role. That sounds small. It isn’t. At most companies, the founder builds their own Claude Code workflows – maybe a daily planning routine, maybe a sales digest – and those stay private. The tooling compounds for one person. Everyone else starts from scratch. We took the opposite approach. Every skill that works for leadership gets generalized and pushed to the repo. A skill to plan your day has a personal config section so anyone can use it. A skill to scan Slack update channels takes personally configured instructions for what to scan for. The BDR team has prospecting skills. The FDE team has subscriber diagnosis skills. The marketing team generates release content. The repo is open and anyone can contribute. ## What makes it stick Getting AI tools into the hands of non-engineers requires deliberate effort. The conventional answer is to build a platform. But there’s another way: build a culture where people just start using the tools because the environment makes it easy and the norms make it expected. A few things made this work at Lorikeet without a dedicated team. First, we connected Claude Code to everything early. The Lorikeet MCP for internal and external use was spun up quickly because we saw the potential. HubSpot, Linear, Grain, Notion, Slack, BigQuery, Google Calendar – they’re all [wired in](/integrations). When someone sits down with Claude Code for the first time, they can immediately do useful work because the connections are already there. Second, early advocates made it their business to get people set up. Loom walkthroughs, start-up guides, open availability to help. This wasn’t someone’s job title. It was people choosing to spend time onboarding their colleagues because they were genuinely excited about what the tools could do. Third, we removed the spend constraint. We openly allowed people to blow through Claude Code limits. Our team did hard work to battle Anthropic to make the pricing work for company-wide usage. Not many companies treat that as a priority. Most ration access, which sends an implicit signal: this tool isn’t really for you. ## Agency plus tools equals dynamite One of our senior engineers put it in a way that stuck: “Lorikeet over-indexes on hiring people who have agency and who do things without being told. We always saw this, but when coupled with being given the keys to Claude Code and access to logs, repo, database, and all the MCPs, we saw magic occur. Endlessly curious, high-horsepower people with agency plus Claude Code equals dynamite.” That’s the part that’s hard to replicate with a platform. You can build the best possible infrastructure for AI adoption, but if your culture doesn’t reward people for using it in unexpected ways, adoption plateaus at “the things the platform team thought of.” Agency is the multiplier. Here’s what 70 people with agency and Claude Code have built at Lorikeet, without anyone telling them to: an internal podcast that synthesizes information from key data sources and distributes it company-wide. Highly contextual microsites for subscribers and internal teams. A tools site at [lorikeet.tools](http://lorikeet.tools) that started as an experiment. Countless Notion documents that in another company would have taken a week of analyst time. A non-engineer on our sales team is building a daily planning tool that pulls from her calendar, Slack, HubSpot, and Notion goals to suggest how to block her day. Nobody assigned that project to her. ## The real question The question isn’t whether you need a platform team to make AI tools work for non-engineers. It’s whether you’ve hired people who will figure it out on their own if you give them the right access and get out of the way. The platform is a solution to a problem that agency solves first. At 70 people, we don’t have the luxury of building internal publishing platforms or design system overhauls for a tools catalog. What we have is a shared repo, a culture of contribution, and a bias toward just building something when you see a gap. As we grow, some of this will need more structure. But the foundation – hire people with agency, give them the real tools, remove the constraints, let them surprise you – is something that doesn’t scale out of relevance. It’s how you get a company where everyone builds, not because you gave them a platform, but because that’s the kind of person they are. --- # How to Calculate the True Cost of Agent Turnover URL: https://www.lorikeetcx.ai/blog/true-cost-of-agent-turnover **Agent turnover costs 3-4x more than most teams estimate. Learn how to calculate the full cascade — from SLA drops to burnout-driven attrition — and model it for your team.** In 2023, a mid-size e-commerce company lost 12 agents from a 40-person support team over six months. Their finance team calculated the damage at $180,000 - roughly $15,000 per departed agent in recruiting and training costs. By the end of the year, the real number was closer to $700,000. The difference wasn't a math error. It was a model error. Most organizations calculate agent turnover the same way: take the cost to recruit, hire, and train a replacement, multiply by the number of agents who left, and call it a day. SHRM pegs replacement cost at 50-200% of annual salary depending on role complexity. For a contact center agent earning $45,000, that's $22,500 to $90,000 per head. Painful, but bounded. The problem is that this model treats each departure as an isolated event. It isn't. ## The cascade nobody models When an agent leaves a 30-person team, 29 people now handle the same volume. That's a 3.4% capacity reduction on paper. In practice, the impact compounds in ways that spreadsheets don't capture. Start with wait times. A team handling 25 tickets per agent per day at a 6-minute average handle time has roughly 10% slack built into the schedule. Remove one agent and that slack evaporates. Remove three and you're running a queue that never clears — and the [financial damage of that growing backlog](https://lorikeet.tools/backlog-cost) compounds daily. Erlang C models - the standard math behind workforce planning - show that wait times don't increase linearly with utilization. They spike exponentially once utilization crosses 85%. A team running at 80% utilization that loses two agents might jump to 92%, and average wait times can triple. Longer wait times drag SLA compliance down. If your target is 80% of contacts answered within 60 seconds, losing even a small amount of capacity can crater that number. One insurance company tracked their SLA compliance dropping from 82% to 61% over three months after losing four agents from a team of 35. They didn't connect the dots until a quarterly review. Then CSAT follows. Research from the Customer Contact Council found that the single strongest driver of customer dissatisfaction is having to contact a company more than once to resolve an issue. When remaining agents are rushing through interactions to clear a growing queue, first-contact resolution drops. Customers call back. The queue grows further. ## Burnout is a feedback loop Here's where the standard model breaks down completely. The agents who stayed are now working harder, handling angrier customers (who waited longer), and watching their metrics slip. Occupancy rates above 85% sustained over more than a few weeks are strongly correlated with agent burnout and - eventually - more turnover. This is the cascade: departures cause overload, overload causes burnout, burnout causes more departures. Contact center turnover already runs between 30-45% annually as an industry average, according to NICE and ContactBabel's annual reports. Teams that experience a turnover spike often see it accelerate rather than stabilize. A workforce management director at a financial services firm described it as "the drain circling faster." They lost 8 agents in Q1. By the end of Q2, they'd lost 14 more. Their exit interviews all cited the same thing: unsustainable workload after the first wave of departures. ## The hidden line items Beyond the cascade, there are costs that never appear in turnover calculations because they belong to other budgets. New agents operate at roughly 60-70% efficiency during their first three months, even after formal training ends. That means every new hire displaces less volume than the person they replaced, extending the overload period for tenured agents. If your average time-to-hire is two months and ramp time is three months, you're looking at five months of reduced capacity per departure - not the two weeks most models assume. Quality costs are real too. New agents generate more escalations, more supervisor interventions, more rework. One B2B SaaS company found that agents in their first 90 days had a 40% higher escalation rate than tenured agents. Each escalation consumed 15-20 minutes of a senior agent's time, further reducing available capacity. There's also institutional knowledge loss. A five-year agent who knows the product edge cases, the workarounds for system limitations, and the particular needs of key accounts carries value that doesn't transfer through a training manual. When that person leaves, resolution times for complex tickets increase across the team because the informal knowledge network has a gap. ## Multiplier math When you layer these effects together, the true cost of agent turnover typically lands at 3-4x what teams estimate using replacement-cost-only models. Take a concrete example. A 50-agent team with an average salary of $48,000, handling 20 tickets per agent per hour, loses 8 agents over a quarter. The standard calculation: 8 agents times $35,000 average replacement cost equals $280,000. The fuller picture includes overtime costs to cover the gap (often 15-25% of base pay for remaining agents over several months), SLA penalties or credits issued to customers during the degraded service period, lost revenue from customers who churned due to poor experience, reduced productivity during the 3-5 month ramp period for each new hire, and additional turnover triggered by the overload on remaining staff. That $280,000 becomes $800,000 to $1.1 million when you account for the cascade. ## Why the number matters The reason to calculate this accurately isn't academic, it changes decisions. At $35,000 per departure, investing $200,000 in retention programs or automation looks like a tough sell. At $100,000 per departure including cascading effects, that same investment pays back in two prevented departures. Different math leads to different resource allocation. It also reframes what "solving" turnover means. If the cascade is the real cost driver, then reducing the impact of each departure matters as much as preventing departures in the first place. Teams that can absorb attrition without the service quality spiral - through better tooling, [automation of routine volume](https://lorikeetcx.ai), or more flexible staffing models — fundamentally change the economics even if their turnover rate stays the same. ## Running your own numbers The gap between estimated and actual turnover cost is large enough that it's worth modeling for your specific team. We built an [Agent Turnover Calculator](https://lorikeet.tools/attrition-impact) that takes your team's actual parameters - size, volume, handle times, SLA targets, salary, replacement costs - and models the full cascade: wait time increases, SLA compliance drops, CSAT impact, utilization spikes, burnout risk thresholds, and total cost with a line-item breakdown. Most teams that run it find their true cost is 3-4x their back-of-envelope number. The ones who act on that revised number tend to make different investments than the ones still budgeting based on replacement cost alone. Contact center turnover isn't going to zero. But the difference between a team that understands its real cost and one that doesn't is the difference between treating attrition as a line item and treating it as the operational risk it actually is. --- # AI Readiness Checklist: Is Your CX Team Ready? URL: https://www.lorikeetcx.ai/blog/ai-readiness-checklist-cx-teams **Most CX teams skip readiness assessment before deploying AI. Score yourself across six dimensions — from knowledge management to change management — with our free assessment.** Klarna's 2024 earnings call made a bold claim: their AI assistant was doing the work of 700 agents. The stock jumped. LinkedIn lit up. And within six months, Klarna quietly started rehiring human agents because resolution quality had cratered and customer satisfaction scores were falling. The technology worked fine. The organization wasn't ready for it. This pattern repeats across the industry. A VP of Support gets budget approval, signs a contract with an AI vendor, launches in eight weeks, and watches the whole thing underperform. Not because the model was bad, but because the team didn't have the operational foundation to make it work. The help center was outdated. The ticket taxonomy was a mess. Nobody had decided which conversations should even be automated. The AI inherited every dysfunction the team already had, and amplified it. Technology readiness is one dimension of AI adoption. It might not even be the most important one. ## The 700-agent fallacy Klarna's story gets cited as both a success and a cautionary tale depending on who's telling it, but the real lesson is simpler than either camp admits. They treated AI deployment as a technology project. It was an organizational one. When an AI agent handles a billing dispute, it needs more than a language model. It needs a knowledge base that accurately reflects current policy. It needs routing logic that knows when to escalate. It needs a team that's been trained to handle the cases AI can't, which are harder and more ambiguous than the ones it can. It needs someone monitoring quality, not just deflection rates. Most CX teams skip straight to "which vendor should we buy" — a decision with [its own complex trade-offs](https://lorikeet.tools/build-vs-buy) — without answering any of these questions first. ## Six dimensions, not one After working with dozens of support teams through their AI rollouts, a pattern emerges. The ones that succeed tend to be strong across six dimensions, not just the obvious technical ones. **Knowledge management** is the first and most predictive. If your help center doesn't exist, or exists but hasn't been updated in six months, AI will confidently serve stale answers to your customers. One e-commerce brand had 340 help articles, of which 40% referenced a returns policy they'd changed two quarters earlier. They launched AI anyway. Their CSAT dropped nine points in the first month. A [knowledge base health check](https://lorikeet.tools/kb-evaluator) before deployment would have caught it. **Process maturity** is the second. Teams that have documented their workflows - when to escalate, how to handle exceptions, what "resolved" actually means - give AI something to follow. Teams that run on tribal knowledge and Slack threads are asking AI to guess. It will guess wrong. **Data readiness** is the third, and it's more mundane than it sounds. It's not about having a data lake or a BI team. It's about whether your tickets are categorized consistently, whether you track resolution time in a way that means something, whether you can actually measure if AI is helping or hurting. A surprising number of teams can't answer "what percentage of our tickets are password resets" with any confidence. **Team structure** is the fourth. AI doesn't replace agents uniformly. It handles the repetitive, well-documented cases and leaves humans with the complex, emotional, ambiguous ones. That's a fundamentally different job than most support agents were hired for. Teams that haven't thought about how roles change post-AI end up with frustrated agents handling only escalations, burning out faster, and leaving. **Change management** is the fifth and most commonly ignored. The support team needs to believe this will make their jobs better, not eliminate them. The product team needs to feed AI-surfaced insights back into the product. Leadership needs to define what success looks like beyond "reduce headcount." Without alignment across these groups, AI projects stall after the pilot. **Technology readiness** is the sixth. Yes, it matters. Your ticketing system needs decent APIs. Your tech stack needs to support integration. But notice where it falls in the list. It's necessary and insufficient. ## The assessment gap The strange thing about AI buying cycles in CX is how little assessment happens before the purchase. In other enterprise categories - ERP, CRM, security - nobody would deploy without a readiness audit. You'd map your current state, identify gaps, build a remediation plan, then buy. With AI, teams skip straight to vendor demos. The urgency is understandable. Every board deck has an "AI strategy" slide. Every competitor claims they're already doing it. The pressure to move fast is real. But speed without readiness isn't speed, it's rework. The teams that "move fast" by deploying AI on top of a broken knowledge base spend the next quarter cleaning up the mess, rewriting articles, rebuilding trust with customers who got bad answers. The teams that spend four weeks getting their foundation right before deploying end up live faster and with better results. ## Scoring yourself honestly The hardest part of readiness assessment is honesty. Every CX leader thinks their knowledge base is "pretty good." Every team believes their processes are "mostly documented." These vague self-assessments are worthless. What works is forcing binary answers. Do you have a customer-facing help center, yes or no? Have your help articles been reviewed in the last 90 days, yes or no? Can you report on ticket volume by category with data you trust, yes or no? Binary questions eliminate the comfortable middle ground where teams convince themselves they're ready when they're not. This is why we built the [AI Readiness Scorecard](https://lorikeet.tools/ai-readiness) - a free tool that walks through all six dimensions with yes/no questions and produces a score out of 100. It takes five minutes. The median score across the hundreds of CX leaders who've taken it is 54. Most teams are closer to ready than they think in some dimensions, and further away than they'd admit in others. The value isn't the number itself. It's the specificity. A score of 62 tells you less than knowing you're strong on technology and process maturity but weak on knowledge management and change management. That's an actionable gap analysis. That tells you where to spend the next month before you sign a vendor contract. ## Foundation first The CX teams that will get the most from AI in 2026 and 2027 aren't the ones buying the fanciest tools. They're the ones doing the unglamorous prep work right now - auditing their knowledge base, documenting their escalation paths, cleaning up their ticket taxonomy, having honest conversations with their team about how roles will evolve. AI readiness isn't a technology question, it's an operational one. The teams that treat it that way will deploy faster, see better results, and avoid the expensive rewind that comes from launching before you're ready. --- # Build vs Buy AI Support: A Framework for CX Leaders URL: https://www.lorikeetcx.ai/blog/build-vs-buy-ai-support-framework **The 3-year cost of building in-house AI support is 2-4x what teams budget. Compare build vs buy with real numbers — including rebuild cycles, opportunity cost, and time to value.** In March 2024, OpenAI released GPT-4 Turbo with a 128K context window. Three months later, Claude 3.5 Sonnet landed with meaningfully better reasoning. By December, Google shipped Gemini 2.0. Every one of those releases changed the architecture assumptions underneath any AI support system built on the previous model. If your engineering team built an AI support agent on GPT-4 in early 2024, they've already had to make at least two significant rebuild decisions by now. That's the part nobody talks about when the "build" conversation starts. ## The V1 trap Every competent engineering team can build a V1. That's the problem. A basic retrieval-augmented generation setup that pulls from your knowledge base and generates answers to simple questions is genuinely not that hard. A senior engineer can have a working prototype in a few weeks. It answers easy questions, the demo looks great, and leadership signs off on a roadmap. What happens next is predictable. The prototype handles the 30% of tickets that are simple lookups - password resets, return policies, hours of operation. But the remaining 70% require understanding context, following multi-step processes, taking actions in backend systems, and knowing when to escalate. Building for that second tier is a fundamentally different problem than building V1, and it typically takes 3-5x the original build timeline. The V1 trap works because it creates a false sense of progress. You've spent $80K in engineering time, you have a working demo, and walking away feels like waste. So you keep building. This is the sunk cost fallacy wearing a technical disguise. ## The rebuild cycle nobody budgets for The 3-year cost model for building in-house AI support looks nothing like what gets approved in the original business case. Year 1 is the build. Two engineers, six months to get V1 live, another four months to iterate it toward production quality. At a fully loaded cost of $175K per engineer, that's $350K before you've handled a single real ticket at scale. You start getting meaningful AI resolution around month 10. Year 2 is the first rebuild. The foundation model you built on has been superseded. A new architecture - maybe agentic workflows, maybe a different retrieval approach - offers meaningfully better performance. Your team spends three months evaluating, two months rebuilding core components, and the rest of the year stabilizing. You also need at least 0.5 FTE on ongoing maintenance, prompt tuning, and edge case management. That's another $260K. Year 3 is the second rebuild, or the year your best engineer on the project leaves and you discover how much institutional knowledge walked out the door with them. Either way, you're spending $175K-$260K again. The 3-year total lands somewhere around $500K-$870K depending on team size and salary levels. And that's just the direct cost - it doesn't account for the opportunity cost of those engineers not building your actual product. ## Opportunity cost is the real number A Series B support technology company allocated two senior engineers to build an internal AI support agent in early 2024. Those engineers spent 14 months on the project before the company switched to a vendor solution. The AI agent they built handled roughly 20% of inbound volume by the time they sunset it. The CTO later estimated the opportunity cost at over $600K when you factored in the product features those engineers would have otherwise shipped. Two senior engineers over 14 months is meaningful product velocity - features that drive revenue, reduce churn, or open new markets. This is the calculation that rarely makes it into the build-vs-buy spreadsheet. Engineering capacity is finite. Every sprint spent maintaining an AI support system is a sprint not spent on your core product. For most companies, AI-powered customer support is not the product, it's infrastructure. ## When building actually makes sense Building makes sense in exactly one scenario: when your support operation is so unique that no vendor can accommodate it, and AI support is a core competitive differentiator for your business. If you're a developer tools company where the support experience is indistinguishable from the product experience, that might be you. If you're running a consumer fintech app where regulatory requirements create genuinely novel support workflows that no horizontal platform can handle, maybe. For the other 95% of companies, the support workflows are more similar than different. Customers ask questions, need status updates, want to make changes to their accounts, and occasionally need complex multi-step resolutions. The specifics vary by industry, but the patterns don't. That's exactly the kind of problem where buying a purpose-built solution beats building a custom one. ## The vendor equation The buy side of the comparison is simpler math. Integration takes 2-4 weeks of engineering time for [most modern platforms](https://lorikeetcx.ai), not months. Ongoing administration is a part-time responsibility for a CX operations person, not a full-time engineering role. And the vendor absorbs the model upgrade cycles, the architecture shifts, and the maintenance burden. A typical vendor engagement runs $4K-$6K per month depending on volume, with time to value measured in weeks rather than months. Over three years, that's $144K-$216K - less than half the build cost in most scenarios, with AI handling tickets from month 2 instead of month 10. That eight-month gap in time to value matters more than it looks on a spreadsheet. Eight months of tickets handled by AI at even 30% automation is meaningful headcount capacity freed up, faster response times delivered, and operational data collected that improves the system. ## Running your own numbers The framework is straightforward: compare the 3-year fully loaded cost of building (including rebuild cycles, maintenance, and opportunity cost) against the 3-year cost of buying (including integration, administration, and vendor fees). Then factor in time to value - when each option starts actually resolving tickets. We built an interactive calculator at [lorikeet.tools/build-vs-buy](https://lorikeet.tools/build-vs-buy) that runs this comparison with your actual numbers. Plug in your engineering salaries, estimated build timeline, and vendor costs. The default scenario shows a $344K gap over three years ($522K build vs $178.5K buy), but the point isn't the default - it's what happens when you use your real inputs. Adjust the rebuild assumptions, change the engineer count, update the salary numbers. The gap narrows or widens, but in almost every scenario, it doesn't close. ## The decision that matters The build-vs-buy question isn't really a technical question, it's a capital allocation question. Your engineering team can build an AI support agent. The question is whether they should, given what else they could build with that same time and budget. The companies getting this right are the ones who separate ego from analysis. Yes, your engineers are talented enough to build it. No, that doesn't mean building it is the highest-value use of their next 18 months. Run the numbers with honest assumptions about rebuild cycles and maintenance burden. The answer is usually clear, and it's usually buy. But before you sign a vendor contract, make sure your team is [ready to deploy AI effectively](https://lorikeet.tools/ai-readiness) — buying a platform doesn't help if the operational foundation isn't there. --- # What's Your Support Backlog Really Costing You? URL: https://www.lorikeetcx.ai/blog/support-backlog-cost **Support backlogs are a financial risk, not just an ops metric. Learn how to calculate the true cost — churn, escalation multipliers, SLA penalties — and make the business case.** In Q3 2024, a mid-market SaaS company with 12,000 customers let its support backlog grow to 1,400 tickets. The VP of Support reported it in the weekly ops review as a throughput problem: they were 23% over target queue depth. Leadership noted it, asked for a recovery plan, and moved on. By the end of the quarter, the company's net revenue retention had dropped 4 points. The connection between those two facts never made it into the same slide deck. This is how most companies treat support backlogs. They live in ops dashboards, measured in ticket counts and average handle times. They get discussed in support team standups, not finance reviews. And that framing - backlog as operational metric - is quietly draining millions from businesses that think they're tracking the right numbers. ## The invisible bleed A 500-ticket backlog doesn't feel like a financial emergency. It feels like a staffing problem, maybe a training gap. Something the support team will work through. But each of those 500 tickets represents a customer waiting, and waiting customers do predictable things. Research from the Harvard Business Review found that customers who had their issue resolved in under five minutes were more likely to make future purchases than customers who never had a problem at all. The inverse is equally true. Customers stuck in a queue don't politely wait for their turn. They start evaluating alternatives. They tell colleagues. They churn. If your average customer lifetime value is $15,000 and your backlog-driven churn rate increases by even 2%, a 500-ticket backlog isn't a queue management problem, it's a six-figure quarterly leak. ## SLA penalties are the obvious part Most support leaders can point to their SLA penalty exposure. It's contractual, it's measurable, it's the number finance already asks about. But SLA penalties are typically the smallest component of true backlog cost, and treating them as the whole picture is like measuring flood damage by counting broken windows. The real cost structure breaks down across five categories, and most organizations are only tracking one of them. Churn risk compounds daily as tickets age. Escalation costs spike because delayed tickets generate angrier customers who demand more senior attention - a ticket that would have taken 8 minutes at day one takes 35 minutes at day five, plus a supervisor review. Overtime spend creeps in as teams try to dig out. Prolonged overload drives [agent burnout and turnover](https://lorikeet.tools/attrition-impact), further reducing capacity. And missed expansion revenue never shows up in any report because the upsell conversation that would have happened didn't. A B2B support organization running 3,000 tickets a month with a 4-day average wait time and $200 average cost per ticket isn't just spending $600,000 on support. The loaded cost - including the churn, the escalation multiplier, the lost revenue - can run three to five times the direct operational number. ## Why the math stays hidden Support leaders aren't ignoring this. They're working with the tools and frameworks they've been given, and those frameworks treat support as a cost center with throughput metrics. Ticket volume. First response time. CSAT. Resolution rate. None of these translate directly to dollars lost. This creates a structural problem when it's time to make the business case for investment. A support director asking for four additional headcount can say "our backlog is growing and CSAT is dropping." A support director who can say "our backlog is costing us $340,000 per quarter in preventable churn alone, and here's the math" gets a fundamentally different conversation with the CFO. The gap between those two pitches isn't analytical skill. It's that the second pitch requires connecting support data to financial data in a way that most organizations haven't built the plumbing for. ## The escalation multiplier The least intuitive cost driver in a backlog is what happens to ticket complexity over time. A straightforward billing question submitted on Monday is a 6-minute resolution. That same ticket, untouched until Thursday, has a customer who's already called back twice, filed a complaint on social media, and demanded to speak with a manager. The original issue hasn't changed. The cost to resolve it has tripled. This escalation multiplier is well-documented in contact center research but rarely quantified at the company level. Forrester found that tickets older than 72 hours cost 2.4x more to resolve than tickets handled within 24 hours. The labor cost alone doubles. Factor in the customer relationship damage and the downstream retention impact, and you're looking at a 3-4x multiplier on every ticket that ages past the three-day mark. A backlog isn't a static queue, it's a compounding liability. Every day you carry it, the per-ticket cost increases. ## Headcount math vs. backlog math Here's the argument that support leaders should be making but often can't, because they don't have the numbers. Say you're carrying a persistent 600-ticket backlog with a 5-day average wait time across a 4,000-ticket monthly volume. Your direct cost per ticket is $18, but your loaded cost - factoring in escalation multipliers, churn probability at current wait times, and SLA exposure - is closer to $67. That's a $268,000 monthly cost overshoot driven primarily by the backlog itself. Four additional agents at a fully loaded cost of $65,000 each annually would cost $260,000 per year. The backlog is costing more than that every single month. The ROI math isn't close, but it never gets presented this way because the backlog cost is invisible and the headcount cost is a line item in the budget. The same logic applies to automation investments. When every unresolved ticket has an implicit daily carrying cost, the payback period on any tool that reduces backlog depth gets dramatically shorter — whether that's additional headcount or [AI automation](https://lorikeet.tools/build-vs-buy). But you can only run that calculation if you know what the backlog is actually costing. ## Turning queue depth into a dollar figure The formula isn't complicated. You need your customer lifetime value, a reasonable churn sensitivity estimate for your industry, current backlog depth, average wait time, SLA penalty terms, cost per ticket, and escalation rate. Multiply those together with the right weightings and you get a number that belongs in the quarterly business review, not just the support team standup. We built a [Backlog Cost Estimator](https://lorikeet.tools/backlog-cost) that does exactly this calculation. You plug in your numbers and get a breakdown across churn risk, SLA penalties, operational overhead, missed revenue, and overtime spend. It takes about two minutes and produces the kind of output you can put in front of a CFO. The point isn't the tool itself. The point is that support leaders need to start speaking in financial terms about what is fundamentally a financial problem. A 500-ticket backlog isn't a number on a dashboard, it's a dollar amount leaving your business every day, and the longer you measure it in ticket counts instead of revenue impact, the harder it is to get the investment to fix it. Support backlogs have been framed as an operational inevitability for too long. They're a quantifiable financial risk, and the organizations that treat them that way are the ones that actually get the budget to solve them. --- # Is Your Knowledge Base Ready for AI? A Health Check URL: https://www.lorikeetcx.ai/blog/knowledge-base-ready-for-ai **Your knowledge base is the #1 predictor of AI agent performance. Learn how poisoned language, duplicates, and thin content degrade AI — and how to audit before you deploy.** When a mid-market fintech company deployed their AI support agent last year, the first week looked great on paper. Resolution rates climbed, response times dropped, and the team started redirecting their attention to complex cases. Then the contradictions started surfacing. A customer asked about reimbursement timelines and got two different answers in the same conversation - one pulled from a 2022 policy article, another from an updated FAQ that covered the same topic with different numbers. The AI wasn't broken. It was doing exactly what it was told. The knowledge base just happened to be telling it two different things. This pattern repeats across almost every AI deployment in customer support. The technology works. The knowledge underneath it doesn't. ## The invisible problem Human agents are remarkably good at compensating for bad documentation. They skim an article titled "Escalation Procedures - Updated Jan 2021," recognize it's stale, and check with a colleague instead. They see "refer to manager for approval" and know that policy changed six months ago. They read an article that explains what the refund policy is but not how to actually process one, and they fill in the gaps from experience. AI agents don't do any of this. They take your knowledge base at face value. Every article is equally authoritative. Every instruction is meant to be followed. Every piece of content is current until someone tells them otherwise. This is why knowledge base quality is the single strongest predictor of AI agent performance. Not the model. Not the prompt engineering. Not the integration architecture. The knowledge. It's one of [six dimensions of AI readiness](https://lorikeet.tools/ai-readiness), and the one where most teams score lowest. ## What "bad" actually looks like The problems aren't dramatic. Nobody's knowledge base is full of obviously wrong information. The issues are subtle, accumulated over years of organic growth, and completely invisible to the humans who work around them every day. **Poisoned language** is the most common culprit. Articles written for internal consumption are littered with phrases that make perfect sense to a human agent but create bizarre customer experiences when an AI reads them literally. "Check the customer's tone and use your judgment" becomes an AI trying to assess sentiment and make autonomous decisions about escalation. "See Sarah in billing for exceptions" becomes the AI telling a customer to contact Sarah. "Use the internal portal to verify" leads the AI to instruct customers to access tools they can't see. **Duplicate coverage** creates contradiction. Over time, different authors write articles covering overlapping topics. The return policy lives in "Returns & Exchanges," "Refund Policy," and "Customer Satisfaction Guarantee" - each with slightly different details, different timelines, different exceptions. A human agent knows which one is canonical. An AI treats all three as equally valid, sometimes synthesizing them into a response that matches none of them. **Thin content** might be the hardest to spot because it looks complete. An article states that customers can request a plan change within 30 days. It doesn't explain what happens on day 31, whether there's a fee, how long the change takes to process, or what the customer should expect during the transition. A human agent reads between the lines or asks a colleague. An AI either makes something up to fill the gap or gives an answer so vague it's useless. **Stale articles** compound every other problem. They don't just contain outdated information - they actively compete with current articles for the AI's attention, creating conflicts that surface as hallucinations or contradictory responses. ## Why companies discover this too late The standard AI deployment timeline looks something like this: evaluate vendors, run a pilot, measure results, scale. Knowledge base readiness gets a quick mention in the implementation checklist, maybe a weekend of cleanup, and then the team moves on to integration work. The cleanup is almost always cosmetic. Someone archives the obviously obsolete articles, rewrites a few titles, and calls it done. The structural problems - the poisoned language embedded in hundreds of articles, the duplicate coverage patterns, the thin content that looks fine at a glance - survive intact. These problems don't show up in pilot metrics because pilots run on a narrow slice of topics with close human oversight. They show up at scale, when the AI is handling the long tail of customer questions and pulling from articles that nobody's reviewed in two years. By that point, the team is debugging AI behavior when they should be debugging content. Teams that [built in-house](https://lorikeet.tools/build-vs-buy) feel this most acutely — without vendor support to fall back on, every content issue becomes an engineering ticket. ## The audit that actually matters A meaningful knowledge base audit isn't a content review. It's a structural analysis. You're not reading articles for accuracy - you're scanning for patterns that will break AI comprehension. The questions that matter are specific. How many articles cover overlapping topics? What percentage contain internal-facing language? How many articles state a policy without explaining the associated process? What's the distribution of article freshness, and how many stale articles compete with current ones for the same queries? These are quantifiable signals. You can scan for them systematically rather than relying on someone to read 400 articles and catch every instance of "use your judgment" buried in paragraph six. This is exactly why we built the [Knowledge Base Evaluator](https://lorikeet.tools/kb-evaluator). Upload a CSV export of your help center articles and it scans for poisoned language, duplicates, thin content, and structural issues - then returns an AI readiness score with specific flagged articles. The data never leaves your browser. It takes about two minutes, and it surfaces problems that would take a team days to find manually. Companies that thought their knowledge base was in good shape routinely discover that 30-40% of their articles have at least one structural issue that would degrade AI performance. ## Fix the foundation first The companies that get the most value from AI support aren't the ones with the most sophisticated technology. They're the ones that did the unglamorous work of fixing their knowledge base before they deployed. They rewrote articles for an AI audience - removing internal jargon, replacing judgment calls with explicit decision trees, consolidating duplicate coverage into single authoritative sources. They built processes to keep content current rather than letting it drift. The readiness gap between companies that audit their knowledge base proactively and those that discover problems in production is measured in months of remediation and customer trust that's harder to rebuild than it was to maintain. Running a health check before deployment isn't just good practice, it's the difference between an AI agent that works and one that confidently gives your customers the wrong answer. --- # AI That Makes Your CX Team Better, Not Smaller URL: https://www.lorikeetcx.ai/blog/ai-that-makes-your-cx-team-better-not-smaller **Most AI vendors pitch replacing your support team. The better play is making them sharper. Here's why augmentation wins and what Coach does differently.** ## The Klarna Experiment In February 2024, Klarna announced that its AI chatbot had handled 2.3 million customer conversations in a single month, doing the work of 700 full-time agents. CEO Sebastian Siemiatkowski called it a revolution. Investors cheered. The fintech press wrote it up as a blueprint for every support org on the planet. By mid-2025, Klarna was [hiring human agents again](https://www.fastcompany.com/91468582/klarna-tried-to-replace-its-workforce-with-ai). Customer satisfaction had dropped. Complex issues were going unresolved. The AI that replaced 700 people could not replicate the thing those people actually did well: exercise judgment under pressure. Siemiatkowski admitted publicly that the cuts went too far. The company pivoted to what it now calls a "human-AI partnership," with AI handling routine queries and humans taking everything that requires empathy, discretion, or escalation. Klarna is not an outlier. It is a preview. ## The Rehiring Wave In February 2026, [Gartner predicted](https://www.gartner.com/en/newsroom/press-releases/2026-02-03-gartner-predicts-half-of-companies-that-cut-customer-service-staff-due-to-ai-will-rehire-by-2027) that by 2027, half of the companies that cut customer service staff because of AI will rehire for those same functions. Not because AI failed entirely, but because the "replace everyone" strategy produces a specific, predictable failure mode: the operation gets cheaper and worse at the same time. The data underneath that prediction is striking. Only 20% of companies that reduced headcount did so primarily because of AI. The rest blamed economic pressure and cost-cutting. But once AI became the narrative, every reduction got framed as automation success, whether it was or not. Salesforce is instructive here. CEO Marc Benioff told the [Logan Bartlett podcast](https://fortune.com/2025/09/02/salesforce-ceo-billionaire-marc-benioff-ai-agents-jobs-layoffs-customer-service-sales/) in September 2025 that his support team went from 9,000 to 5,000, crediting AI agents. What he also said, less quoted: there is now an "omnichannel supervisor" helping those agents and humans work together. In other words, even the company selling AI replacement landed on a hybrid model. Forrester puts a number on the regret: [55% of employers](https://www.forrester.com/press-newsroom/forrester-impact-ai-jobs-forecast/) who made AI-driven cuts say they moved too fast. ## What Agents Actually Do The replacement thesis rests on a misunderstanding of what customer service agents spend their time on. If you believe the job is mostly answering FAQ-level questions, then yes, a language model can do it cheaper. Tier-1 queries like password resets, shipping status checks, and basic returns policy questions are genuinely automatable. But anyone who has managed a real support team knows those tickets are not the ones that determine whether customers stay or leave. The hard work is the insurance claim where the customer is frustrated and the policy is ambiguous. The billing dispute where three systems show conflicting data. The onboarding call where a new customer is ready to churn 48 hours after signing up. A [Stanford and MIT study](https://www.cnbc.com/2023/04/25/stanford-and-mit-study-ai-boosted-worker-productivity-by-14percent.html) of over 5,000 contact center agents found that when AI was used to assist agents rather than replace them, productivity rose 14%. Newer agents saw a 35% increase in chats resolved per hour. Requests to speak with a manager dropped 25%. The AI compressed ramp-up time, meaning a two-month-old agent performed like an eight-month veteran. That is not a story about AI doing the job instead of humans. It is a story about AI making humans better at a job that remains fundamentally human. ## The Real Problem Nobody Talks About Contact centers have a 40-45% annual turnover rate. The [average agent stays 14.3 months](https://www.insigniaresource.com/research/call-center-turnover-rates/). That means the typical support organization is replacing nearly half its workforce every year, losing institutional knowledge, retraining constantly, and watching quality fluctuate with every cohort change. This is the problem that keeps CX leaders up at night, not whether AI can answer a password reset question. The existential challenge is maintaining consistent quality across a team that is always partially new. Traditional QA barely helps. Most contact centers [manually review 1-2% of interactions](https://callminer.com/faq/what-is-call-center-quality-assurance). A QA analyst listens to a handful of calls per agent per month, fills out a scorecard, and delivers feedback days or weeks after the conversation happened. The sample is too small to be statistically meaningful. The feedback loop is too slow to change behavior. And the process itself is expensive enough that scaling it means hiring more QA people, which puts you right back where you started. This is where the AI-for-replacement crowd gets the diagnosis right but the prescription wrong. They see that support operations are inefficient and conclude the answer is fewer humans. The actual answer is better-supported humans. ## Coaching Over Cutting Think about how the best sports teams operate. When a player underperforms, the response is not to cut the roster. It is to review film, identify patterns, and coach. The best teams invest in making their existing players better because they understand that talent development compounds in ways that talent replacement does not. Support organizations need the same approach, and AI makes it possible at a scale that was never realistic before. Instead of reviewing 1-2% of tickets, AI can evaluate 100% of conversations against a consistent rubric. Instead of a QA analyst spending 30 minutes per ticket review, an AI agent can assess quality, flag deviations from protocol, and identify coaching opportunities across thousands of interactions in real time. Instead of feedback arriving two weeks after the conversation, it can surface the same day. The data supports the impact. Agents who receive [personalized AI-driven coaching report 91% job satisfaction](https://www.supportbench.com/coach-agents-using-qa-data-without-making-it-punitive/), compared to 57% for those getting generic feedback. Regular coaching sessions grounded in QA data [improve agent performance by 25-30%](https://www.amplifai.com/blog/call-center-turnover) while reducing attrition by 20-40%. Read those numbers again. A 25-30% performance improvement without changing a single person on the team. A 20-40% reduction in the turnover problem that makes everything else harder. This is not incremental, it is structural. ## The Skeptic's Entry Point If you lead a CX team and your CEO just told you to "look at AI," you are probably dreading vendor calls. Every demo starts the same way: here is how many agents we can replace, here is the cost savings, here is the deflection rate. The pitch assumes you want fewer people on your team. But you built that team. You hired people who care about customers. You know that your agents' ability to handle a difficult conversation with empathy is the reason your retention numbers look the way they do. The last thing you want is a vendor telling you those people are a cost center to be optimized away. You are not wrong to feel that way. The buyers who understand that empathy is a competitive advantage are the ones building the most durable customer relationships. The question is not whether AI has a role in your operation. It does. The question is what role. The most productive starting point is not automation. It is visibility. ## What Visibility Looks Like A CX manager at a healthcare platform recently described her workflow before AI-powered QA. She would pull batches of conversations manually, paste them into ChatGPT, and ask for summaries and theme groupings. Hours of work to get a rough sense of what was happening across her team. The analysis was limited to whatever she had time to sample, which was never enough. With an AI agent evaluating every conversation, she could ask a single question: "What were the primary friction points customers reported across all conversations tagged with early pay last week?" She got sentiment analysis, exact quotes, and a breakdown of which interactions frustrated customers and which satisfied them. The analysis that used to take hours took minutes, and it covered the full dataset instead of a biased sample. That manager is now running this analysis across every product line. She presents findings to the company showing which products drive the most support friction, turning contact rate data into product improvement signals. Her support team did not get smaller. It became the source of customer intelligence for the entire organization. This is the shift that AI skeptics miss when they refuse to engage, and that AI evangelists miss when they fixate on headcount reduction. The highest-value application of AI in customer service is not doing the work instead of your team. It is giving your team the information and tools to do work that was previously impossible. ## From QA to Continuous Improvement The old QA model is pass/fail. Did the agent follow the script? Did they hit the required talking points? The model produces compliance, not growth. The better model treats every conversation as training data. When an AI agent escalates a ticket and a human resolves it, the system should analyze what the human did, aggregate those solutions across similar tickets, and surface the patterns. Over time, this creates a feedback loop: the AI gets better because humans teach it, and humans get better because the AI identifies their strengths and gaps. One company's CXO described it this way: "When the AI can't answer and it gets escalated, go look at what the human said that resolved it. Aggregate across those conversations. The agent said basically the same thing. That is the first thing you should focus on, go fix your content." That is not a replacement dynamic. That is a partnership where each side makes the other more effective. ## Where Lorikeet Coach Fits This is what we built [Lorikeet Coach](https://www.lorikeetcx.ai/blog/launching-coach) to do. Coach is an AI agent that evaluates every support conversation, both human and AI-handled, against a customizable quality rubric. It identifies where agents follow protocol and where they deviate. It scores empathy, accuracy, policy adherence, and resolution quality across 100% of your tickets, not a 2% sample. But evaluation is only the starting point. Coach translates QA data into coaching opportunities. It surfaces the specific conversations and patterns that managers need to have productive one-on-ones with their team. It identifies knowledge gaps and recommends content updates. It detects when quality is drifting before the numbers show up in your CSAT scores. For teams that also use Lorikeet's AI concierge, Coach closes the loop entirely. When it identifies a pattern of escalations around a specific topic, it can recommend workflow changes, draft updated reference material, and test the fix before it goes live. The human agents inform the AI improvement, and the AI improvement reduces the load on human agents. For teams not using Lorikeet for automation, Coach works as a standalone product. It plugs into your existing helpdesk, evaluates your human agents' conversations, and gives your managers the data they need to run a better team. No automation required. No agents replaced. We built it this way because we believe the companies that win in customer experience over the next decade will not be the ones that cut the deepest. They will be the ones that invested in making their people exceptional while using AI to handle the work that does not require human judgment. The pitch from every other vendor is: give us your tickets and we will make your team smaller. Our pitch is different. Give your team better tools and they will make your company better. --- # When Your AI Support Platform Hits a Ceiling URL: https://www.lorikeetcx.ai/blog/when-your-ai-support-platform-hits-a-ceiling **Decagon works for logo-rich SaaS. But teams in regulated industries hit a ceiling: no compliance guardrails, black-box reasoning, and rigid workflows. Here is what is actually missing.** Last quarter, the CEO of a $400K-ARR financial services company told his CX leader something blunt. His board had been pressuring him to switch to Decagon or Sierra. Bigger brands, bigger funding rounds, the logic that "nobody gets fired for buying IBM." He flew his team out to spend a day with the incumbent vendor instead. After 90 minutes of live product testing, he turned to his exec sponsor and said: "Now I have an answer for the board about why we should not switch." That story is not unusual. We hear versions of it regularly from CX leaders at regulated companies who evaluated Decagon and found a gap between the pitch and what their business actually needs. This is not a hit piece. Decagon is a well-funded company with legitimate enterprise customers. But there is a structural mismatch between what Decagon sells and what complex CX organizations require. ## The logo wall Visit [decagon.ai](https://decagon.ai) and the first thing you see is logos: Duolingo, Chime, Notion, Rippling, Substack, Eventbrite. It is an impressive collection and it does exactly what it is supposed to do: signal credibility through association. The logos carry the page because the copy does not. Run Decagon's messaging through any structured analysis and a pattern emerges. The language is almost entirely about resolution rates and automation percentages. Chime achieved 70% AI resolution. Rippling saw 32% ticket deflection. These are real numbers from real companies. But notice what is absent. There is no language about compliance. No mention of regulatory frameworks. No proof points from healthcare, insurance, or financial services companies operating under HIPAA, PCI-DSS, SOX, or state-level consumer protection mandates. The social proof is strong. The substance beneath it is narrow. ## Where it works Decagon built a good product for a specific segment. SaaS companies with high ticket volume, relatively standardized queries, and technically sophisticated teams get real value. Their Agent Operating Procedures (AOPs) let you define workflows in natural language. Their integrations with tools like Zendesk and Intercom are functional. For a Series B SaaS company handling password resets, billing questions, and feature requests, Decagon can automate a meaningful percentage of volume. Their $250M Series D at a $4.5B valuation in early 2026 reflects genuine market traction. They have reportedly crossed 100 enterprise customers and added Deutsche Telekom, Avis, and Block to their roster. That is real momentum. The question is not whether Decagon works. It does, for the segment it was built for. The question is what happens when a company's CX needs grow beyond that segment. ## The compliance gap Pull up Decagon's website and search for "compliance." Search for "regulated." Search for "audit trail" or "guardrails" or "HIPAA" or "PCI." The results are thin to nonexistent. This is not an oversight. It reflects a product architecture decision. Decagon was built for tech-forward consumer companies where the cost of a wrong AI response is a frustrated user, not a regulatory fine. When your biggest customers are Duolingo and Notion, compliance infrastructure is not your first priority. But for a healthcare company where an AI agent's response about medication could trigger an FDA review, or a financial services firm where an incorrect statement about account terms violates TILA, the absence of compliance tooling is not a nice-to-have gap. It is a disqualifier. Fenergo reported that regulatory fines in the first half of 2025 totaled $1.23 billion, a 417% increase over the prior year. The cost of getting AI-assisted customer interactions wrong in regulated environments is not a bad CSAT score. It is an enforcement action. In our experience working with regulated companies, the compliance requirement breaks down into three specific capabilities that most AI CX platforms lack. **Guardrails that are both built-in and customizable.** Generic safety filters are not enough. A health insurance company needs guardrails that prevent the AI from making coverage determinations. A fintech needs guardrails that flag when a conversation approaches credit decision territory. These rules vary by jurisdiction, by product line, and sometimes by individual customer segment. **Full audit trails with explainable reasoning.** Regulators do not accept "the AI decided" as an explanation. Every response needs to be traceable to the knowledge, policy, and logic that produced it. Multiple reviews of Decagon cite a "black box" problem: you cannot always see why the AI did what it did. That makes it difficult to review conversations, adjust behavior, or investigate when something goes wrong. **Deterministic behavior in high-stakes moments.** There are interactions where probabilistic AI output is unacceptable. When a customer asks about their insurance coverage limits or their loan terms, the answer must be precisely correct. This requires the ability to enforce deterministic business logic alongside LLM-powered conversation. Not one or the other. ## The rigidity problem Decagon's Agent Operating Procedures are marketed as natural-language workflow configuration for non-technical teams. In practice, users report a different experience. Implementation demands Agent Engineers and weeks of setup. Getting even basic tasks configured requires engineering support. One G2 reviewer noted that compared to platforms that let non-technical users adjust workflows, tone, or logic, Decagon "can feel rigid once deployed, which becomes a problem when you need to adapt quickly to new products, policies, or edge cases." For companies in regulated industries where policies change with every regulatory update, quarterly product launch, or jurisdiction expansion, this rigidity compounds. A CX leader we spoke with during a competitive evaluation put it directly: Decagon's product walkthrough felt accessible for less technical audiences, but the underlying system required more Python knowledge than expected. The pitch is low-code. The reality is not. This creates a dependency loop. Your CX team identifies a policy change that requires updating the AI agent's behavior. They cannot do it themselves. They file a ticket with your engineering team or with Decagon's professional services. Days pass. Meanwhile, the agent is giving outdated responses to real customers. In regulated environments, those outdated responses can create liability. ## The pricing question Decagon does not publish pricing. Based on marketplace data and reports from prospects we have spoken with, median annual contracts sit around $400,000, with a $50,000 minimum threshold below which Decagon's sales team will not engage. For companies with massive ticket volume and straightforward use cases, this pricing can pencil out on a per-resolution basis. But the total cost of ownership extends well beyond the contract. Factor in the engineering resources required for implementation, the ongoing need for technical staff to maintain and update workflows, and the professional services costs for any non-trivial configuration change. One prospect told our sales team that Decagon informed them their deal was "too small to take on." That is a legitimate business decision for Decagon, but it reveals who the product is built for and who it is not. If your annual CX automation budget is under $200K, you are likely not their target customer. ## The cloud-only constraint Decagon operates exclusively as a cloud-hosted platform. For many SaaS companies, this is fine. For enterprises in regulated industries, it can be a hard stop. Healthcare organizations subject to HIPAA, financial institutions with data residency requirements, and government-adjacent entities with FedRAMP considerations often need deployment flexibility: the ability to run within their own infrastructure or within specific geographic boundaries. Decagon's architecture does not accommodate this. Rasa, one of Decagon's competitors, has built its positioning explicitly around this gap, offering self-hosted deployment with deterministic business logic for regulated industries. The fact that an entire category of competitors exists specifically to address Decagon's deployment model limitation tells you something about the size of the gap. ## The concierge question In March 2026, Decagon launched what they called "proactive agents" with customer memory. This followed Sierra launching similar capabilities the previous year. Both announcements validated a category that had already been in production elsewhere: AI agents that do not just answer questions but proactively engage with customers based on context, history, and behavioral signals. The timing is notable because it highlights a pattern. Decagon's product development appears to follow market leaders rather than lead. Their recent blog post arguing that "MCP is not enough and tools need guardrails" arrived months after other vendors shipped guardrails into general availability. Their "Duet" product launched after competitors had already deployed similar capabilities. Following fast is a viable strategy. But for CX leaders evaluating vendors today, it means Decagon's roadmap is reactive. You are buying a product that will eventually ship what competitors already have, not a product that is defining where the category goes next. ## Who actually switches We have seen three patterns among companies that evaluate or leave Decagon. **Pattern 1: The regulated company that never should have been there.** A fintech or health tech company signs with Decagon based on the logo wall and the resolution rate promise. Six months in, they discover the compliance tooling does not exist and their legal team starts flagging AI responses that create liability. The switching cost is real, but the regulatory risk of staying is higher. **Pattern 2: The company that outgrows the product.** A SaaS company starts with Decagon for basic ticket deflection and it works well. As they expand into more complex use cases, they hit the ceiling of what the platform can do without heavy engineering investment. The total cost of ownership quietly exceeds what they budgeted. **Pattern 3: The company that never gets started.** A prospect evaluates Decagon and finds the minimum contract threshold, the engineering requirements, or the lack of compliance infrastructure is a blocker before they ever sign. These companies are the largest segment and the hardest to see because they never become Decagon customers in the first place. ## Evaluation questions If you are evaluating Decagon or any AI CX platform for a complex or regulated environment, these questions will surface the gaps fastest. **Ask them to show you their guardrails configuration.** Not the concept. The actual interface where your CX team would set compliance rules without engineering support. Ask how quickly a policy change can be reflected in live agent behavior. **Ask them to trace a single AI response back to its source.** Pick a complex customer interaction from your own data and ask the vendor to show you exactly which knowledge, policy, and logic produced the response. If the answer involves the phrase "we are working on that," you have your answer. **Ask them about deployment options for your regulatory environment.** If you have data residency requirements, ask specifically how they are met. "Our cloud is SOC 2 compliant" is not the same as "we can deploy within your infrastructure." **Ask them who on your team will own day-to-day agent management.** If the answer requires dedicated engineering headcount, factor that into your total cost of ownership and compare it against platforms where CX teams manage the agent directly. ## Where Lorikeet fits [Lorikeet](https://www.lorikeetcx.ai) was built specifically for the companies that outgrow platforms like Decagon. We work with financial services companies, healthcare organizations, and insurance providers where every AI interaction happens inside compliance boundaries. Our guardrails are both built-in and fully customizable. Your CX team can configure compliance rules, jurisdiction-specific policies, and brand guidelines without writing code and without filing an engineering ticket. Changes go live immediately. Our AI agent's reasoning is fully auditable: every response traces back to the specific knowledge, policy, and guardrail that produced it. We deploy across chat, email, and voice with sub-second latency. We resolve tickets end-to-end, not just deflect them. And we do it at a price point that does not require a $400K annual commitment to get started. The companies that choose us over Decagon do so because they need AI that works within the constraints of their industry, not AI that was built for a different industry and retrofitted with a compliance checkbox. That is a structural difference, not a feature comparison. [**Talk to Lorikeet about AI customer support built for regulated industries.**](https://www.lorikeetcx.ai/get-a-demo) --- # From First-Gen Chatbot to AI Agent: What to Look for in Your Second Platform URL: https://www.lorikeetcx.ai/blog/from-first-gen-chatbot-to-ai-agent **Already deployed AI for customer service and been disappointed? Here is what experienced, skeptical second-time buyers should look for when evaluating their next AI agent platform.** Last February, a fintech CX leader I know sat across from her board and presented a slide titled "AI-First Support: Year One Results." The company had deployed one of the marquee AI chatbot vendors twelve months earlier. Deflection rate: 48%. Headline cost savings: $1.3 million. But then she showed the next slide. Customer satisfaction had dropped eleven points. Escalation-to-resolution time had increased because agents receiving handoffs had zero context. Repeat contacts on the same issue were up 34%. The board asked one question: what are we doing about this? She is not an outlier. She is the new normal. ## The graduation cohort. Most AI customer service vendors are still selling to first-time buyers: companies that have never deployed AI for support and are dazzled by a slick demo. But a growing segment of the market has already bought, deployed, measured, and been disappointed. They tried Zendesk AI, or Fin, or Ada, or a custom build wired to OpenAI. They got a deflection engine dressed up as automation. Now they are shopping for their second platform. This cohort is different. They have scar tissue. They know what a vendor demo hides. They know the difference between a contained conversation and a resolved one. And they are asking harder questions than they did the first time around. [Gartner found that only 8% of customers used a chatbot in their most recent service interaction](https://www.gartner.com/en/newsroom/press-releases/2023-06-15-gartner-survey-reveals-only-8-percent-of-customers-used-a-chatbot-during-their-most-recent-customer-service-interaction), and just 25% of those said they would use one again. That is not a technology adoption problem. That is a product quality problem created by the first generation of tools. ## Deflection is not resolution. The original sin of first-gen chatbots was measuring success by how many conversations they kept away from human agents. Deflection. Containment. The metric itself reveals the design philosophy: the AI exists to protect your team from customers, not to help customers. This worked just well enough to sell. A VP could show the CFO a deflection rate and claim savings. But the savings were an illusion. [Gartner research found that only 1 in 7 customer service queries are fully resolved through self-service](https://www.cxtoday.com/contact-center/only-1-in-7-customer-service-queries-resolved-with-self-service-gartner-study-finds/). Even for issues described as "very simple," only 36% were handled without a human. The rest bounced back into the queue, often angrier than when they started. The distinction between deflection and resolution is the single most important concept for second-time buyers to internalize. Deflection asks: did we avoid a human interaction? Resolution asks: did the customer's problem get solved? When you evaluate your next platform, open the reporting dashboard before anything else. If you cannot see resolution rate and deflection rate as separate, distinct metrics, walk away. The vendor has already told you what they optimize for. ## The Klarna warning. Klarna became the poster child for AI-first customer support in 2024, announcing its AI could replace 700 agents and handle 2.3 million conversations per month. By mid-2025, the company was [rehiring human agents after customer satisfaction cratered](https://www.bloomberg.com/news/articles/2025-05-08/klarna-turns-from-ai-to-real-person-customer-service). CEO Sebastian Siemiatkowski told Bloomberg the company had gone "too far in the wrong direction." His diagnosis: "Cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality." Klarna's mistake was not adopting AI. It was adopting AI that could only do one thing: answer questions from a knowledge base. When customers needed refunds processed, account details changed, or billing disputes investigated, the bot could not act. It could only talk. Customers received what [CNBC described as "generic, repetitive, and insufficiently nuanced replies"](https://www.cnbc.com/2026/04/01/ai-chatbot-customer-service-complaints-refunds.html) and flooded complaint channels. The lesson is structural, not cautionary. First-gen tools were built to summarize FAQs and route tickets. That architecture has a ceiling, and most companies hit it within twelve months. ## What broke. Across conversations with dozens of CX leaders who are on their second or third AI vendor, the failure modes cluster into five patterns. **No memory across conversations.** Each ticket is treated as a standalone event. The AI does not know the customer asked about the same issue last week, was promised a callback, or has contacted support four times in ten days. [74% of customers expect bots to remember past interactions](https://www.zoom.com/en/blog/chatbot-statistics/), but only 28% say this matches their experience. **No ability to take action.** The bot can tell a customer what the refund policy is. It cannot process the refund. It can explain how to update billing information. It cannot update it. For anything requiring a write operation to a backend system, the conversation dead-ends with "let me transfer you to an agent." **Broken handoffs.** When the AI does escalate, the human agent receives a ticket with minimal or no context. The customer repeats everything. The agent has no visibility into what the AI already tried. The experience feels worse than no AI at all, because the customer already invested time explaining the issue once. **Brittle at the edges.** The demo worked on the ten most common queries. But real support traffic includes billing disputes (where [Gartner found only 17% chatbot resolution](https://www.gartner.com/en/newsroom/press-releases/2023-06-15-gartner-survey-reveals-only-8-percent-of-customers-used-a-chatbot-during-their-most-recent-customer-service-interaction)), multi-step workflows, regulatory edge cases, and emotional customers. First-gen tools handle the easy middle and collapse on everything else. **Impossible to maintain.** One prospect we spoke with described their custom OpenAI integration as "finicky," requiring constant maintenance every time documentation changed or a workflow needed updating. Another called their legacy chatbot setup "a house of cards." When the underlying model updates, the behavior shifts unpredictably, and there is no testing infrastructure to catch regressions before they hit customers. ## The evaluation playbook. Second-time buyers have earned a more sophisticated evaluation framework. Here is what the best ones are doing. **Test with your worst tickets, not your best.** Pull 200 real tickets from your queue, including the messy ones, the multi-turn conversations, the policy-ambiguous cases, and the edge cases that trip up your human agents. Any vendor confident in their platform will welcome this. If they insist on using their own demo data, that tells you everything. **Trigger an escalation during the demo.** Then switch to the agent interface. Did the human agent get full context? Can they see the entire conversation thread? Do they know what the AI already tried and what the customer already said? The handoff is where most platforms disintegrate. **Ask to see a write operation.** Not a knowledge base lookup. Not a canned response. Ask the AI to process a refund, update an account, cancel and rebook a reservation. If the vendor hesitates or says "that is on our roadmap," you are looking at another deflection engine. **Check the analytics for resolution, not containment.** Can the platform tell you what percentage of conversations ended with the customer's problem actually solved, verified by outcome data rather than assumed by conversation closure? Can it show you which intents are rising, where resolution fails, and which workflows cause repeat contacts? **Ask about the knowledge management burden.** [61% of support leaders report backlogs in editing outdated knowledge articles](https://www.accio.com/business/gartner_customer_service_trends), and that directly limits AI effectiveness. How does the platform handle knowledge that is stale, conflicting, or missing? Does it fail silently, or does it surface the gap? ## Resolution means action. The fundamental architectural difference between first-gen chatbots and actual AI agents is the ability to take action in the real world. A chatbot reads from your systems. An AI agent reads and writes. It processes the refund, updates the address, escalates with full context to an internal team, and follows up to confirm the issue was resolved. The [AI agent market is growing at 46.3% CAGR](https://www.blueprism.com/resources/blog/future-ai-agents-trends/), nearly double the chatbot market's 23% growth rate, because buyers have learned that answering questions is table stakes and solving problems is the actual product. This is not a marginal upgrade. It is a category shift. When your AI can take action, the economics change. The customer gets a resolved issue in minutes instead of hours. The human agent team handles genuinely complex cases instead of re-doing work the bot could not finish. And the cost per resolution drops because you are not paying twice for the same interaction. [Gartner predicts agentic AI will autonomously resolve 80% of common customer service issues by 2029](https://www.gartner.com/en/newsroom/press-releases/2025-03-05-gartner-predicts-agentic-ai-will-autonomously-resolve-80-percent-of-common-customer-service-issues-without-human-intervention-by-20290). The companies already operating this way are not waiting for 2029. They are building the muscle now. ## What Lorikeet gets right. We built [Lorikeet](https://www.lorikeetcx.ai) for this exact buyer: the CX leader who has already been burned and is not interested in another deflection engine. Lorikeet is an AI customer support platform that resolves tickets end-to-end across chat, email, and voice. That means processing refunds, updating accounts, coordinating with third parties, and handling complex multi-step workflows. Not summarizing FAQs. Not containing conversations. Resolving them. Three things matter to second-time buyers, and they are all core to how we built the product. **Resolution over deflection.** Lorikeet measures and optimizes for end-to-end resolution. The agent does not succeed by avoiding a human interaction. It succeeds when the customer's problem is actually solved. That is a fundamentally different design philosophy than containment-first platforms, and it shows up in how the product behaves with edge cases, complex queries, and multi-turn conversations. **Action-oriented architecture.** Lorikeet agents take action: calling APIs, processing transactions, updating records, and coordinating with internal teams. Our [Team of Agents](https://www.lorikeetcx.ai/blog/team-of-agents) capability lets a primary agent spawn secondary agents that contact third parties, investigate issues, and report back to resolve the customer's problem entirely. A customer reports a missing delivery. Lorikeet contacts the logistics provider, gets the status, and reports back. Minutes, not days. **Context that survives the handoff.** When Lorikeet does escalate to a human, the agent gets the full conversation, every action the AI took, and a summary of what was tried and what remains. No customer repetition. No blind handoff. The human picks up where the AI left off, with complete context. We work with companies in fintech, healthcare, and other regulated industries where getting the answer wrong has real consequences. That is a forcing function for quality that FAQ-based systems never face. ## The second purchase is smarter. The first time a company buys AI for customer support, they buy the vision: fewer tickets, lower costs, happier customers. The second time, they buy the architecture: what can it actually do, how does it handle failure, and how does it prove it worked? [64% of customers say they would prefer companies did not use AI for customer service](https://www.gartner.com/en/newsroom/press-releases/2024-07-09-gartner-survey-finds-64-percent-of-customers-would-prefer-that-companies-didnt-use-ai-for-customer-service). That number is not a rejection of AI. It is a rejection of bad AI. It is the accumulated frustration of millions of people who were deflected, misunderstood, and forced to repeat themselves by first-gen tools that measured success by how many conversations they avoided. The companies that win the next phase of AI customer service will not be the ones with the most sophisticated language models or the largest knowledge bases. They will be the ones whose AI can actually solve problems. Take action. Remember context. Prove resolution. If you are shopping for your second platform, you already know this. Trust what you learned. Demand more than deflection. And test everything with your hardest tickets, not your easiest ones. [**See how Lorikeet handles the tickets your last platform could not**](https://www.lorikeetcx.ai/get-a-demo)**.** --- # Customer Stories --- # Arbor Scaled Human-Quality CX With Universal AI Concierge URL: https://www.lorikeetcx.ai/customer-stories/lorikeet-helped-arbor-automate-complex-support-flows-across-13-states **Lorikeet helped Arbor automate complex support flows across 13 states — matching human CSAT and navigating sensitive, regulated energy workflows.** **From complex energy questions to clean handoffs. All in minutes.** “The Lorikeet team loves untangling complex models — they know exactly what issues we’re solving for.” ## Key Stats - 89% — CSAT - 1:1 — AI CSAT on par with human agents - 4.7 — Trustpilot rating, up from 4.6 before Lorikeet ## Why Lorikeet ### Arbor needed AI that could solve hard problems — not just deflect Arbor helps U.S. consumers in 13 deregulated states automatically lock in lower fixed electricity rates on a continuous basis. Their support team handles a steady stream of brand and energy questions — from decoding bills to explaining rate options — plus complex, state-specific inquiries tied to brokerage rules, utility breakdowns, and market conditions. When Arbor started exploring AI, they weren’t looking to deflect — they wanted a partner with their shared support philosophy and the ability to understand and navigate the consumer energy landscape. “Most AI companies talk about deflection like it’s the goal,” said Jessica, Arbor’s Head of Customer Operations. But Jessica’s team wanted the opposite. They wanted to solve hard problems, especially when the customer stakes were high and emotionally-charged. They considered Sierra and other well-known players, but most felt built for retail and e-commerce — not for navigating the often-complex world of retail energy supply and handling nuanced workflows. Lorikeet stood out by focusing on exactly what Arbor needed: AI agents that could handle nuance, de-escalate emotional moments, and follow SOPs without breaking a sweat. ## Deep integration ### Arbor was up and running by the end of the week Adopting a new tool is the easy part — it’s integration and onboarding where things usually get stuck. With Lorikeet, Zendesk was connected in under 15 minutes, and custom API endpoints were deployed within days. But Arbor needed more. Their knowledge base wasn’t just state-specific — it spanned state, utility provider, and energy nuances. So the Lorikeet team built a bespoke Notion integration to sync that evolving, hyper-specific content into the system in real time. ## Agent training ### Lorikeet agent treats Arbor customers as if they were their own The real nailbiter? Putting your AI agent to the test. The Lorikeet team built out automated workflows like tagging tickets and logging CRM notes, alongside an FAQ system that blended Arbor’s dynamic knowledge base with real-time customer data. The agent was trained on dozens of utility-specific nuances across 13 states and rigorously tested to ensure it reflected Arbor’s voice, policies, and edge cases. ## Setting guardrails ### No endless loops. Only a clear handoff. Lorikeet was configured with clear escalation logic for sensitive workflows like rate discrepancies and product bugs — where emotional tone and financial stakes make AI self-awareness and discernment critical. When confidence drops or policies require human judgment, the agent self-escalates cleanly, avoiding loops or vague replies — and routing tricky account issues to the right human at the right time. ## The results ### Lorikeet performs at the same level as Arbor’s human agents Customers praise “Olive” by name in Trustpilot reviews, often unaware they were chatting with an AI agent. What stands out isn’t just speed or tone — it’s nuance. Olive explains complex energy concepts, rate switching, and user-specific rate histories and savings in plain terms customers can understand. And when cancellation requests come in, Olive handles unique edge-case workflows with the same accuracy and empathy — turning support into a moment of trust and a reason to stick with Arbor. > “We believe deeply in the value we offer, and we don’t want customers to miss out just because they didn’t have the context. Sometimes, all it takes is the right conversation to help them stay." --- # How Summ Manages Tax-Time Spikes URL: https://www.lorikeetcx.ai/customer-stories/how-summ-efficiently-handles-support-spikes-during-tax-time **Learn how Summ (prv. Crypto Tax Calculator) uses Lorikeet AI to handle massive support spikes during tax season—maintaining speed, accuracy, and CX quality.** **97% faster resolutions. Tax-time panic solved.** Lorikeet’s AI agent is fantastic at connecting customers to the information they want and providing them with near-immediate answers. Its responses are surprisingly brilliant, answering questions in ways we hadn't considered based on its training on our knowledge base. We learn from them, and our customers get answers much faster. It’s a win/win. ## The challenge ### The challenge Summ (prv. Crypto Tax Calculator) simplifies complex cryptocurrency tax reporting across multiple global markets, including Australia, the US, UK, Canada, Europe, and many other countries. The rapid growth of cryptocurrency trading has created a surge in customer support requests, often involving intricate and technical tax-related questions. With customer inquiries ranging from simple FAQs to complex troubleshooting issues that could span weeks or even months, Summ customer support team has faced overwhelming volumes, particularly during peak tax seasons. Despite an extensive internal knowledge base, users have struggled to find the right information and specific answers to their questions due to their complexity and the sheer volume of scenarios. ## The solution ### The solution Summ looked to Lorikeet to handle both the progressive increase in support volume and sharp spikes during multiple tax seasons spread across the countries they serve. Lorikeet integrated seamlessly with their existing Intercom helpdesk, significantly easing initial deployment. After some quick training on their existing knowledge base, they were able to deploy Lorikeet’s AI agent to immediately start answering common FAQs. The efficiency gain for the team and improvement to the customer experience were instant, but Lorikeet’s value really shone when it came to automating critical, previously highly manual workflows. The ability to automate many of the steps involved in processing refunds, like checking eligibility and escalating approved cases for processing, saw Lorikeet quickly become an incredibly valuable extension of SUMM’s small but mighty support team. ## The result ### The result Lorikeet drastically reduced chat and email first-response times from approximately 30 minutes with human agents to under 1 minute. Not only was Lorikeet’s AI agent able to surface the correct information to answer customers’ questions 97% faster, it did so with surprisingly personalized and insightful responses. Its handling of email and chat tickets quickly became an additional onboarding and training resource for new human agents joining the team. Additionally, its automation significantly relieved the support team from routine tasks, allowing them to focus on more complex cases. By adding Lorikeet to their human team, SUMM has been able to improve overall operational efficiency and resource allocation, dramatically easing the tedious and difficult task of forecasting support capacity and coverage with surges in volume during tax seasons. We call that a win/win. --- # How Breeze Scaled Support 40% with Lorikeet AI URL: https://www.lorikeetcx.ai/customer-stories/how-breeze-used-lorikeet-to-build-for-support-scale-from-the-start **Discover how Breeze resolved 40% of complex support tickets autonomously in 30 days using Lorikeet—scaling CX without scaling headcount.** "We were especially impressed by the way Lorikeet integrates with the tools we already use, saving us valuable time and effort." ## The challenge ### The challenge Breeze helps businesses and consumers convert fiat to stablecoin, and stablecoin to fiat. With best-in-class compliance, accounting, analytics, and insights, Breeze brings the best of web2 into the world of web3. As a scaling business, we were at a pivotal point in figuring out how to handle customer support while still focusing on growing our company. With a small team, it was impossible to manage everything manually, and we knew we needed a system that was secure and could scale as we grew. ## The solution ### The solution Lorikeet helped us build a highly automated customer support operation from the get go. With their workflow builder, we were able to set up a system that worked seamlessly with our existing tools, making it easy to handle more customer queries without adding extra staff or feeling overwhelmed. ## The result ### The result In 30 days Lorikeet’s agent has been able to independently resolve 40% of our complex support volume (including >90% independent resolution of the tickets it chooses to solve), including questions about KYC reviews, transaction statuses and decline. Thanks to Lorikeet, we’ve been able to provide fast, reliable support even as our customer base grows. We’re still a small team, but Lorikeet has given us the confidence and the foundation we need to scale efficiently. --- # How Flex Uses AI to Handle 4x Chat Spikes During Rent Week URL: https://www.lorikeetcx.ai/customer-stories/lorikeet-helps-flex-scale-support-with-ai-agents-that-match-urgency-with-accuracy **Lorikeet helps Flex scale support with AI agents that match urgency with accuracy — doubling CSAT and resolving rent-related tickets at record speed.** **4x chat surge during rent week. Zero compromise on care.** “We tested AI solutions head-to-head and Lorikeet was a winner in every metric.” ## Key Stats - 2x — CSAT compared to previous support tool - 4x — Chat volume during rent week vs the rest of the month - 50% — Decrease in median conversation duration to get to resolution ## Why Lorikeet ### For Flex, AI accuracy isn't a feature, it's the job Flex helps renters stay on track by letting them split monthly payments and avoid late fees. But rent is more than a transaction — it’s tied to housing security. That means high urgency, high emotion, and a high volume of support tickets a month. Flex began the search after deciding to migrate off Ada, whose one-size fits-all flows couldn’t match the policy precision renters rely on. In a space where timing and rule accuracy directly affect housing stability, guessing was not an option. They needed an AI system that could stay calm under pressure, follow strict rules, and know when to escalate. Flex put Lorikeet in a head-to-head test against Decagon — and Lorikeet won across every metric that mattered. It didn’t just reply fast; it got it right. Trained on Flex’s policies, workflows, and tone, Lorikeet now handles the bulk of high-volume tickets and escalates with context when needed — giving renters clear answers, and Flex peace of mind. ## Deep integration ### From Day 1, Flex & Lorikeet worked side by side Integrations are where AI rollouts usually get messy. But in this case? The setup was easy. Lorikeet plugged straight into Flex’s existing API and authentication flow — tapping into 40+ customer attributes across account, transactions, property integration, and feature availability to deliver hyper-personalized responses. From there, Lorikeet ingested SOPs and backend data sources quickly and efficiently. What started as a fast integration became a shared learning loop — one that surfaced content gaps, trimmed duplication, and made the agent smarter with every ticket. ## Agent training ### Flex onboarded Lorikeet as if it were a human agent When rent’s due, every second counts — and so does every word. Flex didn’t just need automation; they needed an AI that could interpret date and mathematical terms to best help their customers in high-stress moments. So the teams tackled edge cases together: building logic to interpret “today” vs. “Friday,” adding disambiguation prompts for vague payment questions, and flagging tickets by urgency or sentiment. The lesson? More information isn’t always better — the right information, delivered clearly, is. Lorikeet was trained like part of their team, with workflows shaped by how Flex actually operated — not just what the SOPs said. ## Setting guardrails ### No delays. No dead ends. No loops. Lorikeet was trained with strict escalation logic to flag high-risk tickets immediately, including urgency signals, emotional language, or complaint sentiment. When confidence dips or discretion is required, Lorikeet steps aside fast — never trapping someone in a loop. ## The results ### When rent’s due, chats spike. Lorikeet handles the surge. Flex needed an AI solution that could keep up with the pace and pressure of the making a monthly rent payment. With Lorikeet, customers get faster, clearer answers— and it supports Flex’s brand mission of giving people more financial control and peace of mind each month. Now the Flex support team has more time to focus on solving more complex customer challenges, and scaling their business. > “We’re excited to explore other ways Lorikeet can help us grow Flex — like through email or voice support. That way our human agents can focus on solving more complex customer challenges.” --- # How Lorikeet Helped Amber Achieve 99% Accurate Ticket Routing with AI URL: https://www.lorikeetcx.ai/customer-stories/how-amber-uses-lorikeet-to-get-the-right-tickets-to-the-right-agents-including-their-ai-agent **See how Amber uses Lorikeet to triage support tickets with 99% accuracy—reducing delays, boosting team efficiency, and improving customer satisfaction.** **Goodbye, manual ticket triage. Hello, 99% accurate ticket routing.** "Lorikeet is the ultimate tool for any support team looking to boost efficiency and streamline ticket management." ## The challenge ### The challenge Amber is a new kind of electricity retailer designed to unlock the value of renewables for everyone. For just $22/month, Amber gives Australian households direct access to 30-minute wholesale electricity prices - the same prices that all the big retailers are paying. The wrong teams were getting assigned issues that they couldn’t solve, causing delays and frustrating for both our team and our customers. Support requests that we receive are typically time sensitive and need to be handled by the right team in order to get timely answers and solutions for our customers. ## The solution ### The solution Utilising the triage workflow in Lorikeet meant we were able to build a custom workflow using our existing logic. Over time we were able to make improvements to the workflow which were highlighted by Lorikeet’s AI. ## The result ### The result Lorikeet’s fast ticket routing has been a total lifesaver. With 99% accuracy, we know the right person is getting the job every time, and our team has never been more efficient. The AI does the heavy lifting, so we can focus on helping customers, not tracking down misplaced tickets. --- # Linktree Scales CX With Transparent, Outcome-Driven AI URL: https://www.lorikeetcx.ai/customer-stories/learn-how-linktree-s-support-team-cut-first-response-time-to-1-minute-and-automated-complex-workflows-with-lorikeet **Learn how Linktree’s support team cut first response time to 1 minute and automated complex workflows with Lorikeet — all with full transparency and fail-safes.** **Scaling with intention means taming high-volume, time-sensitive tickets — with total control.** “Support's not about doing more with less — it's about doing better with more. With Lorikeet we’re scaling with intention.” ## Key Stats - 90% — CSAT, paid customers - 1 min — First AI response time - 90% — Subscription and billing reviews resolved ## Why Lorikeet ### With AI transparency baked in, exceptional Linktree support isn’t just possible — it’s the default Linktree helps millions of creators, entrepreneurs, and brands consolidate their digital presence — curating everything from products to live events, in one flexible link. Their support team handles urgent, high-stakes tickets — like broken QR codes during a live show or login issues derailing sales. With millions of users and time-sensitive edge cases, they needed an AI solution that could act fast and get it right. They looked at Intercom Fin, Decagon, and Assembled AI, but they lacked the auditability or control needed for high-traffic, brand-sensitive support. “If an AI gives the wrong answer, I need to know why,” said Tom, Senior CX Manager. He wanted full transparency to trace decisions, fix errors, and fine-tune outcomes to workflows. At Linktree, every second counts, and they finally found an AI partner that can keep up. Lorikeet delivers multi-step logic, granular configuration, and total visibility into how and why every decision is made so Linktree can focus on what really matters — delivering truly exceptional, high-quality experiences to their customers. ## Deep integration ### After a smooth integration, an eye-opening moment hit... Switching to Lorikeet wasn’t just a plug-and-play integration — it was a mindset shift. After quickly plugging into Intercom so Lorikeet could work alongside Linktree’s human agents, the team now had a new way of thinking about AI: not as a product tool, but as an independent, collaborative agent. ## Agent training ### When it’s all hands on deck, Lorikeet is trained to mitigate high-stakes issues Lorikeet gave the Linktree team flexibility to build SOPs that reflected how their support actually works. Early workflows included account access issue investigations (with context captured before handoff), refund and cancellation flows tied to 72-hour policies, and dynamic conversations that skip questions their customers already answered. ## Setting guardrails ### Lorikeet’s workflows give Linktree the confidence that things will be done right Linktree’s support experience has to be fast — but never at the cost of trust. Their customers rely on Linktree to drive traffic and revenue in real time. Any delay or confusion can hurt their customers’ business. Lorikeet is configured with clear guardrails: it escalates when a user asks for a human, steps back after multiple failed attempts to resolve, and screens for frustration signals before continuing. ## The results ### Lorikeet clears the blockers, so Linktree can focus on real customer impact Linktree’s support team is faster, sharper, and more focused where it counts. Because Lorikeet handles the bulk of transactional queries, often so seamlessly that customers don’t realize they’re not talking to a person — human agents now spend more time per conversation. And that’s by design. With routine tickets off their plate, agents go deeper on complex, high-impact cases — helping users get more value from Linktree. The result? Exceptional, personalized experiences, at scale. > “When Lorikeet hands a ticket over to a human, the investigation’s already been done — the human agent can take the action and get a speedier resolution.” --- # How Lorikeet Helped Magic Eden Nearly Double CSAT to Near Human Level from their previous AI solution URL: https://www.lorikeetcx.ai/customer-stories/how-nft-marketplace-magic-eden-use-lorikeet **See how NFT marketplace Magic Eden improved CSAT and scaled customer support by using Lorikeet’s AI agent for fast, high-quality and complexity resolutions.** **Customer satisfaction nearly doubled after switching to Lorikeet.** "Customer support is incredibly important to Magic Eden. We were very impressed with the CSAT scores Lorikeet delivered, but we also highly valued Lorikeet being a thought partner on helping us take our customer support to the next level." ## The challenge ### The challenge Magic Eden is a hub for all things NFT (non-fungible tokens): a launchpad for new NFT projects, a marketplace for trading (including supporting fungible swaps), customizable royalties with mobile app and multichain compatibility (Solana, Ethereum, Polygon, Bitcoin, just to name a few). Magic Eden was facing a surge in customer support ticket volumes in light of its growing position as an NFT platform leader, cryptocurrency bull market, and growing interest and usage of the Magic Eden App. They needed a solution that can address these surges while maintaining high conversation quality. Their previous AI agent solution left a lot to be desired on their customer satisfaction scores (CSAT) at ~45% vs 78% human-level. ## The solution ### The solution Through an integration with Intercom, Magic Eden was able to easily set up conversation and “FAQ”-type workflows on the Lorikeet platform, with API integrations on the platform that allowed Lorikeet to help search and confirm transactions on multiple block explorers (e.g., Solana, Ethereum) and resolve complex queries around swap issues, NFT visibility issues and more. It was also able to respond to users in multiple languages. ## The result ### The result Within the first month of rolling out, Magic Eden achieved a CSAT of 74% with the workflows it rolled out: ~30 ppts higher than Magic Eden’s previous AI Agent and 4 ppts away from human CSAT scores. > "Lorikeet’s customer satisfaction scores are now nearly on par with that of our human agents. They can help users troubleshoot step by step complex blockchain issues and can independently check “on-chain” transaction status on the 10+ blockchains we support." --- # How Eucalyptus Improved Ops with Lorikeet Topics URL: https://www.lorikeetcx.ai/customer-stories/how-eucalyptus-improved-operationally-with-topics **Discover how Eucalyptus used Lorikeet Topics to gain actionable insights, streamline support workflows, and improve operational efficiency.** **The healthtech team that 3x ticket load and still lifted CSAT by 10 points.** “Topics in Lorikeet help us stay focused on the big picture—surfacing the most impactful issues while we scale quality support with AI.” ## The challenge ### The challenge As a leading digital health provider, Eucalyptus offers products and services across multiple brands and healthcare verticals, serving a wide range of patient demographics. With hundreds of distinct support topics and varied resolution paths, they needed a strategic, data-driven approach to scale their support operations while maintaining high-quality patient experience and meeting regulatory requirements. ## The solution ### The solution Eucalyptus uses Topics in Lorikeet to scan through historical customer support interactions to surface common trends and identify key drivers of volume. They were also able to identify tickets that Lorikeet was not yet trained to handle, with Unmatched Topics. This helps them understand what matters most to their customers, how frequently specific issues occur, and where Eucalyptus could further scale their support with AI. ## The result ### The result Topics in Lorikeet continue to guide Eucalyptus as they focus on the highest-impact support areas for their patients. Unmatched Topics, in particular, helped uncover issues earlier and lead to faster resolution. For example, Eucalyptus identified an issue customers faced when applying a new discount code. By spotting this pattern early, they were able to resolve the underlying problem and send proactive communications to affected customers. --- # How Lorikeet Helped Eucalyptus Cut Response Time from 20 Hours to 90 Seconds URL: https://www.lorikeetcx.ai/customer-stories/how-eucalyptus-use-lorikeet-to-deliver-healthcare-support-that-puts-patients-first **Learn how Eucalyptus delivers fast, patient-first healthcare support with Lorikeet—cutting response times from 20 hours to 90 seconds with no quality loss.** **Deliver healthcare support that puts patients first** "We're delivering faster, more efficient support to our patients like never before, with no compromise to patient experience." ## The challenge ### The challenge Eucalyptus is a digital health provider dedicated to putting patients first. They serve hundreds of thousands of patients across Australia, UK, Germany and Japan clinics that offer comprehensive treatments and ongoing care across various health domains. With a rapidly growing patient base, the need to manage a high volume of patient support interactions across multiple digital clinics is imperative to the success of the scalability of Eucalyptus’ operations. As a trusted healthcare provider, compromising on patient satisfaction is not an option. Scaling at this pace stretched the capacity of the existing team - median first response times often sat at over 20 hours. ## The solution ### The solution Eucalyptus turned to Lorikeet to anchor a transition to higher quality, faster, and more scalable support. ## The result ### The result There were zero compromises on quality. The integration between our customer support tool and Lorikeet has been seamless and we can really see how they are revolutionizing support for teams across HealthTech. We measure CSAT closely and Lorikeet’s AI agent consistent scores as high or higher than our human agents. What’s more, Eucalpytus patients get a response from Lorikeet within 90 seconds, transforming their support experience. --- # How Eucalyptus Boosted CSAT 10pts with Lorikeet Triage URL: https://www.lorikeetcx.ai/customer-stories/how-eucalyptus-boosted-csat-10pts-with-lorikeet-triage **Learn how Eucalyptus increased CSAT by 10 points using Lorikeet’s triage workflow to route tickets accurately and resolve support issues faster.** "Lorikeet’s triaging capabilities mean urgent medical questions are escalated to clinicians and dealt with quickly. This gives us the confidence to scale our healthcare offering safely. The uplift in patient satisfaction is icing on the cake." ## The challenge ### The challenge Eucalyptus is a digital health provider dedicated to putting patients first. They serve hundreds of thousands of patients across Australia, UK, Germany and Japan clinics that offer comprehensive treatments and ongoing care across various health domains. With a wide range of complex issues coming in to the support team, Eucalyptus struggled to get the right tickets in front of the right agents. What’s more, some are medically urgent and have to be reviewed promptly by clinical staff. Manually triaging each ticket was time consuming, expensive, and error-prone. ## The solution ### The solution Eucalyptus leveraged Lorikeet’s AI to implement skills-based triage. Lorikeet reviews every ticket and routes it to the appropriate agent tier based on the topics raised, customer tone, and ticket complexity. To make triaging valuable for Eucalyptus requires inputs from a combination of customer issues & sentiment vectors; this was previously not possible with incumbents that only allow for triaging based on 1x predefined field. This allows Eucalyptus to ensure the highest stakes tickets go directly to the most experienced and effective agents. ## The result ### The result After implementing skills-based triage, Eucalyptus saw a 10ppt increase in their CSAT score. The uplift was driven by getting the most complex cases to the best agents, which was impossible to do effectively without Lorikeet’s powerful triaging logic. At the same time, they were able to guarantee medically urgent tickets reached the clinical team quickly and efficiently. --- # How Easykind Delivered 24/7 Support with Lorikeet URL: https://www.lorikeetcx.ai/customer-stories/how-easykind-delivered-24-7-patient-support-without-increasing-headcount **See how Easykind scaled 24/7 patient support using Lorikeet’s AI agent—maintaining quality care without growing headcount or increasing costs.** **Patients get help in minutes, not hours – no extra hires needed.** “What made me most excited was launching Buddie quietly, and within minutes, patients were using it and getting immediate answers – at all times of the day.” ## The challenge ### The challenge Within months of launching in 2023, easykind’s patient base had grown far beyond what two Patient Support Coordinators (PSCs) could handle. Every morning, they spent hours manually triaging the inbox, and even on their best days, the team could only promise a 24-hour response time. It was a "hustle" and "struggle" every day, according to Steph, easykind’s Head of Patient Support. Phone lines were naturally limited to 9 am - 5 pm, leaving night owl and early rise patients with nowhere to turn. Yet easykind’s goal of affordable care, delivered with five-star service, left no room for slower responses or scripted, impersonal answers. Adding headcount wasn’t a viable option, and any technology used to automate responses had to clear strict compliance hurdles: it must never reveal specific product details, it should instantly escalate tickets when patients mention potential side effects, and speak in the clinic’s warm, “mate-ish” voice. easykind needed a way to scale its service, stay compliant, and maintain its human touch without increasing the size of the Patient Support team. ## The Solution ### The Solution easykind turned to Lorikeet and were able to solve two key problems within weeks. First, they trained Lorikeet’s agent to triage emails from patients. Lorikeet now reads every incoming email, tags it as a Tier 1, 2, or 3 issue, and drops it directly into the right Patient Support Coordinators (PSCs) queue—eliminating the team’s daily (and previously, very manual) triage ritual. Second, “Buddie,” a customer-facing AI agent embedded in easykind’s logged-in patient portal, answers FAQs around the clock without compromising compliance rules enforced by the Australian government. Trained to operate under strict guardrails, Buddie will never mention a product by name, and if a patient reports concerning side effects, it immediately directs them to stop taking the medication, call emergency services, and escalates to a human agent. Lorikeet’s engineering team even built a Salesforce Service Cloud integration so easykind could keep its existing CX tech stack, proving that Lorikeet could move as fast as easykind needed. ## The result ### The result The impact was immediate. Email response times dropped 92% from 24 hours to roughly two. Patient Support Coordinators (PSCs) now arrive to a pre-sorted inbox and spend their day delivering five-star service instead of triaging emails. Call volume has begun trending down—patients who previously would try calling at 9 pm or 3 am can now chat with Buddie 24/7, getting answers on the spot. Steph estimates they would need four additional hires to match the new capacity autonomously handled by Buddie, a cost avoided entirely. Most important, service quality hasn’t suffered: Buddie‘s friendly, on-brand sign-offs (e.g., “So happy to help—have a great day!”) ensure patients continue to recieve human-quality support without putting compliance at risk. In short, easykind gained a 24/7 support arm, protected its five-star reputation, and freed its team to focus on higher-value patient care — all without adding headcount. --- # How Eucalyptus Managed Support Spikes with Lorikeet URL: https://www.lorikeetcx.ai/customer-stories/how-eucalyptus-uses-lorikeet-to-turn-chaos-into-calm-during-their-biggest-moments **Learn how Eucalyptus used Lorikeet’s AI to respond to patient crises fast—deploying complex workflows and keeping CSAT stable under pressure.** "Without Lorikeet we would have been overwhelmed with enquiries - and our patients would have been waiting hours for critical information." ## The challenge ### The challenge Eucalyptus is a digital health provider dedicated to putting patients first. They serve hundreds of thousands of patients across Australia, UK, Germany and Japan clinics that offer comprehensive treatments and ongoing care across various health domains. Changes in government regulation have meant Eucalyptus needed to make changes to patients’ treatments plans. In the same period, Eucalyptus also experienced issues in its last mile logistics network - which creates serious disruption when the delivery contains temperature sensitive and life changing medication. ‍These disruptions are genuinely concerning for patients and lead to large and unexpected spikes in ticket volume. This puts pressure on response times, support costs and CSAT. ## The solution ### The solution With the no-code workflow builder in Lorikeet, we were able to train the AI agent to handle patient complex patient questions in minutes. After one government announcement we had a complex, multi-step workflow built, tested and launched within 45 minutes of the news breaking. ## The result ### The result The rapidly deployed workflows allowed Lorikeet’s agent to handle the spike in volume smoothly. Patients were grateful for the quick and empathetic responses, strengthening their trust in the company. Response times didn’t spike, and CSAT remained stable. Automated triage enabled any medical issues arising from the outages to be quickly put in front of the clinical team. --- # How Hnry Went From Zero to 70% Automation Through the Hardest Week of the Year URL: https://www.lorikeetcx.ai/customer-stories/how-hnry-went-from-zero-to-70-automation-through-the-hardest-week-of-the-year **Hnry deployed Lorikeet on Australian tax, its hardest jurisdiction, and went from a standing start to around 70% of conversations automated in the peak week of the Australian financial year.** “We actually started with Lorikeet with the hardest jurisdiction for year-end, like the nuances of Australian tax.” ## Key Stats - 70% — of conversations automated at peak - 17,000 — conversations in the first month ## The hardest weeks ### End of financial year hits Hnry from three directions at once. Every year, Australian end of financial year hits Hnry from three directions at once. Support volume climbs as customers try to understand their tax position. Onboarding surges, because everyone waits for the new financial year to sign up, which pushes first payments, ID checks, and AML checks up alongside it. And on top of that, it’s also when the team is working through the actual mechanics of reviewing and lodging thousands of income tax returns. Alex MacAvoy, Head of Customer Operations, is blunter about the scale of it. It is, she says, the most challenging week of the year. ## The plan ### Simple questions go to the agent, judgement calls go to the team. Hnry’s view going in was that not every one of those conversations needed a person. A lot of questions have one factual answer, and when that’s the case, most customers would rather get it straight away than wait in a queue for someone to type it out. So the plan was to let the agent take those questions, the simple, repeatable ones, and free up the team for the conversations that actually need a person: where the situation’s complex enough that it needs real judgement, or where a customer just needs to talk to someone rather than be given an answer. That’s also what gives the team more time to get ahead of things instead of just reacting. A simple question gets answered on the spot, any time of day, no queue. And the time saved lets the team spot a pattern in someone’s account before it becomes a problem, or notice an expense a customer hasn’t claimed and let them know, rather than waiting for the customer to bring it up. ## Hardest jurisdiction first ### Hnry deployed Lorikeet on Australian tax first, the hardest jurisdiction it has. Australian tax rules carry enough nuance that a correct answer often depends on information only the customer holds, and Hnry relies on them to provide it accurately. As Staci puts it, between the rules, the nuances, and how quickly customers can get confused and frustrated, it adds up to the hardest support environment the company has. It is also the jurisdiction Hnry chose to deploy Lorikeet on first. Australians grow up expecting a tax return process that runs through an accountant and ends in a refund, self-employed or not. Many are counting on that refund, or at least waiting on it, which adds urgency to the conversation. New Zealanders don’t carry the same expectations, so when an Australian customer’s experience doesn’t match what they’re used to, the frustration shows up faster. That mismatch between expectation and outcome shows up clearly in one of the tickets that costs the team the most time: a customer wanting to understand why they’ve been taxed at the rate they have. Answering it means walking through how their self-employed income adds up, accounting for things like other income sources and deductions, and explaining how all of that comes together to land them on the specific rate that applies to them. The question comes up often enough, and the answer follows a consistent enough shape, that Lorikeet now handles it as a structured FAQ-style response rather than passing it to a human. ## What happened ### From a standing start to around 70% automation in the peak week. Hnry’s agent went live in mid-May 2026, from a baseline of zero automation. In the first month it handled 17,000+ support conversations. By the peak week of Australian end of financial year, Lorikeet was automating around 70% of them. The customer’s issue was closed out by the agent alone. Ticket volume climbed through late June and peaked at the turn of the financial year. Automation climbed with it, from roughly 58% across the full period to around 70% in the week the load was heaviest. ## What's next ### New Zealand and the UK are live, with a wider knowledge base next. Lorikeet is already live in New Zealand and the UK as well as Australia. The current focus is broadening the knowledge base so it can answer a wider range of questions in each market, and looking at ways to integrate it more deeply into how customers manage their accounts. --- # How Carmoola Resolves 60% of Support End-to-End with Lorikeet URL: https://www.lorikeetcx.ai/customer-stories/how-carmoola-resolves-60-of-support-end-to-end-with-lorikeet **Carmoola built Katie, an AI agent for the FCA-regulated car finance journey, lifting end-to-end resolution from 30% to 60% of inbound conversations across WhatsApp, chat and email.** "I didn’t believe it was possible, I really will be honest. But the interactions that I see every day are of another level now." ## Key Stats - 60% — of inbound conversations resolved end-to-end - +60% — uplift in conversion - 90% — of outbound conversations resolved end-to-end ## Old friends, at scale Buying a car is, for most people, the biggest purchase they make after a house. It is exciting and stressful at the same time. Carmoola, a UK fintech and car finance lender that lets people pay for their next car straight from their phone, built its whole model around that moment: customers apply in an app, learn their budget up front, and tap to pay at any of 9,000+ UK dealerships. The promise is that they walk away from every interaction feeling like an old friend, not a credit score. “When we interact with our customers, we want them to feel like old friends. We want them to feel like we’re grateful they chose Carmoola,” says Lucinda Bentley, Head of Customer Operations at Carmoola. That promise is easy to keep with a small team and a handful of customers. It gets harder with every new applicant. Carmoola had always automated where it could, with a lean team of expert problem-solvers at the core, but the technology had a ceiling. Its existing automation, built on a knowledge base and rigid, form-driven prompts, resolved about 30% of inbound questions. “Those interactions were fine, but they weren’t meaningful,” says Bentley. “They didn’t have all the other stuff that a human interaction with our customer would have.” The question was how to scale the warmth, not just the throughput, on the channel Carmoola’s customers actually reach for: WhatsApp. ## Meet Katie Carmoola went to market, ran a proof of concept with Lorikeet, and “felt like we’d really met our match,” says Bentley. The resolution-based pricing sealed it. “It’s not how our previous model worked, but it felt then that we were all incentivised towards working towards the same goal.” What they built is Katie, an AI agent that works across the regulated car finance journey. Carmoola gave her a personality that matches the brand, and the team ran with it. Bentley was sceptical going in. “I didn’t believe it was possible, I really will be honest,” she says. “But the interactions that I see every day are of another level now.” ## Answers to outcomes Katie works across WhatsApp, chat and email, inbound and outbound, around the clock. A customer sitting up in bed at 2am, comparing two cars, can ask her what changing their deposit does to their plan, a fiddly calculation, and get an instant, correct answer. “Coming into work the next day and the customers have then signed their agreements and they’re ready to buy just feels incredibly impressive,” says Bentley. This is where the complexity shows. Carmoola operates in consumer credit regulated by the UK’s Financial Conduct Authority (FCA), where an answer about affordability or a missed repayment is a decision, not a lookup. When a customer is struggling to make a repayment, Katie reasons through it. “What is the best solution that we offer this person, and then driving the conversation towards […] the right outcome for the customer,” is how Bentley describes it. The customer “comes away from it feeling like Carmoola really cares,” and “all of that solution has happened without needing human intervention.” Because Katie carries context, the experience stays continuous. She remembers a customer mentioned the car was for their wife’s birthday, and follows up on pickup day: “I’m so glad your wife will be so excited. When are you going to surprise her with the car?” Bentley calls it “stuff that we’ve dreamed about in customer support, the full picture, the full context to create a fully personalised customer experience.” Katie also reaches out. When a customer stalls partway through or needs to reapply, she makes contact and guides them through, the same agent picking up where the conversation left off, rather than a no-reply email blast. > “When we interact with our customers, we want them to feel like old friends.” ## Real results, fast From day one with Lorikeet, Katie resolved 40% of inbound conversations end-to-end, up from 30% and at a higher satisfaction level. That figure now sits at 60% of all inbound on WhatsApp, chat, and email, and Carmoola expects roughly 75% by the end of the quarter. Coach helps the team see which escalated conversations to automate next, so the number keeps climbing. “It doesn’t feel like there’s a limit to what we can do,” says Bentley. The outbound results were just as clear. Run as an A/B test, Katie’s proactive outreach lifted one of their conversion metrics by 60% and resolved 90% of those conversations end-to-end, without adding to the team’s workload. The proof Carmoola cares about most shows up on Trustpilot. “One of our goals was to get a Trustpilot review where someone mentioned Katie,” says Bentley. “The first day that we got one for Katie, it was obviously a huge celebration, and now they’re coming in pretty much daily.” Some customers ask for her by name: “We had last night, ‘Where’s Katie? I would like to talk to Katie.’” Internally, the team has been freed for the work that needs them. “It’s relieved a lot of pressure on the team,” says Bentley. They handle the edge cases, shape Katie’s responses, and design new outbound moments. “It’s like having your best agent on their best day, available all the time.” > "It’s like having your best agent on their best day, available all the time." ## What's next Carmoola is extending Katie to handle third-party contacts, building toward car recommendations, and deepening the concierge experience so the right message reaches the right customer at the right moment. "It feels like we've done so much already, but I also feel like we're just getting started," says Bentley. --- # How Firm360 Turned Its Support Stack Into a Self-Improving System With AI URL: https://www.lorikeetcx.ai/customer-stories/how-firm360-turned-its-support-stack-into-a-self-improving-system-with-ai **Firm360 runs a fortnightly loop that mines Lorikeet MCP, Slack and recent conversations for knowledge gaps, publishes the missing articles, and lets Auto QA keep them honest, so the gaps shrink with every cycle.** "With Lorikeet I didn't have to make that choice. Lorikeet takes the edge off the support volume, then the team can absorb the rest." ## Headcount questions ### The headcount question Firm360 started how the best businesses often do: someone fed up with the tools they had, with an idea for a better solution, and a small group of like-minded professionals willing to try it first. What began as a tool one accountant needed for his own firm has grown into a platform used by accounting firms across the United States. Today, the company faces a familiar challenge: increases in customer volume require increases in team capacity. When Gonzalo Cayo joined a year ago to structure operations, the choice was a Customer Success Manager to drive renewals and upsell, or dedicated support agents to take tickets off the people doing onboarding and customer success. ## Self-improving ### A self-improving knowledge base The most distinctive thing about how Firm360 runs Lorikeet is what happens to the knowledge base. The team treats Lorikeet as a contextual data layer, not a frontend. Every two weeks, Firm360 runs analysis against three sources at once: Lorikeet’s MCP server, Slack, and recent customer conversation transcripts. The output is a ranked list of knowledge base gaps with draft answers attached. Once new articles are published, Lorikeet’s Auto QA reviews them continually. The next time a similar conversation comes in, Lorikeet handles it end-to-end. With each cycle the gaps decrease. This is the loop that compounds. Most AI support deployments stall when a knowledge base isn’t updated, because nobody owns the work of finding gaps. At Firm360, the work of finding gaps is automated, and the work of closing them is heading the same way. > “The MCP server works really, really well. Right off the bat it works. The gaps get less and less every time.” ## Lorikeet for operations ### From support to operations At Firm360, the teams Lorikeet assists are not support agents. They are onboarding specialists, CSMs, and ops leads, all answering tickets between other priorities. By helping reduce the load, Lorikeet acts as both customer operations and customer experience. It makes sense that the next use case is internal. Lorikeet also lives in Slack, where the Firm360 sales team can ask product questions during demos. Months of historical Q&A in that channel become the seed context for an internal-facing knowledge base that the sales team can talk to in real time to improve and scale capability without scaling the team. > “Now the team can ask Lorikeet, get responses in the moment, and immediately act on customer needs. Hiring is a full-time job, onboarding is a full-time job. Lorikeet makes everyone more efficient.” --- # How Wonderschool Answers Every Parent Call With AI URL: https://www.lorikeetcx.ai/customer-stories/how-wonderschool-answers-every-parent-call-with-ai **Wonderschool put Lorikeet on its inbound parent line and moved the answer rate from roughly one call in ten to every call, while absorbing the scam traffic that made up about half of inbound volume.** “After one month live, our voice AI is answering every single parent call. 100% of calls answered and handled, where answer rate used to be around 10%.” ## Key Stats - 100% — of parent calls answered, up from around 10% - ~50% — of inbound volume was scam noise, now absorbed with zero triage time - 0% — of previous scam calls make it to the support team. ## The challenge Wonderschool runs the marketplace and operating system that connects families to childcare across the US. Thousands of providers, parents in every state, and one daily moment that matters more than any other: a parent picking up the phone to ask whether there’s a spot for their child this week. Until early 2026, most of those calls hit voicemail. The Helping Hands line, Wonderschool’s inbound parent support channel, was answering roughly one in ten. Nine in ten parents either hung up, left a message no one had time to return, or moved on to another option entirely. The genuine ones were buried inside a flood of Google verification scam calls that Wonderschool’s triage team would have to filter through before reaching a real parent ready to enroll. For Wonderschool’s providers, the lost calls were enrollment intent walking out the door. For Wonderschool’s growth team, they were invisible. ## Answering every call In early March 2026, Wonderschool deployed Lorikeet as the voice agent on the Helping Hands line. Within the first full month live, the answer rate moved from 10% to 100%, with Lorikeet picking up every parent call. Inside a single call, Lorikeet chains across multiple decisions: it qualifies the parent, queries Wonderschool’s provider database for the specific school, evaluates whether that school has open spots, and either shares enrollment details directly, texts an enrollment link, or hands off to Wonderschool’s outbound SMS workflow so the school can follow up. If the requested provider is full or unresponsive, the agent surfaces alternative childcare nearby. The whole orchestration happens in one conversation, not five. ## A contextual agent The Helping Hands results come from Lorikeet handling the full parent call path within one conversation. Unlocking this meant going beyond standard knowledge base reference, ensuring Lorikeet had access to datapoints throughout Wonderschool’s system. The result was an agent built against live context rather than static knowledge. Wonderschool’s engineering team worked with Lorikeet to ensure the agent had the necessary endpoints to read live availability, hours, pricing, and contact information back to the parent in natural speech, decide whether to send an enrollment link by SMS, and choose when to hand off to the school. A single call can chain across three or four tool calls without breaking the conversational flow. > “Lorikeet MCP is super helpful - it’s great and very easy to work with.” ## Scam calls cut in half That same contextual architecture turned an endemic inbound scam problem into a non-event. Anyone with a Google maps number will be aware of the verification scam - a bot that calls your number claiming to be Google asking you to update your number. These scams are roughly half of inbound volume on the line. Lorikeet handles all of them and closes them out before anyone sees them. Those calls never reach a queue or waste a minute of triage time, so Wonderschool’s team focuses entirely on parents who actually want to enroll. Across the first month, the share of calls coming from genuine parent enrollment inquiries more than doubled, climbing from 14% in week one to 30% by month-end. Hang-ups fell from 43% to 9% as the opening script, voice, and pacing tightened through the first two weeks. Of every genuine parent call, 85% reached a useful outcome inside the same conversation. 41% were fully resolved by Lorikeet and a further 44% were handed off cleanly to the school via outbound workflows. For Wonderschool’s providers, the operational result is that no inbound parent contact gets dropped. The growth team is now seeing rising enrollment applications in the states the Helping Hands line covers. > "Lorikeet absorbs around 50% of calls that were scam noise. Zero human triage time wasted." ## Where complex helps Voice support for a childcare marketplace doesn’t tolerate generic answers. A parent asking about a spot for their two-year-old next month needs the real availability, the real address, and the clear next step. Hallucinated availability, off-policy pricing, or a generic “let me take a message” defeats the purpose of answering the call at all. Two design choices keep the bar high. First, every answer Lorikeet gives is grounded in live contextual data, not in a static knowledge base. Second, when the agent doesn’t have the tools available to handle a specific request, or the question is beyond its current scope, it hands off to Wonderschool’s outbound workflows so a team member at the school takes over, with the parent’s request and contact details already captured. The handoff matters as much as the resolution. A clean handoff with full context is the difference between a parent who finishes the journey and one who never hears back. With Helping Hands in maintenance mode, Wonderschool is extending Lorikeet into provider-side support: inbound voice and email for childcare providers themselves, covering account access, availability updates, and pricing changes that still need human attention today. Past that, Wonderschool is using Lorikeet as the foundation for marketplace-side growth: outbound voice and SMS campaigns, multi-school search when a parent’s first choice is full, and improved conversion attribution across the parent funnel. With Lorikeet, Wonderschool is ramping up to service parents end-to-end across the whole lifecycle. --- # How Taptap Send Supports Remittance Customers Across 80+ Markets With AI URL: https://www.lorikeetcx.ai/customer-stories/how-taptap-send-supports-remittance-customers-across-80-markets-with-ai **Taptap Send uses Lorikeet across a remittance business spanning more than 80 markets, where a support answer often depends on corridor, currency and compliance rules that differ by country.** "Lorikeet stood out for the quality of the product. The speed of iteration is impressive." ## Key Stats - 35+ — sending markets served - 80+ — receiving markets across four continents - 15 — languages supported ### Support built for people moving money home Taptap Send is one of the largest cross-border payments platforms in the world, processing billions of dollars each year for millions of customers. Built AI-native from the ground up, it is engineered for speed, low cost and reliability, delivering money across more than 35 send markets and over 80 receive markets spanning Asia, Latin America, the Middle East and Africa. Founded by the team behind GiveDirectly and Segovia, Taptap Send is committed to bringing the cost of cross-border payments below 3%, in line with the UN Sustainable Development Goal. It keeps reducing the friction and cost of getting money into emerging markets, offering one of the widest networks of on and off-ramps, whether bank, card, mobile money or cash pick-up. Lorikeet partners with Taptap Send to help deliver that experience to customers around the world. ### Real-time context across email, voice and chat Taptap Send evaluated multiple AI customer service providers before choosing Lorikeet. The deployment runs on European-hosted infrastructure integrated with Taptap Send’s systems, with real-time customer data access across email, voice and chat. The AI agent handles remittance-specific processes, including business day calculations and compliance checks, and generates responses grounded in each customer’s profile and transaction history. When a customer gets in touch, the agent retrieves the relevant account and transaction data before responding, applying the right policy for each geography while maintaining the guardrails needed when dealing with people’s money. ### Coach turns goals into action The ease of set-up and optimisation using Lorikeet’s Coach stood out for Taptap Send. Rather than configuring endless workflows by hand, the team states its goals, gets recommendations and has Coach take action. That has sharpened their view of current performance and made it far easier to improve. > "We don't have to go in and configure endless workflows. Instead, the team can use Coach to state our goals, get recommendations, and have it take action for us." ### Reassurance during a peak period During a busy sending period, a customer messages support worried that a transfer to Senegal will not arrive in time for Eid. The AI agent retrieves her transaction instantly, sees it is on track with the local partner, and reassures her that it should arrive within 24 hours. When she checks back the next day, the agent recognises her, recalls the earlier conversation and confirms the money was delivered that morning. She gets specific, personalised updates without re-explaining her situation or waiting in a queue. As Taptap Send expands into new markets and corridors, AI agents are central to delivering consistent support across languages, time zones and regulatory environments, giving the company a foundation to scale its impact without sacrificing quality. > "Right now Lorikeet is handling 700 chats. I'm not reviewing them, and it's not saying anything wrong. It's a very different product." Director of Business Operations --- # Articles and Glossary # AI in Customer Service: Everything You Need to Know in 2026 URL: https://www.lorikeetcx.ai/articles/ai-in-customer-service-everything-you-need-to-know **AI customer service uses artificial intelligence to handle, route, and resolve support interactions -- evolving from FAQ chatbots to action-taking agents that close tickets without human intervention.** AI in customer service is the use of artificial intelligence to handle, route, and resolve customer support interactions across chat, email, voice, and SMS. The technology has moved beyond FAQ chatbots into action-taking agents that process refunds, update accounts, and close tickets without human intervention. According to [McKinsey](https://www.mckinsey.com/capabilities/operations/our-insights/the-next-frontier-of-customer-engagement-ai-enabled-customer-service), advanced AI deployments reduce service interactions by 40-50%. - Only 14% of customer issues fully resolve through traditional self-service, per [Gartner](https://www.gartner.com/en/newsroom/press-releases/2024-08-19-gartner-survey-finds-only-14-percent-of-customer-service-issues-are-fully-resolved-in-self-service) - most chatbots deflect rather than resolve - Action-taking AI agents achieve 55-70% first contact resolution versus 10-25% for knowledge-base-only tools - Cost per resolution drops from $8-12 to $1-3 when AI handles routine requests end-to-end - 64% of customers would prefer companies not use AI in service, per [Gartner](https://www.gartner.com/en/newsroom/press-releases/2024-07-03-gartner-reveals-64-percent-of-customers-would-prefer-that-companies-didnt-use-ai-for-customer-service) - but that number reflects bad chatbot experiences, not good AI The real divide in 2026 is not between companies using AI and companies that are not. It is between companies whose AI resolves issues and companies whose AI deflects them. The difference shows up in every metric that matters: resolution rate, handle time, cost per ticket, and customer satisfaction. ## What Are the Three Generations of AI in Customer Service? AI in customer service has evolved through three distinct phases: scripted chatbots, knowledge-base AI, and action-taking agents. Each generation addressed a different problem, and each produced measurably different outcomes. Most companies are stuck in generation two. The leaders have moved to generation three. ### Generation 1: Scripted Chatbots Decision-tree bots that match keywords to pre-written responses. They handle 5-10 specific scenarios and break the moment a customer goes off-script. These dominated 2015-2020 and trained customers to type "speak to agent" immediately. ### Generation 2: Knowledge-Base AI Retrieval-augmented systems that search help articles and generate answers. Zendesk, Freshdesk, and Intercom's Fin operate here. They handle a wider range of questions but cannot take action. When a customer says "cancel my subscription," they suggest the customer log in and do it themselves. The ceiling is 10-25% autonomous resolution. ### Generation 3: Action-Taking Agents Platforms like [Lorikeet](https://www.lorikeetcx.ai) connect directly to CRMs, payment systems, and order management tools. The AI reads and writes to these systems mid-conversation - processing refunds, updating addresses, modifying subscriptions. One agent handles the customer while others [contact third parties in parallel](https://www.lorikeetcx.ai/blog/team-of-agents). Resolution rates reach 55-70%. ## Why Does the Deflection-vs-Resolution Distinction Matter? Deflection measures tickets kept away from human agents. Resolution measures tickets actually solved. These are fundamentally different outcomes, and optimizing for one often undermines the other. A deflected customer whose problem remains unsolved contacts you again - or leaves. Deflection-focused AI creates a perverse loop. Customers who get a help article instead of a solution either give up or call back. The cost per resolution stays high because the issue was never resolved. Action-taking AI breaks this loop. When the AI processes a refund or cancels a subscription mid-conversation, the customer leaves with their problem solved. No callback. No second ticket. That is resolution - the only metric that directly reduces cost while improving satisfaction. ## How Should You Evaluate AI for Your Support Team? Skip feature checklists and vendor demos with scripted scenarios. Instead, test AI platforms against your actual top 10 ticket types and measure what percentage get resolved end-to-end without human involvement. That single test tells you more than any sales deck. 1. **Test real workflows, not demos.** Give the platform your actual top 5 ticket types. If it cannot process a refund or update an account during the trial, it will not do it in production. Target above 50% autonomous resolution on routine requests. 2. **Check integration depth.** Does it connect to your CRM, payment processor, and order management system? Read-only access means it can answer questions. Read-write access means it can take action on complex issues. The difference is the difference between generation 2 and generation 3. 3. **Demand auditability.** Can you trace exactly why the AI made each decision? Instruction-based systems with [continuous QA reviewing 100% of tickets](https://www.lorikeetcx.ai/product/coach) are auditable. Self-training black boxes are not. This matters more than most teams realize until a compliance audit happens. 4. **Measure cost per resolution, not cost per ticket.** Deflection tools charge per interaction whether they help or not. Resolution-focused platforms tie cost to outcomes. A $0.99-per-resolution charge sounds cheap until Intercom's Fin handles 10,000 monthly conversations at $9,900 in AI fees alone. ## What Results Does AI Actually Deliver? The performance gap between AI generations is not marginal. It is the difference between slightly faster human-agent workflows and fundamentally different unit economics. The numbers shift within 90 days of deployment. First contact resolution typically moves from 20-30% with traditional chatbots to 55-70% with AI agents that access backend systems. Average handle time drops from 8-12 minutes per interaction to under 3 minutes for routine requests. Cost per resolution falls from $8-12 to $1-3 when AI handles the full workflow. CSAT scores improve by 15-25 points as customers receive instant resolution instead of queue-based responses. These patterns hold across industries. [Eucalyptus automated 80% of first-response emails](https://www.lorikeetcx.ai/blog) without adding headcount. GiveCard served 300,000 people across 60,000 calls in 3 languages - deployed in 48 hours. Magic Eden saw CSAT jump 30 points within the first month of switching. ## What Should You Know Before Deploying AI in Customer Service? You do not need perfect documentation, fully mapped processes, or a complete system overhaul before deploying AI. The companies winning with AI are the ones who started before they were ready and iterated. Waiting for perfect conditions is the most expensive mistake in AI adoption. Start with your highest-volume, most repetitive ticket types - order status, refund requests, account updates. Think of AI configuration as coaching a new team member, not programming a machine. Your first prompt will not work perfectly. That is the process, not failure. Arbor got their AI agent running in a week. Summ automated refund workflows during tax season and achieved 97% faster resolutions. The role that makes this work is the CX Automation Specialist - someone who understands customer workflows and can translate them into AI instructions. ## Key Takeaways - AI customer service has evolved from scripted chatbots to action-taking agents that resolve 55-70% of tickets autonomously - Deflection and resolution are opposite strategies - optimize for resolution rate and cost per resolved issue, not tickets deflected - Test platforms on your real ticket types - any AI that cannot process a refund during trial will not do it in production - Deploy iteratively, starting with high-volume repetitive requests - you do not need perfect conditions to start **See the difference with your own data. **[Lorikeet](https://www.lorikeetcx.ai)** runs evaluations against your real ticket types - not demos with scripted scenarios.** --- # Best Customer Service Software 2026: Ranked by Resolution Rate URL: https://www.lorikeetcx.ai/articles/best-customer-service-software-in-2026 **The best customer service software in 2026 is defined by resolution rate -- the percentage of issues resolved without human intervention -- not feature count.** Customer service software is the technology stack that manages, routes, and resolves customer support interactions across channels like email, chat, phone, and social media. The best platforms in 2026 are defined not by feature count but by resolution rate - the percentage of customer issues resolved without human intervention. According to [McKinsey](https://www.mckinsey.com/capabilities/operations/our-insights/the-next-frontier-of-customer-engagement-ai-enabled-customer-service), AI-enabled customer service can reduce service interactions by 40-50%. - Only 14% of customer issues fully resolve through traditional self-service tools, per [Gartner 2024 research](https://www.gartner.com/en/newsroom/press-releases/2024-08-19-gartner-survey-finds-only-14-percent-of-customer-service-issues-are-fully-resolved-in-self-service) - AI-native platforms achieve 55-70% first contact resolution versus 10-25% for chatbot-only tools - Cost per resolution drops from $8-12 to $1-3 when AI handles routine requests end-to-end - The gap between "AI-assisted" and "AI-resolved" is the defining divide in the 2026 market Every "best customer service software" list includes the same 10 platforms ranked by features. That approach tells you nothing useful. A platform with 300 features and a 15% resolution rate costs more per outcome than a focused platform with 50 features and a 60% resolution rate. Here is a different framework: categorize platforms by what they actually do, then pick based on your specific needs. | Category | Platforms | AI Resolution | Cost/Resolution | Handle Time | Best For | | --- | --- | --- | --- | --- | --- | | AI-Native Resolution | Lorikeet | 55-70% | $1-3 | <3 min | Autonomous end-to-end resolution | | Ticket Management | Zendesk, Freshdesk, Zoho Desk | 10-25% | $6-12 | 7-8 min | Structured queue-based operations | | Conversation Platforms | Intercom, Front | 10-25% | $6-12 | 6-10 min | Product-led SaaS support | ## What Are the Main Categories of Customer Service Software? Customer service software in 2026 falls into three distinct categories: ticket management platforms, conversation platforms, and AI-native resolution platforms. Each category reflects a different philosophy about how customer support should work, and each produces measurably different outcomes. ### Ticket Management Platforms Zendesk, Freshdesk, and Zoho Desk organize support around ticket queues. Agents claim tickets, investigate issues across multiple tabs, and send responses. AI assists by suggesting replies and routing tickets. These platforms are mature, well-integrated, and optimized for human-agent workflows. Resolution depends on agent speed and availability. ### Conversation Platforms Intercom and Front organize support around ongoing conversations rather than discrete tickets. Better for SaaS and product-led teams where support is embedded in the user experience. AI handles front-line messaging. These platforms feel more modern but share the same fundamental limitation: AI that answers questions, not AI that takes action. ### AI-Native Resolution Platforms [Lorikeet](https://www.lorikeetcx.ai) represents the newest category. Instead of managing tickets or conversations, these platforms resolve issues autonomously by connecting to backend systems - CRMs, payment processors, order management. The AI reads and writes to these systems mid-conversation, executing the same actions a human agent would. The difference is not incremental. It is structural. ## How Should You Evaluate Customer Service Software? Skip feature comparison tables. Instead, measure platforms against three metrics that directly impact your bottom line: first contact resolution rate, cost per resolution, and time to resolution. Everything else is secondary. 1. **First contact resolution rate.** What percentage of customer issues get fully resolved in the first interaction without escalation? Ticket management platforms typically hit 10-25% with AI. AI-native platforms reach 55-70%. This single metric determines your staffing needs. 2. **Cost per resolution.** Not cost per ticket - cost per resolved issue. Deflection-based tools charge per interaction whether they help or not. Resolution-based tools tie cost to outcomes. Target: under $3 per resolved issue for routine requests. 3. **Integration depth.** Does the platform read from your systems or read and write? Knowledge-base-only integration means AI can answer "where is my order?" but cannot process a return. Backend integration means AI handles the full workflow. 4. **Auditability.** Can you trace exactly why the AI made each decision? For regulated industries, [instruction-based systems with continuous QA](/articles/what-does-qa-mean-in-customer-service) are auditable. Self-training black boxes are not. This matters more than most teams realize until an audit happens. ## What Results Separate the Categories? The performance gap between categories is not marginal. It is the difference between AI that helps agents work slightly faster and AI that eliminates the need for agents on routine work entirely. Ticket management platforms (Zendesk, Freshdesk) reduce handle time by 15-25% through AI suggestions, bringing average handle time from 10 minutes to 7-8 minutes. Cost per resolution stays in the $6-12 range because humans still do the work. AI-native platforms like Lorikeet achieve handle times under 3 minutes for routine requests, with cost per resolution of $1-3. CSAT scores typically improve 15-25 points within 90 days as customers receive instant resolution instead of queue-based responses. These are not vendor claims. Lorikeet's published customer data shows consistent results: [GiveCard served 300,000+ people across 60,000 calls](https://www.lorikeetcx.ai/blog), and Eucalyptus automated 80% of first-response emails. ## Key Takeaways - Customer service software splits into 3 categories: ticket management, conversation platforms, and AI-native resolution - Evaluate platforms on resolution rate and cost per resolution, not feature count or marketplace size - AI-native platforms achieve 55-70% autonomous resolution at $1-3 per issue versus $6-12 on ticket management tools - For regulated industries, choose auditable instruction-based AI over self-training black-box models --- # Zendesk vs Intercom 2026: Pricing, AI, and What Both Get Wrong URL: https://www.lorikeetcx.ai/articles/zendesk-vs-intercom **Zendesk is a ticket management system built for structured, agent-centric workflows. Intercom is a messaging-first platform focused on conversational automation. Neither fully resolves tickets autonomously.** Zendesk and Intercom represent two different philosophies in customer support. Zendesk is a ticket management system - structured, queue-based, agent-centric. Intercom is a messaging-first platform - conversational, proactive, automation-forward. Zendesk has deeper enterprise features. Intercom has a more modern interface and stronger product-led support capabilities. Both charge premium prices, and both add AI as a layer on top rather than building resolution into the core. - Zendesk starts at $55/agent/month (Suite), Intercom starts at $29/seat/month but scales steeply with add-ons - Intercom's Fin AI handles front-line conversations but charges $0.99 per AI resolution on top of seat costs - Zendesk treats support as ticket queues, Intercom treats it as ongoing conversations - Neither connects to backend systems for autonomous resolution - both primarily surface knowledge base content This comparison matters because teams evaluating Zendesk alternatives often land on Intercom as the "modern" option. The interface is slicker. The messaging model feels more natural. But under the surface, both platforms face the same fundamental limitation: their AI can answer questions about your product, but it cannot take action on behalf of your customers. | | Zendesk | Intercom | Lorikeet | | --- | --- | --- | --- | | Starting price | $55/agent/mo | $29/seat/mo | Per resolution | | AI pricing | $50/agent/mo flat | $0.99/resolution | Per resolution (included) | | AI approach | Agent assistance | Front-line conversations | Autonomous resolution | | Support model | Ticket queues | Messaging/conversations | End-to-end resolution | | In-app messaging | Third-party required | Native | Chat, email, voice, SMS | | Enterprise compliance | Strong | Basic | Auditable instructions | | Reporting depth | Deep (Explore) | Basic | Full audit trail | | Backend system access | No | No | Read + write (CRM, payments, OMS) | ## Where Does Zendesk Beat Intercom? Zendesk outperforms Intercom on structured ticket management, enterprise compliance features, and complex routing for large support operations. If your team handles 50,000+ tickets per month across multiple brands and channels with strict SLA requirements, Zendesk's queue management and escalation workflows are more mature. Zendesk's reporting suite (Explore) offers deeper analytics than Intercom's. Its role-based access controls and audit logging meet enterprise compliance needs that Intercom's simpler permission model does not fully address. For teams with dedicated support operations managers who build complex automated workflows, Zendesk provides more customization surface area. ## Where Does Intercom Beat Zendesk? Intercom wins on proactive messaging, in-app support, and product-led customer engagement. Its messenger widget is embedded directly in your product, enabling contextual help that triggers based on user behavior. If your support strategy includes onboarding flows, in-app tours, and proactive outreach, Intercom does this natively where Zendesk requires third-party tools. Intercom's **Fin AI agent** handles first-line conversations with more conversational fluency than Zendesk's AI. It resolves simple questions effectively by drawing from your help center. The per-resolution pricing ($0.99/resolution) means you only pay when Fin actually answers - though costs become unpredictable at high volumes. For SaaS companies where support is part of the product experience, Intercom's approach often fits better. ## How Do Their AI Capabilities Compare? Zendesk's AI focuses on agent assistance - triaging tickets, suggesting responses, detecting intent and sentiment. Intercom's Fin takes a front-line role, directly answering customer questions before routing to humans. Both rely on knowledge base content for answers. Neither connects to CRMs, payment systems, or order management to take action. Fin's per-resolution pricing creates a different cost dynamic than Zendesk's per-agent AI add-on. At low volumes, Fin is cheaper. At high volumes (10,000+ monthly conversations), costs can exceed Zendesk's flat per-agent model. The deeper issue is shared: per [Gartner](https://www.gartner.com/en/newsroom/press-releases/2024-08-19-gartner-survey-finds-only-14-percent-of-customer-service-issues-are-fully-resolved-in-self-service), only 14% of issues resolve through self-service, and knowledge-base-only AI does not meaningfully change that number. Platforms like [Lorikeet](https://www.lorikeetcx.ai) address this by connecting AI directly to operational systems - reading and writing to CRMs, processing payments, and executing multi-step workflows autonomously. ## What Does the Total Cost Actually Look Like? Both platforms have pricing that looks simple on the surface but scales unpredictably. Comparing total cost requires modeling your actual usage patterns, not just per-seat rates. Zendesk's Suite Team costs $55/agent/month. Advanced AI adds $50/agent. For 20 agents, that is $25,200/year at full AI capability. Intercom's Essential plan starts at $29/seat/month, but Fin AI adds $0.99 per resolution. If Fin resolves 5,000 conversations monthly, that is $4,950/month in AI costs alone - $59,400 annually - on top of seat fees. Intercom becomes significantly more expensive than Zendesk at scale. Teams with high conversation volumes should model both platforms against their actual ticket numbers before committing. ## Key Takeaways - Zendesk suits structured ticket operations with complex routing - better for 50+ agent enterprise teams - Intercom suits product-led SaaS companies where support lives inside the product experience - Intercom's per-resolution AI pricing can exceed Zendesk's flat model at high volumes - model your numbers - Neither platform resolves issues through backend systems - both primarily answer questions from knowledge bases --- # Zendesk vs Freshdesk 2026: Pricing, AI, and Resolution Compared URL: https://www.lorikeetcx.ai/articles/zendesk-vs-freshdesk **Zendesk and Freshdesk are both ticket management platforms -- Zendesk has deeper enterprise features, Freshdesk offers similar core functionality at roughly 70% lower starting cost. Neither resolves tickets autonomously.** Zendesk and Freshdesk are both ticket management platforms used by customer service teams worldwide. Zendesk is the market incumbent with deeper enterprise features and a larger app marketplace. Freshdesk offers similar core functionality at roughly 70% lower starting cost. Neither platform autonomously resolves tickets through backend systems - both primarily route issues to human agents, with AI that suggests responses rather than executes actions. - Zendesk starts at $55/agent/month (Suite Team), Freshdesk at $15/agent/month (Growth) - a 3.5x price gap at entry level - Both platforms achieve 10-25% autonomous resolution rates with their native AI features - Zendesk's app marketplace has 1,500+ integrations versus Freshdesk's 1,000+ in its marketplace - For teams needing AI that acts on backend systems, neither platform resolves the core gap - consider AI-native alternatives This comparison exists because Freshdesk is the most common Zendesk alternative teams evaluate first. The pricing difference is real. But price alone does not determine value. The right question is whether either platform will meaningfully change your team's resolution rate and cost per ticket - or whether you are comparing two versions of the same approach. | | Zendesk | Freshdesk | Lorikeet | | --- | --- | --- | --- | | Starting price | $55/agent/mo | $15/agent/mo | Per resolution | | Free tier | No | Yes (2 agents) | No | | AI add-on cost | $50/agent/mo | Included at lower tiers | Included | | AI resolution rate | 10-25% | 10-25% | 55-70% | | Backend system access | Read only | Read only | Read + write | | Marketplace integrations | 1,500+ | 1,000+ | API + MCP servers | | Agent onboarding time | 3-5 days | ~1 day | No agents needed for routine work | | Best for | Enterprise 100+ agents | SMB teams on a budget | Teams wanting autonomous resolution | ## Where Does Zendesk Beat Freshdesk? Zendesk wins on enterprise features, marketplace breadth, and workflow customization depth. If your team runs complex multi-brand, multi-language support with intricate routing rules, Zendesk's maturity shows. Its reporting is deeper, its API is more flexible, and its ecosystem of third-party tools is larger. Zendesk's **Advanced AI add-on** ($50/agent/month extra) includes intent detection, sentiment analysis, and automated triage that Freshdesk's AI does not match in sophistication. For large teams with 100+ agents, Zendesk's admin controls and permission management handle enterprise complexity better. Its sandbox environments allow testing workflow changes before pushing to production - a feature Freshdesk lacks at most tiers. ## Where Does Freshdesk Beat Zendesk? Freshdesk wins on price-to-value ratio, setup simplicity, and time to productivity for small and mid-sized teams. A 20-agent team pays roughly $3,600/year on Freshdesk Growth versus $13,200/year on Zendesk Suite Team - a $9,600 annual difference before AI add-ons. Freshdesk's interface requires less training. Most agents are productive within a day versus 3-5 days on Zendesk. The platform includes a free tier for up to 2 agents, which Zendesk eliminated years ago. For teams that primarily handle email and chat support without complex routing needs, Freshdesk delivers 80% of Zendesk's functionality at a fraction of the cost. ## How Do Their AI Features Compare? Both platforms offer AI-powered features, but neither fundamentally changes how tickets get resolved. Zendesk's AI suggests responses, classifies tickets, and routes them to the right team. Freshdesk's Freddy AI does similar work at a smaller scale. Neither AI connects to backend systems to process refunds, update accounts, or execute multi-step workflows autonomously. Zendesk's AI is more sophisticated in intent detection and can handle more nuanced routing decisions. Freshdesk's AI is simpler but included in lower tiers without the $50/agent add-on. According to [Gartner](https://www.gartner.com/en/newsroom/press-releases/2024-08-19-gartner-survey-finds-only-14-percent-of-customer-service-issues-are-fully-resolved-in-self-service), only 14% of customer issues fully resolve in self-service - a limitation shared by both platforms. For teams needing AI that takes action beyond suggesting articles, platforms like [Lorikeet](https://www.lorikeetcx.ai) connect directly to CRMs, payment systems, and order management to resolve issues end-to-end. ## What Does the Total Cost Actually Look Like? Per-agent pricing tells a partial story. Total cost of ownership includes base licenses, AI add-ons, marketplace apps, admin hours, and implementation. Here is how the numbers play out for a 25-agent team over 12 months. Zendesk Suite Team runs $16,500 annually at base pricing. Adding Advanced AI ($50/agent) brings that to $31,500. Factor in 1-2 premium marketplace apps at $5-15/agent and an estimated 10 admin hours per week for workflow maintenance. Freshdesk Growth costs $4,500 annually for the same team. Its AI features are included at lower tiers, and admin complexity is lower, cutting maintenance hours roughly in half. The total gap narrows when you factor in Freshdesk's limitations on reporting and enterprise features, but Freshdesk typically costs 40-60% less when comparing equivalent setups. ## Key Takeaways - Zendesk wins on enterprise depth and AI sophistication - worth the premium for 100+ agent teams with complex needs - Freshdesk delivers 80% of Zendesk's core features at 40-60% lower total cost for small to mid-sized teams - Neither platform autonomously resolves tickets through backend systems - both cap around 10-25% AI resolution - Consider whether you are optimizing within the ticket-routing model or need a fundamentally different approach --- # Zendesk Alternatives in 2026: Ticket Routers vs Ticket Resolvers URL: https://www.lorikeetcx.ai/articles/zendesk-alternatives-in-2026 **The best Zendesk alternatives in 2026 fall into two categories: ticket routers that organise work for humans, and ticket resolvers that close issues autonomously using AI agents.** Zendesk alternatives are customer support platforms that replace or supplement Zendesk's ticketing, automation, and AI capabilities. The most common reasons teams switch: pricing that scales faster than value, AI features that deflect rather than resolve, and integration complexity that requires dedicated admins. According to a [2024 Gartner survey](https://www.gartner.com/en/newsroom/press-releases/2024-08-19-gartner-survey-finds-only-14-percent-of-customer-service-issues-are-fully-resolved-in-self-service), only 14% of customer issues fully resolve in self-service - a stat that applies to most legacy helpdesks, Zendesk included. - Zendesk pricing starts at $55/agent/month for Suite Team but climbs to $115+ for AI add-ons and advanced features - AI-first alternatives like Lorikeet price per resolution, not per seat - aligning cost with outcomes - The core question is whether you need a ticket router or a ticket resolver - Switching costs are real but often lower than the annual overhead of maintaining a Zendesk instance with multiple add-ons Zendesk is not a bad product. It is a mature ticket management system with a deep feature set. But "mature" also means legacy architecture, complex pricing tiers, and AI capabilities bolted on rather than built in. If you are evaluating alternatives, the question is not "which platform has the most features?" It is "which platform will resolve the most customer issues without human intervention?" | Alternative | Type | Starting Price | AI Resolution | Backend Access | Best For | | --- | --- | --- | --- | --- | --- | | Lorikeet | AI-native | Per resolution | 55-70% | Read + write | Autonomous end-to-end resolution | | Freshdesk | Ticket management | $15/agent/mo | 10-25% | Read only | Budget-conscious SMB teams | | Help Scout | Ticket management | $22/user/mo | 10-25% | Read only | Email-first small teams | | Zoho Desk | Ticket management | $14/agent/mo | 10-25% | Read only | Zoho ecosystem users | | Intercom | Conversation | $29/seat/mo | 10-25% | Read only | Product-led SaaS companies | ## Why Do Teams Leave Zendesk? Teams leave Zendesk for three consistent reasons: escalating costs as they add AI features and premium tiers, AI automation that deflects to help articles rather than resolving issues, and admin complexity that requires dedicated staff to maintain workflows, triggers, and integrations. Zendesk's pricing model charges per agent per month, which means costs scale linearly with headcount. Adding AI capabilities through their Advanced AI add-on pushes per-agent costs above $100/month. For a 50-agent team, that is $60,000+ annually before factoring in implementation and admin overhead. Teams paying that much expect AI that resolves tickets - not AI that suggests articles. ## Which Zendesk Alternatives Actually Resolve Issues? The alternatives split into two categories: platforms that replicate Zendesk's ticket management with different pricing, and platforms that fundamentally change how tickets get resolved. The first group includes Freshdesk, Help Scout, and Zoho Desk. The second includes [Lorikeet](https://www.lorikeetcx.ai) and newer AI-native platforms. ### Ticket Management Alternatives **Freshdesk** offers similar functionality to Zendesk at lower price points, starting at $15/agent/month. Good for teams that need basic ticketing without Zendesk's complexity. **Help Scout** focuses on simplicity and email-first support, starting at $22/user/month. **Zoho Desk** integrates well with the Zoho ecosystem at $14/agent/month. All three manage tickets well but share Zendesk's fundamental limitation: AI that assists agents rather than replacing routine work. ### AI-Native Alternatives Lorikeet takes a different approach. Instead of routing tickets to humans, its AI connects directly to CRMs, payment systems like Stripe and Adyen, and order management tools. It reads and writes to these systems mid-conversation - processing refunds, updating accounts, checking order status. One agent handles the inquiry while others [contact third parties](https://www.lorikeetcx.ai/blog) in parallel. The system uses explicit, auditable instructions rather than self-training models. ## How Do the Alternatives Compare on Results? Comparing helpdesk platforms on feature checklists misses the point. What matters is measurable impact on resolution rates, handle time, cost, and customer satisfaction. Here is how the categories perform based on industry benchmarks and published customer data. Ticket management alternatives (Freshdesk, Help Scout, Zoho) typically achieve autonomous resolution rates of 10-25% - similar to Zendesk. Handle times stay at 6-10 minutes per interaction. Cost per resolution ranges from $6-12 depending on agent salaries and ticket volume. AI-native platforms like Lorikeet push autonomous resolution to 55-70%, with handle times under 3 minutes and cost per resolution of $1-3 according to [McKinsey research](https://www.mckinsey.com/capabilities/operations/our-insights/the-next-frontier-of-customer-engagement-ai-enabled-customer-service) on advanced AI deployments. ## What Should You Look for When Switching From Zendesk? Migrating from Zendesk requires evaluating four dimensions: data portability, integration compatibility, AI capability depth, and total cost of ownership. Do not just compare monthly per-agent pricing - factor in implementation time, add-on costs, and the admin hours your current Zendesk instance demands. 1. **Data migration path.** Check whether the alternative supports importing Zendesk ticket history, macros, and knowledge base content. Most major platforms offer migration tools or services. Budget 2-4 weeks for a clean migration. 2. **Integration depth.** Zendesk connects to 1,500+ apps through its marketplace. But count how many you actually use. Most teams rely on 5-10 integrations. Verify these work with the alternative before committing. 3. **AI that acts, not just suggests.** Test whether the AI can complete your top 5 ticket types end-to-end. Reading a knowledge base is not the same as processing a refund. 4. **Total cost of ownership.** Add agent licenses, AI add-ons, admin hours, and integration maintenance. Zendesk's "per agent" pricing often doubles when you add features teams actually need. ## Key Takeaways - Zendesk alternatives fall into ticket routers (Freshdesk, Help Scout) and ticket resolvers (Lorikeet) - know which you need - AI-native platforms achieve 55-70% autonomous resolution versus 10-25% for traditional helpdesks - Compare total cost of ownership, not per-agent pricing - Zendesk add-ons can double the sticker price - Test your top 5 ticket types end-to-end before committing to any platform switch --- # Best Help Desk Software 2026: Ranked by Resolution Rate URL: https://www.lorikeetcx.ai/articles/best-helpdesk-software-in-2026 **The best helpdesk software in 2026 goes beyond ticket management -- it takes action: processing refunds, updating accounts, and resolving issues without human intervention.** Helpdesk software is the operational backbone of customer support - the system where tickets are created, routed, tracked, and (ideally) resolved. The best platforms in 2026 go beyond ticket management. They take action: processing refunds, updating accounts, and resolving issues without human intervention. According to [Gartner](https://www.gartner.com/en/newsroom/press-releases/2024-08-19-gartner-survey-finds-only-14-percent-of-customer-service-issues-are-fully-resolved-in-self-service), only 14% of customer issues fully resolve in self-service tools. - AI-capable helpdesks now resolve 40-70% of inbound tickets autonomously, per [McKinsey research](https://www.mckinsey.com/capabilities/operations/our-insights/the-next-frontier-of-customer-engagement-ai-enabled-customer-service) on advanced AI deployments - Average cost per resolution drops from $8-12 (human agent) to $1-3 with AI-first helpdesk platforms - First contact resolution rates above 55% separate action-taking platforms from deflection-based chatbots - The right helpdesk should integrate with your CRM, payment tools, and order management - not just your knowledge base Most "best helpdesk software" lists rank platforms by feature count. That tells you almost nothing about whether the tool will actually resolve your customers' problems. A helpdesk with 200 features but a 15% automation rate is worse than one with 50 features and a 60% resolution rate. Here is what to actually look for - and which platforms deliver. | Platform | Category | Starting Price | AI Resolution Rate | Cost/Resolution | Backend Access | | --- | --- | --- | --- | --- | --- | | Lorikeet | Action-taking | Per resolution | 55-70% | $1-3 | Read + write | | Zendesk | Knowledge-base-first | $55/agent/mo | 10-25% | $6-12 | Read only | | Freshdesk | Knowledge-base-first | $15/agent/mo | 10-25% | $6-12 | Read only | | Intercom | Knowledge-base-first | $29/seat/mo + $0.99/resolution | 10-25% | $6-12 | Read only | | Zoho Desk | Knowledge-base-first | $14/agent/mo | 10-25% | $6-12 | Read only | ## What Makes Helpdesk Software "Best" in 2026? The best helpdesk software resolves issues end-to-end without requiring human handoff for routine requests. That means reading and writing to backend systems like CRMs, payment processors, and order management tools - not just surfacing help articles. Resolution rate matters more than feature count. Traditional helpdesks were designed as ticket routers. Agent receives ticket, reads it, opens 4 other tabs, copies data between systems, sends a response. Modern platforms compress that into seconds. The shift is from **ticket management** to **ticket resolution**. Companies using AI-powered service platforms see handle times drop significantly - from industry averages of 8-12 minutes to under 3 minutes per interaction for routine requests. ## Which Helpdesk Platforms Actually Take Action? Most helpdesk vendors claim AI capabilities, but there is a significant gap between answering questions from a knowledge base and actually executing multi-step workflows across backend systems. Only a handful of platforms can process a refund, update an account, and send a confirmation - all within a single conversation. ### Action-Taking Platforms [Lorikeet](https://www.lorikeetcx.ai) connects directly to CRMs, payment tools like Stripe and Adyen, and order management systems. Its AI reads and writes to these systems mid-conversation - checking order status, processing refunds, updating addresses. The platform uses explicit instructions rather than black-box learning, which matters for regulated industries that need auditable decisions. ### Knowledge-Base-First Platforms Zendesk, Freshdesk, and Intercom offer AI features that primarily surface help articles and suggest responses. They handle high volumes of simple queries well. But when a customer needs an actual account change or refund processed, these platforms typically route to a human agent. The AI assists - it does not resolve. ## What Results Can You Expect From Modern Helpdesk Software? Switching from a ticket-routing helpdesk to an action-taking platform changes your core metrics within 90 days. The improvements are measurable across resolution rates, handle time, cost per ticket, and customer satisfaction. First contact resolution typically moves from 20-30% with traditional chatbots to 55-70% with AI platforms that access backend systems. Average handle time drops from 8-12 minutes to under 2-3 minutes per interaction. Cost per resolution falls from $8-12 to $1-3 when AI handles routine requests end-to-end. CSAT scores improve by 15-25 points as customers get instant resolution instead of being told to "wait for an agent." These numbers are not theoretical. [Lorikeet's customer deployments](https://www.lorikeetcx.ai/blog) show consistent patterns: GiveCard served 300,000 people across 60,000+ calls, and Eucalyptus automated 80% of first-response emails. ## How Should You Evaluate Helpdesk Software? Skip the feature comparison matrix. Instead, test whether the platform can complete your top 10 customer requests end-to-end without human intervention. If it cannot process a refund or update an account during the trial, it will not do it in production either. 1. **Test real workflows, not demos.** Give the platform your actual top 5 ticket types. Measure what percentage it resolves completely versus deflects to a human. Target: above 50% autonomous resolution. 2. **Check system integrations.** Does it connect to your CRM, payment processor, and order management? Read-only access is not enough - it needs write access to take action on behalf of customers. 3. **Audit the AI's decision-making.** Can you see exactly why the AI took each action? Instruction-based systems are auditable. Self-training black boxes are not. This matters for compliance. 4. **Measure cost per resolution, not cost per ticket.** Deflection-focused tools charge per ticket whether they help or not. Resolution-focused tools prove value by the percentage of issues fully closed. ## Key Takeaways - AI helpdesks that access backend systems resolve 55-70% of tickets versus 10-25% for knowledge-base-only tools - Cost per resolution drops from $8-12 to $1-3 when AI handles routine requests autonomously - Test platforms on your real ticket types - demo performance rarely matches production results - Audit AI decision-making before buying, especially in regulated industries requiring [continuous QA](/articles/what-does-qa-mean-in-customer-service) --- # Which AI Support Tool Can Troubleshoot Technical Issues for Customers? URL: https://www.lorikeetcx.ai/articles/which-ai-tool-can-troubleshoot-technical-issues-for-customers **AI tools that troubleshoot technical issues need three capabilities: multi-system diagnosis, multi-step fix execution, and intelligent escalation when issues exceed their scope.** AI support tools that troubleshoot technical issues for customers are platforms capable of diagnosing problems across multiple systems, executing multi-step fixes, and knowing when to escalate to a human specialist. According to [Gartner](https://www.gartner.com/), only 14% of customer service issues fully resolve through self-service today - largely because most tools can explain a problem but cannot investigate or fix it across interconnected backend systems. - Technical troubleshooting requires querying logs, checking account states, and coordinating fixes across multiple systems - most AI tools cannot do this - [Lorikeet](https://www.lorikeetcx.ai) handles multi-step technical resolution: diagnose, verify, fix, and confirm across CRM, payment, and ticketing systems in one conversation - Self-aware AI agents escalate complex issues with full context rather than trapping customers in unhelpful loops - Teams using action-capable AI for troubleshooting see 60%+ first-contact resolution on complex issues, up from 20-30% with chatbots A customer calls about a declined card while traveling overseas. The chatbot suggests they "check with their bank." That is the entire response. Meanwhile, the customer is standing at an airport counter with no way to pay for a transfer. Technical troubleshooting in customer service is not about answering questions. It is about diagnosing what went wrong across interconnected systems and fixing it while the customer is still on the line. Most AI tools cannot do that. Here is what separates the ones that can. ## What Makes Technical Troubleshooting Different From FAQ Handling? Technical troubleshooting requires the AI to investigate a specific customer's situation across multiple systems, identify the root cause, and execute a fix. FAQ handling matches a question to a pre-written answer. The difference is between a diagnostic process and a lookup function. When a customer reports a failed payment, FAQ handling returns an article about common payment errors. Technical troubleshooting means the AI checks the customer's payment method status, queries the transaction processor for the decline code, examines whether fraud detection triggered, verifies the account is in good standing, and takes corrective action based on what it finds. Each step depends on the previous one. An [Ipsos](https://www.ipsos.com/) survey found only 35% of customers say chatbots usually solve their problem - and technical issues are where the gap is widest because they require investigation, not recitation. ## How Does Lorikeet Handle Multi-Step Technical Issues? Lorikeet handles multi-step technical issues through its Team of Agents architecture, where specialized agents work across systems simultaneously while maintaining a single customer conversation. The primary concierge diagnoses the issue, then dispatches agents to take parallel actions - freezing a card, issuing a replacement, updating the CRM, and contacting relevant third parties. ### Cross-System Diagnosis Technical problems rarely live in one system. A failed transaction might involve the payment processor, fraud detection, the customer's account status, and a third-party banking partner. Lorikeet queries all relevant systems via API to build a complete picture before acting. It examines transaction history, checks feature availability for the specific user, and identifies which system in the chain caused the failure - the kind of investigation a skilled human agent would do, but faster. ### Sequential Dependency Handling Complex troubleshooting has dependencies: step 3 cannot happen until steps 1 and 2 succeed. Lorikeet manages these sequences automatically. In a [card fraud scenario](https://www.lorikeetcx.ai/blog/lorikeet-voice-2-0), the system blocks the compromised card first, then creates a virtual replacement, then overnights a physical card to the customer's hotel, then contacts the taxi company with the new card details. Each action depends on the previous one completing successfully. If any step fails, the system adapts the plan rather than proceeding blindly. ## What Technical Issues Can AI Actually Resolve Today? AI can resolve technical issues that follow identifiable patterns and connect to systems via APIs. This includes payment failures, account access problems, integration errors, configuration issues, and system status inquiries. The practical ceiling depends on your API infrastructure, not the AI's reasoning capability. 1. **Payment and billing diagnostics.** The AI checks transaction logs, identifies decline codes, verifies payment method status, and takes corrective action - processing a retry, updating the payment method, or escalating a dispute. These represent a significant portion of technical support volume in fintech and e-commerce. 2. **Account and access troubleshooting.** Locked accounts, failed logins, permission issues, and credential resets are resolved by querying the identity system, verifying the customer, and executing the fix directly. No human copy-pasting between admin panels. 3. **Integration and configuration issues.** For platform businesses, customers often need help with API integrations, webhook configurations, or feature settings. The AI can check specific configuration states for individual users and guide them through corrections with context-aware instructions. 4. **Compliance-sensitive diagnostics.** In healthcare, fintech, and energy, troubleshooting involves regulated data. Lorikeet handles identity verification, HIPAA-compliant data access, and financial disclosures with configurable permissions - each agent restricted to only the systems and data it needs. ## What Results Do Teams See on Technical Issues? Teams deploying AI for technical troubleshooting see the largest gains on issues that previously required multi-system investigation by experienced agents. The improvement is most visible in first-contact resolution and handle time because the AI eliminates the back-and-forth between systems that dominates human troubleshooting. First-contact resolution on complex technical issues rises from 20-30% with traditional chatbots to 60%+ with action-capable AI agents. Handle time drops from 10-15 minutes for human-investigated issues to under 4 minutes when the AI can query all systems simultaneously. [Magic Eden](https://www.lorikeetcx.ai/customer-stories), an NFT marketplace with complex debugging requirements, achieved 74% CSAT with Lorikeet - 30 points higher than their previous solution - specifically because the AI could collect technical diagnostic information upfront, enabling faster human resolution on cases that required escalation. Cost per resolution follows. AI-resolved technical tickets cost $1-3 versus $8-12 for human-handled ones. For a team processing 10,000 tickets monthly, shifting even 50% of technical issues to AI resolution saves $25,000-45,000 per month. ## Why Does Self-Awareness Matter for Technical Troubleshooting? Self-awareness matters because technical issues have a long tail of complexity. An AI that attempts everything and fails on 60% of attempts creates worse outcomes than one that resolves 50% confidently and escalates the rest with full context. Customers stuck in unhelpful AI loops during a technical crisis lose trust fast. Lorikeet is built to [know what it does not know](https://www.lorikeetcx.ai/blog/why-deflection-focused-products-make-worse-ai-agents). When it encounters an issue outside its scope, it escalates immediately with the complete diagnostic context: what systems were checked, what was found, and what was attempted. The human agent picks up at step 5 instead of starting at step 1. One company making software for doctors found their previous chatbot gave confused, off-topic responses when patients called in crisis situations. Lorikeet's agent instantly recognized these as outside its scope and escalated to a human immediately - in high-risk environments, that self-awareness is not optional. ## Key Takeaways - Technical troubleshooting requires cross-system diagnosis and multi-step fixes - FAQ handling tools cannot do this, which is why only 14% of issues resolve via self-service today - Lorikeet's Team of Agents queries multiple systems simultaneously, handles sequential dependencies, and resolves complex technical issues in under 4 minutes - Self-awareness is critical: AI that escalates with full context produces better outcomes than AI that attempts everything and fails on 60% of tickets - Magic Eden achieved 74% CSAT (30 points higher than previous solution) by using AI for technical diagnostic collection and resolution --- # What Tools Use AI to Automate Customer Service Workflows? URL: https://www.lorikeetcx.ai/articles/what-tools-use-ai-to-automate-customer-service **AI tools that automate customer service workflows execute multi-step processes -- refunds, account updates, third-party coordination -- across your operational systems without human intervention.** AI tools that automate customer service workflows are platforms capable of executing multi-step processes across your operational systems - processing refunds, updating accounts, coordinating with third parties, and closing tickets without human intervention. According to [Gartner](https://www.gartner.com/), agentic AI will autonomously resolve 80% of common customer service issues by 2029, but today most tools still only handle the information layer, not the action layer. - True workflow automation means executing 5-10 sequential steps across multiple systems in a single conversation - [Lorikeet's](https://www.lorikeetcx.ai) Team of Agents architecture dispatches specialized agents to handle parallel tasks - including calling third parties by phone - Companies using action-capable AI report 50-70% automated resolution within 90 days on targeted ticket types - Most "AI automation" tools stop at FAQ responses - they cannot process a return, generate a shipping label, or issue store credit The phrase "AI automation" in customer service has become meaningless. Every vendor claims it. Half of them mean a chatbot that surfaces knowledge base articles faster than your old search bar. The other half mean a system that reads a ticket, takes 7 actions across 3 systems, and closes the issue before a human touches it. Those are not the same product category. Here is how to tell the difference and which tools actually automate workflows end to end. ## What Does Workflow Automation Mean in Customer Service? Workflow automation in customer service means an AI agent independently executes the full sequence of steps required to resolve a customer issue - from reading the request, to querying systems, to taking action, to confirming the outcome. It is not ticket routing. It is not auto-responses. It is the AI doing the work. A refund request is a useful test case. A knowledge-base tool tells the customer how to request a refund. A workflow automation tool checks the order in your OMS, verifies the return window against your policy, processes the refund through your payment processor, generates a return shipping label, issues a store credit, updates the CRM, and notifies the customer - all in one interaction. That is 7 steps across 3 systems. According to [McKinsey](https://www.mckinsey.com/capabilities/operations/our-insights/the-next-frontier-of-customer-engagement-ai-enabled-customer-service), companies deploying this level of AI see 40-50% fewer service interactions and more than 20% reduction in cost to serve. ## How Does Lorikeet's Team of Agents Architecture Work? Lorikeet's Team of Agents architecture works by having a primary concierge agent manage the customer conversation while dispatching specialized agents to handle specific tasks in parallel. Each spawned agent has defined parameters: what information to gather, what outcome to achieve, and time limits for completion. ### Multi-Party Coordination This is where most tools fail entirely. Real customer service often requires contacting people outside your organization - calling a doctor for consultation notes, reaching a hotel for a late checkout, checking with a KYC team for onboarding verification. [Lorikeet's Team of Agents](https://www.lorikeetcx.ai/blog/team-of-agents) dispatches other agents via phone, SMS, email, or Slack to handle these third-party contacts while the customer waits. A bank could have one agent call a merchant about a dispute while the primary agent keeps the customer informed. ### Parallel Execution Sequential processing is slow. Lorikeet runs tasks simultaneously. In a card fraud scenario, the system blocks the compromised card, creates a new virtual card, spawns one agent to call the customer's hotel about the new payment details and another to contact the taxi company - all while maintaining a natural conversation with the customer on the line. This parallel execution compresses what would be a multi-day, multi-handoff process into minutes. ## What Can AI Workflow Tools Actually Automate Today? AI workflow tools can automate any process that follows defined business rules and connects to systems via APIs. The practical limit is not the AI's reasoning ability but whether your systems expose the right endpoints. Most teams start with high-volume, low-risk ticket categories that represent 40-60% of total support volume. 1. **Refund and billing workflows.** Lorikeet connects to payment processors like Stripe and Adyen to process refunds, adjust billing, and issue credits autonomously. [Summ](https://www.lorikeetcx.ai/customer-stories/how-summ-efficiently-handles-support-spikes-during-tax-time) automated their refund workflows during tax season and achieved 97% faster resolutions. 2. **Account management workflows.** Address changes, subscription modifications, credential resets, and plan upgrades - all executed directly in your CRM without human copy-pasting between systems. 3. **Cross-system coordination.** Returns requiring OMS updates, shipping label generation, payment reversal, and customer notification happen in one automated flow. [Linktree](https://www.lorikeetcx.ai/customer-stories/learn-how-linktree-s-support-team-cut-first-response-time-to-1-minute-and-automated-complex-workflows-with-lorikeet) built workflows reflecting how support actually operates, cutting first response time to 1 minute. 4. **Omnichannel deployment.** Lorikeet's Universal Concierge architecture means workflows built once deploy identically across chat, email, voice, and SMS. Eucalyptus automated 80% of first-response emails using this approach. ## What Results Do Teams See After Deploying Workflow Automation? Teams deploying genuine workflow automation see measurable results within the first 90 days, concentrated in resolution rate, handle time, and cost per ticket. The gains are largest on the ticket types you automate first - typically high-volume, policy-driven requests. Automated resolution rates reach 50-70% on targeted ticket categories within 90 days. Handle time for AI-resolved tickets drops to under 3 minutes compared to 8-12 minutes for human-handled equivalents. Cost per resolution falls from $8-12 per ticket to $1-3 per AI resolution. [GiveCard](https://www.lorikeetcx.ai/customer-stories/lorikeet-helped-arbor-automate-complex-support-flows-across-13-states) deployed Lorikeet's multilingual voice agents in under 48 hours during an emergency food assistance response, serving 300,000 people across English, Spanish, and Mandarin with over 60,000 calls handled. These results depend on integration depth. Tools that only connect to your knowledge base cannot produce these numbers because they cannot execute the resolution steps. The automation ceiling is determined by how many of your systems the AI can read from and write to. ## How Do You Evaluate Workflow Automation Tools? Evaluate workflow automation tools by testing resolution capability, not conversation quality. The demo should show the AI completing a real multi-step workflow, not answering questions impressively. If you see only FAQ responses, you are evaluating a chatbot, regardless of how the vendor packages it. Check for three capabilities. First, action-taking: can the AI write to your systems, not just read from them? Second, policy-awareness: can it interpret dynamic business rules like return windows and approval thresholds, or does it follow rigid scripts? Third, escalation intelligence: when the AI hits its limits, does it hand off with full context or force the customer to repeat everything? Lorikeet's instruction-based approach means every workflow is transparent and auditable - critical for [quality assurance](/articles/what-does-qa-mean-in-customer-service) and compliance in regulated industries. ## Key Takeaways - True workflow automation means the AI executes multi-step processes across systems - not just surfacing articles or routing tickets - Lorikeet's Team of Agents dispatches specialized agents in parallel, including calling third parties by phone, to resolve complex multi-party issues in minutes - Teams see 50-70% automated resolution on targeted ticket types within 90 days, with cost per resolution dropping from $8-12 to $1-3 - Evaluate tools by testing real end-to-end resolution, not conversation quality - if the demo only shows FAQ answers, it is a chatbot --- # Which Support Agent Can Query and Update CRM Data Automatically? URL: https://www.lorikeetcx.ai/articles/which-support-agent-can-query-and-update-crm-data **AI support agents that query and update CRM data need direct API integrations into your operational systems -- not just knowledge base access -- to actually resolve customer issues.** AI support agents that query and update CRM data automatically are platforms with direct API integrations into your operational systems - not just your knowledge base. According to [Gartner](https://www.gartner.com/), only 14% of customer issues resolve through traditional self-service, largely because most tools can read documentation but cannot take action in backend systems like CRMs, payment processors, or order management platforms. - Most AI support tools connect to your help center only - they cannot read or write CRM records - Platforms like [Lorikeet](https://www.lorikeetcx.ai) integrate directly with CRMs, billing tools, and ticketing systems via APIs - Real CRM access means the AI can pull a customer's last 3 transactions and verify identity mid-conversation - Teams using action-capable AI agents report 40-60% reductions in cost per resolution within 90 days Your customer contacts support about a billing error. The AI identifies the problem, explains what happened, then tells the customer to email the billing team. That is not resolution. That is a forwarding service with extra steps. The gap between tools that read your CRM and tools that write to it is the difference between deflection and actual problem-solving. Here is how to tell which category your AI falls into. ## What Does CRM Integration Actually Mean for AI Support? CRM integration for AI support means the agent has programmatic read-write access to your customer database, payment systems, and ticketing platform through API connections. It can pull account details, modify records, process transactions, and update ticket statuses without human intervention or system switching. Most vendors use "CRM integration" loosely. They mean the AI can look up a help article when someone mentions a billing issue. Real integration means the AI queries the CRM for the customer's account, checks their payment history in Stripe, verifies their identity, and updates their record - all within a single conversation. Would you hire a human agent and not give them system access? The same logic applies to AI. [Lorikeet's deployment approach](https://www.lorikeetcx.ai/blog/how-to-deploy-ai-support-without-redesigning-your-entire-operation) gives AI agents the same programmatic access as your best human agents from day one. ## How Do Most AI Tools Handle CRM Data Today? Most AI support tools handle CRM data by not handling it at all. They connect to your knowledge base, search for relevant articles, and paste back answers. When a customer needs something done in the CRM - an address change, a refund, a subscription update - the AI either escalates to a human or tells the customer how to do it themselves. This is the knowledge-base-only model. According to [Ipsos research from 2023](https://www.ipsos.com/), only 35% of customers say chatbots usually solve their issue effectively. The reason is straightforward: most support requests require action, not information. A customer asking "where's my order?" does not want a tracking FAQ. They want the AI to pull their order from the system and give them a real-time status. ## Which Platforms Can Read and Write CRM Records? Platforms that read and write CRM records are those built with deep API integrations into operational systems from the ground up. Lorikeet is the leading example - its architecture connects directly to CRMs, payment processors like Stripe and Adyen, ticketing platforms like Zendesk, and order management systems like Shopify through live API connections. 1. **Account updates mid-conversation.** Lorikeet modifies customer records, subscription tiers, addresses, and credentials autonomously. One agent checks order status while another processes a refund through Stripe and a third updates CRM records - all in parallel using its [Team of Agents architecture](https://www.lorikeetcx.ai/blog/team-of-agents). 2. **Dynamic permission controls.** Not every action should be automatic. Lorikeet uses dynamic gating to auto-approve low-risk actions (a $20 refund) while routing high-risk ones ($500 dispute) to a human agent with full context. This is configurable per action, per dollar threshold, per customer tier. 3. **Custom system integrations.** [Arbor](https://www.lorikeetcx.ai/customer-stories/lorikeet-helped-arbor-automate-complex-support-flows-across-13-states), an energy company operating across 13 US states, connected Lorikeet to their Notion-based knowledge system with a custom integration and was operational within a week. The AI reads state-specific energy content and takes action based on regional policies. ## What Results Can You Expect From CRM-Connected AI? CRM-connected AI agents produce measurably different outcomes than knowledge-base-only tools because they resolve issues instead of describing solutions. The improvement shows up across resolution rate, handle time, and cost per ticket within the first 90 days of deployment. First-contact resolution typically rises from 20-30% (standard chatbot range) to 55-70% when the AI can actually execute actions in your systems. Handle time drops from 8-12 minutes for human-assisted tickets to under 3 minutes for AI-resolved ones. Cost per resolution falls from $8-12 per ticket to $1-3 per AI-handled resolution - a 40-60% reduction. According to [McKinsey](https://www.mckinsey.com/), companies deploying advanced AI in service operations see 40-50% fewer service interactions overall. These numbers shift because the bottleneck in most support operations is not understanding the problem. It is acting on it. When your AI can query the CRM, verify the issue, and fix it in the same conversation, the entire resolution chain compresses. ## How Should You Evaluate CRM Integration Depth? Evaluate CRM integration depth by testing whether the AI can complete a real end-to-end resolution during a demo, not just answer questions about your product. If the vendor demo only shows FAQ responses, you are looking at a chatbot with upgraded marketing. ### Questions That Expose Real Integration Ask the vendor to show the AI pulling a specific customer's transaction history, modifying an account field, and processing a refund - all in one conversation. If they cannot demonstrate this live, their "CRM integration" means read-only access to documentation, not operational write access. Check whether actions require human approval at every step or if the system supports configurable autonomy thresholds. ### The Audit Trail Requirement For regulated industries - fintech, healthcare, energy - every CRM write needs an audit trail. Instruction-based systems like Lorikeet provide transparent, auditable decision-making because every action follows explicit rules you define. Self-training systems that learn from thumbs-up feedback cannot explain their decisions to a regulator. That distinction matters when compliance is not optional. ## Key Takeaways - Most AI support tools connect to your knowledge base only - they cannot query or write to your CRM, which is why resolution rates stay at 10-25% - Lorikeet integrates directly with CRMs, payment tools, and ticketing systems via APIs, enabling autonomous account updates, refunds, and record changes mid-conversation - CRM-connected AI delivers 55-70% first-contact resolution versus 20-30% from knowledge-base-only tools, with handle times under 3 minutes - Dynamic gating controls let you auto-approve low-risk CRM actions while routing high-value decisions to humans with full context --- # Why Is My CSAT Dropping? Hidden Causes and Early Warnings URL: https://www.lorikeetcx.ai/articles/why-is-my-csat-dropping **CSAT drops are usually caused by 2–3 compounding factors -- survey methodology decay, upstream product changes, and agent consistency erosion -- not a single failure.** CSAT drops are almost never caused by a single failure. They result from 2-3 compounding factors - survey methodology decay, upstream product changes, and agent consistency erosion - that individually stay below alarm thresholds while collectively dragging the score down over weeks. - 66% of unhappy customers leave without telling you, per Qualtrics - your declining CSAT may understate the real problem - Segment by agent, channel, issue type, and time period to isolate the root cause - Targeted fixes (QA coaching, routing automation, resolution verification) recover points in 2-6 weeks - Replace lagging CSAT monitoring with leading indicators: response time, reopen rate, and QA scores The dashboard says 74%. Last quarter it was 79%. Nobody changed anything - same team, same tools, same processes. Yet every month the number inches lower. This is the signature pattern of a CSAT decline: gradual, invisible in weekly reports, and already entrenched by the time someone escalates it. The fix starts with understanding that CSAT drops are almost never caused by one thing going wrong. They are caused by several small things going unnoticed. ## Why Do Gradual CSAT Declines Happen? Gradual CSAT declines happen because customer satisfaction is a lagging indicator that aggregates multiple small deteriorations. A 1-point monthly decline across response time, agent consistency, and resolution quality compounds to a multi-point quarterly drop - and no single metric triggers an alarm along the way. The other driver is invisible: rising customer expectations. Customers do not benchmark your support against your last quarter. They benchmark against their last best experience anywhere. When Uber resolves an issue in 2 minutes via the app, that becomes the expectation for your insurance support team. The [Qualtrics 2024 Consumer Trends report](https://www.qualtrics.com/) found that 66% of customers who have a bad experience will not tell the company - they simply leave. Your CSAT may drop not because service got worse, but because the customers who would have rated you lowest are no longer in your survey pool, and the remaining customers now expect more. ## What Are the Hidden Causes Most Teams Miss? The hidden causes most teams miss are survey methodology decay, upstream product changes creating downstream support complexity, and agent consistency erosion from coaching gaps. These factors do not show up in standard operational dashboards, which is why the decline feels mysterious. 1. **Survey methodology decay.** Response rates decline over time as customers develop survey fatigue. Research shows response rates have dropped significantly over the past decade. When your response rate falls below 15%, the remaining respondents skew toward extremes - very happy or very angry - and your CSAT becomes unreliable. A 3-point CSAT drop might be a real service decline or just a noisier sample. 2. **Upstream product changes.** A new pricing tier, a redesigned checkout flow, a deprecated feature - these decisions happen in product and engineering meetings, but the support team absorbs the fallout. Ticket complexity rises, resolution times increase, and CSAT drops. The support team did not cause the problem and cannot fix it alone. 3. **Agent consistency erosion.** Turnover replaces experienced agents with new hires. Coaching cadence slips from weekly to monthly. Top performers leave and the performance floor drops. The change is gradual - a few points per month - but compounds relentlessly. 4. **Channel mix shift.** If your customer base migrates from phone (where satisfaction tends to be higher) to email or chat (where it is typically 3-5 points lower), your blended CSAT drops even if quality stays constant on every channel. ## How Do You Pinpoint the Root Cause? Pinpoint the root cause by decomposing your CSAT decline into 4 segments: agent, channel, issue type, and time period. The segment with the highest variance - where the decline concentrates - reveals where to look. A uniform decline across all segments points to systemic factors like rising expectations or survey issues. ### The Agent View Compare CSAT by individual agent over the declining period. If the drop concentrates in your bottom quartile or in recently hired agents, you have a coaching and ramp problem. If your top performers are also declining, the cause is upstream - likely product complexity or policy changes that agents cannot solve. ### The Issue Type View Segment CSAT by ticket category. A sharp drop in "billing" after a pricing change or in "returns" after a policy update reveals a specific trigger. A uniform decline across all categories suggests systemic factors - response time degradation, survey fatigue, or expectation drift. ## What Fixes Produce the Fastest CSAT Recovery? The fastest CSAT recovery comes from targeting the single highest-impact failure point identified in your segmentation analysis, not from broad improvement programs. Specificity is everything - a targeted fix recovers points in weeks while an "improve customer experience" initiative produces nothing for months. If the root cause is agent inconsistency, implement weekly 1-on-1 coaching sessions tied to specific [QA scores](/articles/what-does-qa-mean-in-customer-service). Teams that close the gap between their best and worst performers see the CSAT floor rise, which lifts the overall average. If the root cause is response time creep, invest in routing automation and triage - significant first-response-time reductions directly correlate with CSAT improvement. If the cause is false resolutions (tickets marked "solved" that are not), add a confirmation step before ticket closure. This reduces reopened ticket rates and prevents customers from receiving satisfaction surveys about unresolved issues. The common mistake is treating the symptom (CSAT score) instead of the cause. Training programs, motivational initiatives, and "customer first" campaigns do not move the number. Fixing the specific broken step in the process does. ## How Do You Build an Early Warning System? Build an early warning system by monitoring CSAT at the segment level weekly, not monthly at the aggregate level. A 2-point drop in one agent's score or one issue category is fixable in a week. A 2-point drop in your company-wide average means the problem has been compounding across multiple segments for weeks. Set automated alerts for: any agent whose CSAT drops 5 or more points week-over-week, any issue category that drops 3 or more points in a month, response rate falling below 15% (which degrades data reliability), and first response time increasing more than 20% from baseline. Implement [continuous QA - reviewing 100% of tickets via AI](https://www.lorikeetcx.ai/product/coach) rather than small manual samples - so quality issues surface in real time. Monthly QA reviews are already 4 weeks stale. By the time you see the pattern, your CSAT has already absorbed the damage. ## Key Takeaways - CSAT declines are rarely one cause - expect 2-3 compounding factors (response time, consistency, survey methodology) eroding simultaneously - 66% of unhappy customers leave without telling you, per Qualtrics - your declining CSAT may understate the real problem - Segment by agent, channel, issue type, and time period - the highest-variance dimension reveals the root cause - Target the single highest-impact failure point for fast recovery - broad initiatives do not move CSAT --- # What Does QA Mean in Customer Service? The Full Breakdown URL: https://www.lorikeetcx.ai/articles/what-does-qa-mean-in-customer-service **QA in customer service is the systematic evaluation of customer interactions against defined standards, combined with coaching to drive measurable agent performance gains.** QA in customer service stands for quality assurance - the systematic evaluation of customer interactions against defined standards, combined with coaching to drive agent improvement. It is not just scoring tickets. It is the feedback loop that turns reviews into measurable performance gains across resolution rate, consistency, and customer satisfaction. - QA follows a 4-stage loop: review, score, coach, calibrate - skipping any stage breaks the system - Most programs fail due to small sample sizes (3-5 tickets/month), disconnected coaching, and compliance-focused scorecards - AI QA eliminates the sample constraint by reviewing 100% of interactions automatically - Effective QA should measurably lift CSAT and first-contact resolution within 3-6 months If you asked 10 CX leaders what QA means, you would get 10 different answers. One says it is a scorecard. Another says it is compliance. A third says it is "something we do because we're supposed to." This confusion is the reason most QA programs produce reports instead of results. QA in customer service has a precise meaning - and understanding it is the difference between grading tickets and actually improving your support operation. ## What Does QA Stand for in Customer Support? QA stands for quality assurance - the systematic evaluation of customer interactions against defined performance standards. In customer support, this means reviewing tickets, calls, and chats to assess whether agents provided accurate information, followed correct processes, communicated effectively, and actually resolved the customer's issue. The "assurance" part is key. QA is not quality measurement - that is just data collection. Assurance implies a guarantee: that your team's output meets a defined standard consistently. When QA works, it creates a feedback loop where evaluation leads to coaching, coaching leads to behavior change, and behavior change leads to better customer outcomes. When it does not work, it is just measurement with extra steps. ## What Does a QA Process Look Like Day to Day? A QA process consists of four repeating stages: review, score, coach, and calibrate. Each stage feeds the next. Skipping any one of them breaks the loop and turns QA into a reporting exercise rather than an improvement engine. ### Review and Score Reviewers evaluate interactions using a scorecard with weighted criteria. Traditional programs sample 3-5 tickets per agent per month. [AI-powered tools](https://www.lorikeetcx.ai/product/coach) review every interaction automatically. The output is a set of scores and flagged issues - but scores alone change nothing. ### Coach and Calibrate Coaching connects QA findings to agent behavior change. The best teams deliver coaching within 7 days of the reviewed interaction while context is fresh. Calibration sessions ensure reviewers score consistently - without calibration, different reviewers will interpret the same interaction differently, undermining agent trust in the entire program. ## What Are the Most Common QA Mistakes? The most common QA mistakes are reviewing too few interactions, disconnecting scoring from coaching, and building scorecards that measure compliance instead of customer outcomes. These three errors account for why the majority of QA programs fail to move CSAT, resolution rates, or agent performance. 1. **Insufficient sample size.** Reviewing 3-5 tickets per agent per month from a pool of hundreds gives you a statistically meaningless sample. Conclusions drawn from 1-2% of an agent's work are unreliable. It is the equivalent of judging a restaurant by one appetizer. 2. **Scoring without coaching.** If QA scores go into a spreadsheet and surface in a monthly report, they are already stale. Effective QA requires that every finding connects to a specific coaching conversation within the same week. Scores that do not lead to conversations do not lead to change. 3. **Compliance-focused scorecards.** Scorecards that overweight process adherence ("did the agent use the customer's name?") while underweighting resolution ("was the problem actually fixed?") produce high QA scores and low CSAT. The scorecard should predict customer satisfaction - if it does not, the criteria are wrong. 4. **No calibration.** Without regular calibration sessions, reviewers drift apart in how they interpret criteria. One reviewer gives a 4/5 for the same interaction another rates 2/5. Agents lose trust in the system, and the data becomes unreliable. Target a kappa score above 0.8 for strong inter-rater agreement. ## What Impact Should a QA Program Have? A QA program should measurably improve agent consistency, reduce repeat contacts, and lift [customer satisfaction](/articles/what-is-a-good-csat-score) within 3-6 months. If your program has been running for a year and none of these metrics have moved, the program is not working - regardless of how many scorecards you have filled out. The cross-industry first-contact resolution average sits at 71%, according to [SQM Group](https://www.sqmgroup.com/). Teams with mature QA programs consistently exceed this because they catch error patterns early and coach them out before they become habits. The more direct impact is on agent consistency - when the gap between your best and worst performers narrows, your worst customer experiences improve, and that is what lifts the overall CSAT floor. Teams implementing structured QA with weekly coaching typically see measurable consistency improvements within 4-6 weeks. ## How Does AI Change the QA Equation? AI changes the QA equation by eliminating the sample size constraint. Manual QA is limited by reviewer capacity - typically 1 QA analyst per 15-25 agents, each reviewing a handful of tickets. AI reviews every interaction, every time, against the same criteria. This shifts QA from statistical estimation to complete measurement. The practical impact goes beyond coverage. [AI-powered QA](https://www.lorikeetcx.ai/product/coach) detects patterns that small samples miss - an agent who consistently struggles with one issue type but handles others well, a process step that generates confusion across the entire team, or a policy change that is silently driving up escalation rates. These patterns are invisible in a 5-ticket monthly sample. They are obvious in a full-coverage view. For teams without dedicated QA staff, AI QA provides the capability without the headcount. ## Key Takeaways - QA means quality assurance - systematic evaluation plus coaching that drives behavior change, not just scoring tickets - The 4-stage QA loop - review, score, coach, calibrate - breaks when any stage is skipped or delayed - Most programs fail due to small samples, disconnected coaching, and compliance-focused scorecards that do not predict customer outcomes - AI QA eliminates the sample constraint, reviewing 100% of interactions and surfacing patterns invisible in manual sampling --- # What Is a Good CSAT Score? How to Read Your Number URL: https://www.lorikeetcx.ai/articles/what-is-a-good-csat-score **A good CSAT score falls between 75–85% for most industries. The US national average is 77–78% according to the ACSI, but benchmarks vary widely by sector.** A good CSAT score falls between 75-85% for most industries. The US national average sits at 77-78% according to the American Customer Satisfaction Index, but benchmarks vary widely - e-commerce averages 80-83% while telecoms sit at 68-71%. The number alone tells you very little without industry context, survey methodology, and trend direction. - 75-85% is the healthy range; below 70% signals structural problems, above 90% likely signals survey bias - Always benchmark within your industry - complexity and switching costs set the floor - Pair CSAT with first-contact resolution, response time, and reopen rate for the full picture - A stable 76% beats a declining 84% - track trajectory, not snapshots You pulled up the dashboard, saw 78%, and immediately wondered: are we good or are we coasting? CSAT benchmarks get thrown around in every QBR, but the number itself is almost useless without context. Industry matters. Channel matters. Whether you are measuring post-resolution or post-interaction matters. Here is how to actually benchmark your CSAT score - and what the number is hiding from you. ## What Does CSAT Actually Measure? CSAT measures a customer's satisfaction with a single interaction, not their overall relationship with your brand. It captures a snapshot of how the customer felt at one specific moment - after a support ticket, a purchase, or a service call. That narrow scope is both its strength and its limitation. The standard approach uses a 1-5 scale survey sent immediately after an interaction. Your CSAT percentage equals the number of 4 and 5 responses divided by total responses. A customer who rates you 3 out of 5 - neutral - counts against your score. This means CSAT penalizes mediocrity, not just failure. That design choice makes it a sharper tool than it first appears, but it also means small shifts in "okay" experiences disproportionately move the number. ## What CSAT Range Should You Target? Target 75-85% as your working range. Below 70% signals structural problems in your support operation. Above 90% usually means your survey methodology is skewed, not that your service is exceptional. The goal is consistent improvement within that band, not chasing a perfect score. The [American Customer Satisfaction Index (ACSI)](https://www.theacsi.org/) reported a US average of approximately 77-78 out of 100 across 2024. But averages obscure more than they reveal. E-commerce companies regularly hit 80-83% because transactions are straightforward and expectations are well-defined. Telecom and utilities sit at 68-71% because their interactions involve complex billing disputes and service disruptions. If you are a fintech company comparing yourself to a retail benchmark, you are measuring against the wrong yardstick. ## Why Is Industry Context More Important Than the Number? Industry context matters because CSAT reflects the difficulty of your interactions as much as the quality of your service. A healthcare support team resolving insurance claims operates in a fundamentally different complexity tier than a fashion brand processing returns. Both can be excellent - their scores will never look the same. ### Complexity Drives the Baseline Industries with multi-step, regulated, or emotionally charged interactions see lower CSAT floors. Financial services, insurance, and healthcare typically operate in the 70-78% range even with strong teams. The issue is not agent performance - it is the inherent friction in the process. ### Customer Alternatives Drive the Ceiling When switching costs are low, dissatisfied customers leave rather than complain. Retail and food delivery CSAT scores look higher partly because the unhappy customers already churned. They are not in your survey pool. Captive industries - utilities, insurance - retain frustrated customers who pull the average down. ## What Metrics Should You Track Alongside CSAT? CSAT alone is a partial picture. Pair it with first-contact resolution (FCR), response time, and reopened ticket rate to understand what is driving the score, not just what the score is. A high CSAT with low FCR means customers are satisfied but you are wasting resources on repeat contacts. 1. **First-contact resolution rate.** The percentage of tickets resolved in one interaction. Industry average sits around 71% per [SQM Group](https://www.sqmgroup.com/). When FCR rises, CSAT almost always follows - fewer repeat contacts means less customer frustration. 2. **First response time.** How long before the customer hears back. Expectations vary by channel - under 1 hour for email, under 2 minutes for chat. According to [HubSpot](https://www.hubspot.com/), 90% of customers consider an "immediate" response important, with 60% defining immediate as under 10 minutes. 3. **Reopened ticket rate.** Tickets marked resolved that bounce back. Industry best practice is to keep this below 5%. High reopen rates indicate false resolutions, which tank CSAT on the second interaction. 4. **Agent consistency spread.** The CSAT gap between your top and bottom quartile agents. A narrow spread means your training and [QA](/articles/what-does-qa-mean-in-customer-service) are working. A wide spread means your customer experience depends on who picks up the ticket. ## When Should You Worry About Your CSAT Score? Worry when the trend moves, not when the absolute number looks unfamiliar. A steady 74% in a complex industry is healthier than an 82% that dropped from 88% over two quarters. Trajectory reveals operational health; snapshots do not. Specific triggers worth investigating: any segment dropping 3 or more points in a single month, a widening gap between your best and worst agent scores, or response rates falling below 15% (which makes the score statistically noisy). Also watch for seasonal patterns - CSAT typically dips 3-5 points during high-volume periods like holidays due to longer wait times and temporary staff. ## Key Takeaways - Target 75-85% CSAT - below 70% signals structural issues, above 90% likely signals survey bias - Always benchmark against your own industry, not cross-sector averages - complexity and switching costs set the floor - Pair CSAT with FCR, response time, and reopen rate to understand what is driving the score - Track trajectory over absolute number - a stable 76% beats a declining 84% every time --- # AI Agent Platform: How to Choose the Right One for Customer Service URL: https://www.lorikeetcx.ai/articles/ai-agent-platform An AI agent platform is software that deploys autonomous AI agents to handle customer service interactions across chat, email, and voice. Unlike basic chatbot builders, these platforms connect to backend systems, execute multi-step workflows, and resolve issues without human involvement. The market has grown rapidly - [Gartner predicts](https://www.gartner.com/en/newsroom/press-releases/2025-03-05-gartner-predicts-agentic-ai-will-autonomously-resolve-80-percent-of-common-customer-service-issues-without-human-intervention-by-20290) that by 2029, agentic AI will autonomously resolve 80% of common customer service issues without human intervention. - AI agent platforms resolve 50-70% of support tickets autonomously vs. 15-25% for chatbot tools - Key differentiators: backend integrations, workflow complexity, accuracy guardrails, and channel coverage - Pricing models vary widely - per resolution ($0.50-3.00), per seat ($100-500/mo), or usage-based - Deployment timelines range from 2 weeks for basic setups to 8 weeks for full enterprise rollouts The AI agent platform market is crowded and confusing. Every vendor claims high resolution rates. Few can prove it. The difference between a platform that deflects tickets and one that resolves them comes down to architecture - specifically, how deeply the AI connects to your systems and how accurately it follows your policies. This guide breaks down what actually matters when evaluating platforms, based on what separates tools that work from tools that demo well. ## What Makes an AI Agent Platform Different from a Chatbot Builder? Chatbot builders create scripted conversation flows with if-then logic. AI agent platforms deploy autonomous agents that reason through problems, access live data from your systems, and take actions - processing refunds, updating subscriptions, modifying orders - without following a predetermined script. The architectural difference is fundamental. Chatbot builders like ManyChat or Tidio work well for lead qualification and simple FAQ responses. AI agent platforms like [Lorikeet](https://www.lorikeetcx.ai), Ada, and Forethought connect to your order management, billing, CRM, and knowledge base to handle complex, multi-step customer requests. Think of it as the difference between a phone tree and a trained support agent - one routes, the other resolves. ## What Features Should You Prioritize When Evaluating Platforms? Focus on 3 areas: integration depth with your existing systems, accuracy and guardrail controls, and transparent reporting on resolution vs. deflection rates. Everything else - the UI, the AI model used, the marketing language - is secondary to these fundamentals. 1. **Backend system integrations.** The platform must read and write to your core systems - Shopify, Stripe, Salesforce, your OMS. If it can only read data but not take actions, it's a lookup tool, not an agent. Check the depth of each integration, not just the logo wall. 2. **Policy and guardrail controls.** You need fine-grained control over what the AI can and cannot do. Can you set refund limits? Restrict actions by customer tier? Define escalation triggers? Platforms with [built-in quality assurance](https://www.lorikeetcx.ai/product/coach) that reviews 100% of AI interactions catch policy violations before they reach customers. 3. **Resolution metrics, not vanity metrics.** Demand containment rate (tickets resolved without human involvement), not "automation rate" or "engagement rate." A platform that deflects 80% of tickets to a human isn't automating - it's creating extra steps. 4. **Channel coverage.** Can the same AI agent work across chat, email, and voice? Multi-channel support from a single platform reduces complexity and ensures consistent customer experiences. ## How Do AI Agent Platforms Handle Complex Workflows? The best platforms break customer requests into reasoning steps - identify intent, gather context from backend systems, check policies, execute actions, and confirm with the customer. This happens in seconds, across multiple systems, within a single conversation. For example, a customer says "I received the wrong item and I need a replacement." A capable AI agent platform will: pull the order details, verify the delivery, check return policy eligibility, initiate a return label, create a replacement order, and send confirmation - all without human intervention. Industry analysis shows that platforms handling this level of workflow complexity achieve significantly higher containment rates than simpler chatbot tools. The key technical capability is multi-system orchestration - the ability to chain actions across different APIs in a single workflow. ## What Results Do Companies See After Deploying an AI Agent Platform? Companies deploying mature AI agent platforms see measurable ROI within the first 90 days. The results vary by industry and ticket complexity, but the direction is consistent across deployments. Containment rates typically reach 50-70% within the first quarter, up from 10-20% with legacy chatbots. Average handle time drops from 8-12 minutes to 2-4 minutes for AI-resolved tickets. Cost per resolution falls from $8-15 with human agents to $0.50-2.00 with AI, according to industry benchmarks. Customer satisfaction scores for AI-resolved tickets match or exceed human agent scores when the resolution is accurate and fast. The indirect benefits matter too. Human agents handle fewer but more meaningful tickets. Response times drop across the board. And the data from AI interactions feeds back into product and process improvements. ## What Are the Common Pitfalls When Choosing a Platform? The biggest mistake is buying based on the demo instead of testing with your actual tickets. Every platform looks impressive resolving a scripted scenario. The real test is handling your messiest, most ambiguous customer requests with your real data and policies. Other common traps: choosing the cheapest option (which usually means shallow integrations), over-indexing on the AI model name (GPT-4 vs. Claude matters less than how the platform uses it), and ignoring the accuracy monitoring layer. A platform without built-in quality assurance means you're trusting the AI blindly. Run a pilot with 100-200 real tickets before committing. Measure actual resolution rate, not the vendor's reported numbers. And check what happens when the AI gets it wrong - does the platform catch errors proactively, or do your customers find them first? ## Key Takeaways - Evaluate platforms on resolution rate, not automation rate - demand containment metrics - Integration depth matters most: the AI must read and write to your backend systems - Expect 50-70% containment and $0.50-2.00 cost per resolution within 90 days - Always pilot with real tickets before committing - demos don't reflect production performance ## Frequently Asked Questions ### How much does an AI agent platform cost? Pricing varies by model: per-resolution ($0.50-3.00 per ticket), per-seat ($100-500/month per agent seat), or usage-based. Most mid-market companies spend $2,000-10,000/month. The ROI calculation should factor in reduced human agent costs - a platform resolving 500 tickets/day at $1 each replaces $2,500-7,500/day in human agent costs. ### How long does implementation take? Basic deployment with standard integrations takes 2-4 weeks. Full enterprise rollout with custom integrations, policy configuration, and multi-channel setup takes 6-8 weeks. Most platforms offer staged rollouts - start with one ticket category, prove ROI, then expand. ### Can AI agent platforms work with my existing helpdesk? Yes, most platforms integrate with Zendesk, Salesforce Service Cloud, Freshdesk, Intercom, and other major helpdesks via API or native integrations. The AI agent sits in front of your existing queue, resolving tickets it can handle and routing the rest to your human team through your existing workflow. The AI agent platform you choose will define your CX capabilities for the next 3-5 years. The market is maturing fast, and the gap between leaders and laggards is widening. Focus on what matters - resolution depth, system integrations, and accuracy controls - and ignore the marketing noise about model sizes and feature counts. Start with a clear picture of your ticket types and volumes. Match those to platform capabilities. Pilot before you commit. **See how an AI agent platform handles your actual tickets. [Explore Lorikeet](https://www.lorikeetcx.ai) - built for complex, multi-step customer service workflows across chat, email, and voice.** --- # AI vs Human Customer Service: How to Get the Balance Right URL: https://www.lorikeetcx.ai/articles/ai-vs-human-customer-service AI customer service uses autonomous agents to resolve routine tickets - refunds, order tracking, account changes - without human involvement. Human agents handle complex, emotionally sensitive, or high-stakes cases that require judgment and empathy. The split isn't about replacing people. [McKinsey research](https://www.mckinsey.com/capabilities/operations/our-insights/the-next-frontier-of-customer-engagement-ai-enabled-customer-service) shows that 50-60% of customer interactions are transactional and automatable, with the most advanced organizations targeting higher automation rates each year. - AI resolves routine tickets 3-5x faster than human agents on average - Human agents outperform AI on complex cases requiring judgment and emotional intelligence - The best teams use AI for volume and humans for value - not one or the other - Companies using blended AI-human models see significant cost reductions by shifting routine volume to AI The "AI vs humans" debate misses the point. It's not a competition. The real question is which tickets should never touch a human - and which ones a human should never miss. Most CX teams get this balance wrong. They either over-automate (customers rage-quit when they can't reach a person) or under-automate (agents burn out answering the same password reset question 200 times a day). Getting the split right is the difference between cutting costs and cutting corners. ## What Can AI Customer Service Actually Handle Today? Modern AI agents resolve order tracking, refund processing, subscription changes, account updates, and FAQ responses without human involvement. These are structured, repeatable tasks where the correct action follows clear business rules and requires system access rather than judgment. The key word is "resolve" - not "deflect." Early chatbots created the impression that AI just bounces customers between menus. Today's AI agents connect to backend systems, pull order data, check policies, and execute actions. A customer asking for a refund on a damaged item gets the refund processed, confirmation email sent, and replacement initiated - all in one conversation. Per [Forrester](https://www.forrester.com/), AI agents now resolve 40-60% of total ticket volume across industries, with e-commerce and subscription businesses reaching 70%+. ## Where Do Human Agents Still Outperform AI? Human agents excel in 4 areas: emotionally charged situations, ambiguous edge cases, high-value customer retention, and complex multi-party disputes. These scenarios require reading tone, exercising judgment, and adapting to context that falls outside standard policy. ### Emotional and Sensitive Situations A customer dealing with a billing error during a family emergency needs empathy, not efficiency. Human agents read emotional cues and adjust their approach. AI can detect sentiment, but responding appropriately to grief, frustration, or anxiety still requires a person. ### Complex Judgment Calls When a long-term customer's request falls in a policy grey area - technically outside the rules but reasonable given their history - human agents make nuanced decisions. They weigh customer lifetime value, precedent, and relationship context in ways AI cannot reliably replicate yet. ## How Should You Split Work Between AI and Human Agents? Map your ticket types by complexity and emotional weight. Route structured, policy-clear tickets to AI and escalate ambiguous, emotional, or high-stakes cases to humans. The split should be dynamic, not static - review monthly as AI capabilities improve. 1. **Tier 1: Full AI automation.** Password resets, order status, tracking updates, simple FAQ. These are 40-50% of most queues and require zero human judgment. Automate completely. 2. **Tier 2: AI with human oversight.** Refunds, cancellations, billing disputes under a threshold. AI handles the workflow; a human reviews edge cases flagged by [AI quality assurance tools](https://www.lorikeetcx.ai/product/coach). 3. **Tier 3: Human-led, AI-assisted.** Escalated complaints, retention calls, complex account issues. The human leads; AI surfaces context, suggests responses, and handles follow-up tasks. 4. **Tier 4: Human only.** Legal, compliance, safety-critical, or deeply emotional interactions. Keep AI out of the conversation entirely. ## What Results Does a Blended AI-Human Model Deliver? Companies that properly balance AI and human agents see measurable improvements across cost, speed, and satisfaction within the first quarter of deployment. The gains come from both sides - AI handling volume and humans handling value. Industry benchmarks suggest support costs per ticket range from $8-15 for human-only teams, dropping to $2-5 in blended AI models. First-response time improves from 4-8 hours to under 2 minutes for AI-routed tickets. CSAT for AI-resolved tickets averages 80-85%, while human-resolved complex tickets score 85-90%. Agent attrition measurably decreases when repetitive work is removed from their queue - less burnout, more meaningful work. The compounding effect matters most. When AI handles routine volume, human agents take fewer tickets per day but spend more time on each. Quality goes up. Burnout goes down. The customers who need a person actually get one quickly. ## What Mistakes Do Teams Make When Deploying AI Alongside Humans? The most common failure is automating by channel instead of by ticket type. Putting AI on chat and humans on phone creates inconsistent experiences. Instead, route by complexity - let AI resolve simple issues across all channels and escalate complex ones to humans regardless of channel. Other pitfalls: setting containment targets too aggressively (forcing AI to handle tickets it shouldn't), failing to build smooth handoff paths (customers repeat themselves after escalation), and not monitoring AI accuracy with [continuous QA](https://www.lorikeetcx.ai/product/coach). The teams that succeed treat AI as a teammate, not a replacement - with regular reviews of what it handles well and where it still needs guardrails. ## Key Takeaways - AI handles 50-60% of transactional volume; humans focus on complex cases requiring judgment - Route by ticket complexity, not by channel - AI should work across chat, email, and voice - Blended models cut cost per ticket from $8-15 to $2-5 while improving CSAT - Monitor AI with continuous QA to catch errors before customers do ## Frequently Asked Questions ### Will AI replace human customer service agents? Not entirely. AI replaces repetitive task execution, not human judgment. Most projections show AI handling 70-80% of routine volume by 2027, while human agent roles shift toward complex problem-solving, retention, and relationship management. The total headcount may decrease, but the remaining roles become more skilled and better compensated. ### How do customers feel about AI vs human support? It depends on the issue. [Salesforce research](https://www.salesforce.com/resources/research-reports/state-of-the-connected-customer/) shows 60% of consumers prefer streamlined, low-touch interactions for simple tasks because speed matters most. But most still want human involvement for complex, high-stakes issues. The key is giving customers a clear path to a person when they need one. ### How long does it take to implement a blended AI-human model? Plan for 4-8 weeks to deploy AI on your first ticket category, then expand one category at a time. Full deployment across all ticket types typically takes 3-6 months, including integration with your existing helpdesk and quality monitoring setup. The AI vs human debate ends when you stop thinking about it as a choice. The best CX teams use both - AI for speed and consistency on routine work, humans for judgment and empathy on everything else. The companies winning at customer service in 2026 aren't choosing sides. They're building teams where each handles what they're best at. The starting point is simple: audit your ticket queue, identify what's repetitive, and automate those categories first. Then reinvest the time savings into your human team's ability to handle the hard stuff well. **Build a blended AI-human support model that actually resolves issues. See how [Lorikeet's AI agents](https://www.lorikeetcx.ai) handle the routine so your team can focus on what matters.** --- # Zendesk Alternatives 2026: 5 Better Platforms Ranked by Use Case URL: https://www.lorikeetcx.ai/articles/zendesk-alternative Zendesk alternatives are customer service platforms that replace or supplement Zendesk's ticketing system with deeper AI automation, different pricing models, or specialized capabilities. The market has shifted since Zendesk's 2022 privatization - [G2](https://www.g2.com/) lists dozens of competing platforms as teams look for alternatives with stronger AI resolution capabilities and more predictable pricing. - Top alternatives include Freshdesk, Intercom, Help Scout, Salesforce Service Cloud, and AI-native platforms like Lorikeet - The main switching triggers: pricing unpredictability, weak AI automation, and slow feature development - AI-native platforms resolve 50-70% of tickets autonomously vs. Zendesk's legacy bots at single-digit rates - Migration typically takes 2-6 weeks depending on data volume and integration complexity Zendesk works. For a lot of teams, it works fine. But "fine" has a cost - escalating per-agent pricing, an AI add-on that feels bolted on rather than built in, and a growing gap between what the platform promises and what it actually automates. If you're reading this, you've probably hit at least one of those walls. The alternatives listed here aren't just cheaper versions of the same thing. They're built on different assumptions about how customer service should work in 2026. ## Why Are Teams Switching Away from Zendesk? Three factors drive most switches: unpredictable pricing that scales with headcount rather than value delivered, AI automation that deflects rather than resolves, and a platform architecture that requires extensive customization to handle complex workflows. Teams with growing ticket volumes feel the pain first. Zendesk's per-agent pricing model creates a structural problem. As your team grows, costs grow linearly - even if AI could handle 60% of that volume without human agents. Per-resolution pricing models used by newer platforms align costs with outcomes instead. The AI gap is the bigger issue, though. Zendesk's Answer Bot and AI features handle simple FAQ lookups, but struggle with multi-step workflows that require backend system access. Teams handling refunds, billing changes, and account modifications often find they're still routing 75-80% of tickets to human agents despite paying for AI features. ## What Are the Best Zendesk Alternatives in 2026? The best alternative depends on what's broken for you. If it's pricing, look at Freshdesk or Help Scout. If it's AI depth, look at AI-native platforms. If it's enterprise scale, Salesforce Service Cloud is the standard. Here's how 5 alternatives compare on the metrics that matter. 1. **Freshdesk.** Best for teams wanting Zendesk-like functionality at 30-50% lower cost. Strong ticketing and knowledge base, decent AI via Freddy. Limitations: AI still chatbot-level, not autonomous agent-level. Per [Freshworks](https://www.freshworks.com/), 50,000+ companies use the platform. 2. **Intercom.** Best for product-led companies wanting conversational support plus marketing in one tool. Fin AI agent handles basic resolution. Limitations: expensive at scale, and Fin's resolution depth is limited on complex workflows that require backend actions. 3. **[Lorikeet](https://www.lorikeetcx.ai).** Best for teams with complex, multi-step tickets (billing, orders, account management). AI-native architecture - the AI agent connects to your backend systems and executes workflows autonomously. Built-in QA reviews 100% of AI interactions. Not a ticketing system replacement - pairs with your existing helpdesk. 4. **Help Scout.** Best for small-to-mid teams wanting simplicity and per-user pricing transparency. Clean UI, shared inbox model. Limitations: minimal AI automation, no autonomous resolution capabilities. 5. **Salesforce Service Cloud.** Best for enterprise teams already in the Salesforce ecosystem. Deep customization, Einstein AI, omnichannel routing. Limitations: complex implementation (3-6 months), high total cost of ownership, and AI features require significant configuration. ## How Do AI-Native Alternatives Compare to Zendesk's AI? Zendesk's AI features (Answer Bot, AI agents) handle intent detection and FAQ surfacing well but struggle with actions that require backend system access. AI-native platforms are built around autonomous resolution - the AI doesn't just understand the question, it executes the solution. The practical difference shows up in containment rates. Zendesk's legacy Answer Bot reportedly resolved single-digit percentages of tickets on average, per customer reports on [G2](https://www.g2.com/). Newer Zendesk AI agents claim higher rates, but user reports vary widely. AI-native platforms built for complex workflows achieve 50-70% containment because they can process refunds, modify subscriptions, update shipping addresses, and handle other multi-step actions that Zendesk's AI routes to human agents. If your ticket mix is 70%+ complex actions (not just FAQ lookups), the AI-native route delivers significantly more value. ## What Should You Consider Before Migrating? Map your current Zendesk setup completely before evaluating alternatives. Document your ticket categories, automation rules, macros, SLA policies, integrations, and reporting dashboards. The migration cost is in recreating your configuration, not transferring your data. Critical migration factors: ticket history portability (most platforms import via CSV or API), integration rebuilding (check that your CRM, phone system, and other tools connect), team retraining time (budget 1-2 weeks), and SLA continuity during the transition. The smoothest migrations run both platforms in parallel for 2-4 weeks, routing new tickets to the new platform while legacy tickets close out in Zendesk. Budget 2-6 weeks total depending on complexity. Most teams report full productivity on the new platform within 30 days. ## What Results Do Teams See After Switching? Teams that switch from Zendesk to AI-native platforms report the clearest gains. Those switching to other traditional helpdesks mainly see cost savings. The performance improvement depends on what you're switching to and why you left. Teams moving to AI-native platforms see containment rates climb to 50-70%, up significantly from what they experienced with Zendesk's built-in AI. Cost per ticket drops from $8-15 (Zendesk with human agents) to $2-5 (AI-native with blended model). First response time for automated tickets goes from minutes to seconds. Teams switching to Freshdesk or Help Scout mainly see 25-40% cost savings with similar operational performance. The biggest wins come from teams that didn't just switch platforms but redesigned their support model around AI-first resolution with human escalation for complex cases. ## Key Takeaways - Switch triggers: pricing unpredictability, weak AI automation, slow iteration - AI-native platforms deliver 50-70% containment vs. Zendesk's significantly lower legacy bot rates - Migration takes 2-6 weeks; run parallel for 2-4 weeks to maintain SLA continuity - Match the alternative to your pain point: cost (Freshdesk), AI depth ([Lorikeet](https://www.lorikeetcx.ai)), or scale (Salesforce) ## Frequently Asked Questions ### How much can I save by switching from Zendesk? It depends on where you switch. Freshdesk and Help Scout save 25-40% on licensing costs alone. AI-native platforms may cost similar or more in platform fees but reduce total support costs by 40-60% through automated resolution - fewer human agents needed for routine tickets. ### Can I migrate my Zendesk ticket history to a new platform? Yes, most alternatives support importing Zendesk ticket history via CSV export or API. Full conversation threads, customer profiles, and metadata typically transfer cleanly. Custom fields and complex automation rules need manual recreation on the new platform. ### What if I want to keep Zendesk but add better AI? You can layer an AI agent platform on top of Zendesk rather than replacing it. Platforms like [Lorikeet](https://www.lorikeetcx.ai) integrate with Zendesk via API, handling AI-resolvable tickets before they reach your human queue. This approach keeps your existing workflow while adding autonomous resolution capabilities. The Zendesk alternative you choose should solve the specific problem driving your search. If it's cost, simpler platforms work. If it's AI automation depth, only AI-native platforms will move the needle. And if you're not ready to switch entirely, layering an AI agent on top of your current Zendesk setup is the lowest-risk path to higher resolution rates. Start with a clear list of what's not working. Match that to platform capabilities. Pilot before you migrate. **Want Zendesk-level reliability with AI that actually resolves tickets? [See how Lorikeet works alongside or instead of your current helpdesk.](https://www.lorikeetcx.ai)** --- # Zendesk Competitors 2026: 6 Alternatives Compared by AI Resolution URL: https://www.lorikeetcx.ai/articles/zendesk-competitors Zendesk competitors are customer service platforms that compete with Zendesk for helpdesk, ticketing, and AI-powered support market share. The competitive landscape shifted significantly in 2024-2025 as AI-native platforms entered the market and established players added autonomous agent capabilities. According to [G2's 2025 market data](https://www.g2.com/), Zendesk holds the largest market share in helpdesk software but faces growing competition from both traditional players and AI-first entrants. - Major competitors: Freshdesk, Intercom, Salesforce Service Cloud, HubSpot, and AI-native platforms - The biggest competitive gap: AI resolution depth, not ticketing features - Zendesk's per-agent pricing ($55-169/mo per seat) faces pressure from per-resolution models ($0.99-2.00 per ticket) - Migration between platforms takes 2-6 weeks with parallel operation recommended Zendesk built the modern helpdesk category. But categories evolve. The competitors worth watching in 2026 aren't the ones copying Zendesk's ticketing model with minor tweaks. They're the ones rethinking what a support platform should do when AI can resolve most routine tickets without a human. This guide maps the competitive landscape by what matters most: which platforms actually resolve customer issues vs. which ones just manage the queue. ## Who Are Zendesk's Main Competitors in 2026? Zendesk competes against 3 categories: traditional helpdesk platforms (Freshdesk, Help Scout), all-in-one CX platforms (Salesforce, HubSpot, Intercom), and AI-native resolution platforms (Lorikeet, Ada, Forethought). Each category competes on a different dimension. 1. **Freshdesk (Freshworks).** The closest direct competitor. Similar ticketing, knowledge base, and automation features at 30-40% lower pricing. [Freshworks](https://www.freshworks.com/) reports 50,000+ companies on the platform. Freddy AI handles basic automation but lacks autonomous resolution on complex workflows. 2. **Intercom.** Shifted from messaging tool to AI-first platform. [Fin AI agent](https://www.intercom.com/pricing) charges $0.99 per resolution and claims 50%+ containment. Best for product-led companies wanting conversational support and marketing in one tool. Weaker on voice and complex multi-system workflows. 3. **Salesforce Service Cloud.** The enterprise standard. Deep CRM integration, Einstein AI, and omnichannel routing. Best for teams already invested in Salesforce. Significantly more expensive and complex to implement than Zendesk. 4. **[Lorikeet](https://www.lorikeetcx.ai).** AI-native platform built for complex, multi-step customer service. Connects to backend systems to execute workflows (refunds, billing changes, account updates) autonomously. Pairs with existing helpdesks rather than replacing them. Built-in QA reviews 100% of AI interactions. ## How Does Zendesk's AI Compare to Its Competitors? Zendesk launched outcome-based pricing for AI agents in 2025, charging $1.50-2.00 per automated resolution on top of per-seat licensing. Competitors are split between those matching this model and those with fundamentally different AI architectures. Zendesk's AI strength is breadth - it works across the existing Zendesk ecosystem with minimal setup. The weakness is depth. Complex workflows requiring multi-system actions (processing refunds through Stripe, updating orders in Shopify, adjusting subscriptions) still frequently escalate to human agents. Intercom's Fin handles conversational resolution well but similarly struggles with deep backend actions. AI-native competitors like Lorikeet and Ada are purpose-built for these complex workflows, achieving higher containment rates on action-heavy ticket types. The gap is narrowing as all platforms invest heavily in AI, but architecture matters - platforms built around AI from day one have structural advantages over those adding AI to existing ticketing systems. ## What Are the Key Pricing Differences? Zendesk uses a hybrid model: per-agent seat pricing ($55-169/month) plus per-resolution AI pricing ($1.50-2.00). Competitors vary widely, from pure per-seat (Freshdesk, $15-79/agent) to pure per-resolution (Intercom Fin at $0.99) to custom enterprise pricing (Salesforce). The pricing shift matters strategically. Per-agent models charge you for headcount, which means costs scale with team size regardless of automation. Per-resolution models charge for outcomes, which means costs decrease as AI handles more volume. For a team handling 5,000 tickets/month where AI resolves 50%, the math changes dramatically. Per-agent pricing stays flat whether AI resolves 10% or 60%. Per-resolution pricing rewards higher automation. [Zendesk's hybrid approach](https://www.eesel.ai/blog/understanding-zendesk-ai-pricing-a-complete-pay-per-resolution-guide) charges both, which some teams find penalizes them on both dimensions. ## What Results Do Teams See When Switching from Zendesk? The outcome depends entirely on what you switch to and why. Teams moving to cheaper traditional platforms see cost savings. Teams moving to AI-native platforms see cost savings plus resolution improvements. Teams moving to Salesforce see enterprise capabilities at higher total cost. Freshdesk switches save 25-40% on licensing with comparable feature sets. Intercom switches show 30-50% containment improvement for conversational support-heavy teams. AI-native platform switches show the largest resolution gains - teams report containment rates doubling from 20-30% (Zendesk AI) to 50-70% (purpose-built AI agents) on complex ticket types. Salesforce switches rarely save money but improve cross-team visibility and CRM integration for enterprise operations handling 50,000+ tickets monthly. The riskiest move is switching platforms while simultaneously changing your support model. Migrate first, optimize later. Run parallel systems for 2-4 weeks to maintain SLA continuity. ## How Should You Evaluate Zendesk vs. Its Competitors? Start with your pain point, not a feature checklist. The right competitor depends on what's broken: pricing, AI depth, simplicity, or enterprise scale. No single platform wins across all dimensions. Build an evaluation around 3 questions. First: what percentage of your tickets can be fully automated? If it's under 30%, AI depth doesn't matter yet and cost optimization (Freshdesk) makes sense. If it's 30-60%, AI-native platforms ([Lorikeet](https://www.lorikeetcx.ai)) deliver the highest ROI. Second: do you need a full helpdesk replacement or an AI layer on top? Some teams keep Zendesk for ticketing and add an AI agent for resolution. Third: what's your integration complexity? Teams with 5+ backend systems need deep integration capabilities, which favors platforms with mature API ecosystems. ## Key Takeaways - Zendesk's main competitors split into 3 categories: cheaper (Freshdesk), AI-first (Lorikeet, Ada), enterprise (Salesforce) - Per-resolution pricing ($0.99-2.00) is replacing per-agent pricing ($55-169/mo) across the market - AI-native platforms show 2-3x higher containment rates on complex tickets vs. Zendesk's built-in AI - Evaluate based on your pain point: cost, AI depth, or enterprise scale ## Frequently Asked Questions ### Is Freshdesk really cheaper than Zendesk? Yes, for comparable features. [Freshdesk's](https://www.freshworks.com/) per-agent pricing runs 30-40% lower than Zendesk's equivalent tiers. The gap widens when you add Zendesk's AI pricing on top. However, total cost depends on your specific feature needs and agent count. ### Can I use an AI agent platform alongside Zendesk instead of replacing it? Yes. Platforms like Lorikeet integrate with Zendesk via API, sitting in front of your existing queue. The AI resolves tickets it can handle; the rest flow to your Zendesk-based human team unchanged. This is often the lowest-risk path to improving resolution rates. ### Which Zendesk competitor has the best AI? It depends on your ticket types. For conversational support and FAQ resolution, Intercom Fin performs well. For complex, multi-step workflows requiring backend system actions (refunds, billing, account changes), AI-native platforms like Lorikeet deliver the deepest resolution capabilities. The helpdesk market is fragmenting. The Zendesk-or-nothing era is over. But switching for the sake of switching rarely pays off. Define your specific problem, evaluate 2-3 alternatives against that problem, and pilot before committing. The best choice is the one that solves your actual pain point, not the one with the longest feature list. **Explore how [Lorikeet](https://www.lorikeetcx.ai) works alongside Zendesk to resolve complex tickets autonomously, or as a standalone AI agent for your support operation.** --- # AI Agent for Customer Service: How It Works and What to Expect URL: https://www.lorikeetcx.ai/articles/ai-agent-for-customer-service An AI agent for customer service is an autonomous system that resolves support tickets by understanding customer intent, accessing backend data, and executing actions like refunds, account updates, and order modifications. Unlike chatbots that follow scripts, AI agents reason through multi-step problems. [Gartner predicts](https://www.gartner.com/en/newsroom/press-releases/2025-03-05-gartner-predicts-agentic-ai-will-autonomously-resolve-80-percent-of-common-customer-service-issues-without-human-intervention-by-20290) agentic AI will resolve 80% of common customer service issues without human intervention by 2029. - AI agents execute actions (process refunds, change subscriptions) while chatbots surface information - Deployment takes 2-6 weeks with measurable containment within the first 30 days - Per-resolution pricing ($0.99-2.00/ticket) replaces per-agent models at leading platforms - Quality assurance and guardrails are critical for accuracy on high-stakes customer actions Customer service teams are drowning in routine tickets. Password resets, refund requests, order tracking, subscription changes - the same 10 ticket types make up 70-80% of most queues. Human agents handle them competently but slowly, and the repetition drives attrition. AI agents are built to absorb that routine volume. Not by deflecting customers to a FAQ page, but by doing what a human agent would do: pull up the account, check the policy, take the action, confirm with the customer. The difference is speed, consistency, and scale. ## How Does an AI Customer Service Agent Actually Work? An AI agent receives a customer message, interprets the intent using a large language model, retrieves relevant data from your backend systems, applies your business policies, executes the required action, and responds to the customer. This entire chain happens in seconds. The technical architecture involves 3 layers. The understanding layer uses natural language processing to parse what the customer wants, even when the request is vague or spans multiple issues. The reasoning layer checks policies, evaluates eligibility, and determines the correct action. The execution layer connects to your systems - Shopify, Stripe, Salesforce, your OMS - and performs the action. For example, "I got charged twice" triggers the agent to pull billing records from Stripe, identify the duplicate charge, check refund eligibility, process the refund, and confirm the credit timeline. No human touches the ticket. ## What Types of Customer Issues Can AI Agents Resolve? AI agents handle structured, policy-driven issues with clear resolution paths. This includes order management, billing inquiries, subscription changes, account updates, returns and refunds, shipping modifications, and product information requests. 1. **Order management.** Tracking, modifications, cancellations, and replacements. AI agents connect to your OMS and shipping providers to give real-time answers and take immediate action, cutting order-related handle time by 50-70%. 2. **Billing and refunds.** Duplicate charges, refund processing, payment method updates, and invoice questions. The agent accesses your payment system, validates the issue, and executes the resolution within policy guardrails. 3. **Subscription management.** Upgrades, downgrades, cancellations, and plan changes. AI agents handle the full workflow including proration calculations, confirmation emails, and retention offers when configured. 4. **Account and profile updates.** Address changes, contact information, password resets, and preference updates. These high-volume, low-complexity tickets are ideal for full automation. ## What Results Should You Expect from Deploying an AI Agent? Expect measurable improvements within 30-60 days of deployment. The first metric to move is containment rate - the percentage of tickets resolved without human involvement. Most teams see this climb steadily as the AI handles more ticket categories. Containment rates reach 40-60% within the first quarter for teams deploying across 3-5 ticket categories. Cost per resolution drops from $5-12 (fully human) to $1-3 (AI-resolved), according to industry benchmarks. Average handle time for AI-resolved tickets runs 1-3 minutes vs. 8-12 minutes with human agents. [McKinsey notes](https://www.mckinsey.com/capabilities/operations/our-insights/beyond-the-bot-building-empathetic-customer-experiences-with-agentic-ai) that AI in customer care can improve satisfaction scores by 5-10% when resolution quality is maintained. The compounding effect: as AI handles more routine tickets, human agents take fewer but more complex cases. Their response times improve. Quality on escalated tickets goes up. Agent satisfaction increases because they're solving interesting problems instead of processing repetitive requests. ## What Should You Look for in an AI Agent Platform? Prioritize 3 capabilities: integration depth with your existing systems, policy guardrails that control what the AI can do, and transparent resolution metrics that separate actual resolution from deflection. Integration depth is non-negotiable. If the AI agent can't read and write to your order management, billing, and CRM systems, it's a lookup tool, not an agent. Platforms like [Lorikeet](https://www.lorikeetcx.ai) connect to your backend to execute complete workflows. Guardrails matter equally - you need fine control over refund limits, escalation triggers, and [quality assurance](https://www.lorikeetcx.ai/product/coach) that audits every AI interaction, not just sampled ones. Finally, demand real containment data during evaluation. Run a pilot with 200+ real tickets and measure how many the AI actually resolves vs. how many it deflects to your team. ## How Do You Deploy an AI Agent Without Disrupting Your Team? Start with a single high-volume, low-risk ticket category. Expand one category at a time as the team builds confidence and the AI demonstrates accuracy. Most full deployments take 2-4 months across all ticket types. The deployment pattern that works: pick your highest-volume, most repetitive ticket type (usually order tracking or password resets). Deploy the AI agent on that category only, routing all other tickets to your human team as usual. Monitor for 2 weeks, review accuracy with your QA process, then add the next category. This staged approach avoids the big-bang risk and lets your team adjust gradually. Budget 2-6 weeks for the initial category, including integration setup, policy configuration, and testing. Each subsequent category adds 1-2 weeks. ## Key Takeaways - AI agents execute actions across your systems, not just answer questions from a knowledge base - Expect 40-60% containment within 90 days and cost per resolution under $3 - Start with one high-volume ticket category and expand weekly - Audit AI interactions with [continuous QA](https://www.lorikeetcx.ai/product/coach) to maintain accuracy at scale ## Frequently Asked Questions ### How much does an AI customer service agent cost? Most platforms charge $0.99-2.00 per resolved ticket, with some offering monthly seat-based pricing ($100-500/month). For a team handling 3,000 tickets/month at 50% AI containment, expect $1,500-3,000/month in AI costs - typically less than the cost of 1 full-time human agent. ### How accurate are AI agents at resolving customer issues? Accuracy depends on the platform, your policy configuration, and the ticket type. Well-configured AI agents on structured ticket types (order tracking, refunds) achieve 90-95% accuracy. Complex, ambiguous cases score lower. The key is monitoring every interaction and catching errors before customers do. ### Can an AI agent work with my existing helpdesk? Yes. Most AI agent platforms integrate with Zendesk, Freshdesk, Salesforce Service Cloud, and Intercom via API. The AI agent sits upstream, resolving tickets it can handle and routing the rest to your existing human queue. No helpdesk replacement required. AI agents for customer service are not theoretical - they're in production at thousands of companies handling millions of tickets. The question isn't whether they work. It's whether your team can afford not to deploy them while competitors automate 40-60% of their support volume. Start small. Start with your highest-volume ticket type. Measure results. Expand from there. **See how [Lorikeet's AI agents](https://www.lorikeetcx.ai) resolve billing, order, and account tickets end-to-end. Built for complex workflows, not just FAQ deflection.** --- # Best AI Chatbot for Customer Service in 2026 URL: https://www.lorikeetcx.ai/articles/best-ai-chatbot-for-customer-service The best AI chatbot for customer service is one that resolves tickets - not one that deflects them to your human team. Only [14% of customer service issues](https://www.gartner.com/en/newsroom/press-releases/2024-08-19-gartner-survey-finds-only-14-percent-of-customer-service-issues-are-fully-resolved-in-self-service) are fully resolved by self-service tools today, per Gartner. Most chatbots are expensive FAQ pages. This guide compares what actually works. - Intercom Fin resolves 51-66% of queries on average at $0.99 per resolution - strong for mid-market - Zendesk AI charges $1.50-2.00 per resolution and fits best inside existing Zendesk environments - Ada starts around $30,000/year and targets enterprise volume with custom models - Lorikeet handles complex multi-step workflows across backend systems other tools escalate If you manage support, you have seen the same vendor demo 15 times. The chatbot nails the easy ticket - shipping status, password reset. Then you deploy it, and 60-80% of real tickets hit a wall because they need order data, policy checks, and actions across 3 systems. We compare 6 platforms on resolution capability, pricing, and honest trade-offs. ## What Should You Measure When Comparing AI Chatbots? Measure resolution rate, not deflection rate. A resolved ticket means the customer's problem is fixed without a human stepping in. A deflected ticket means the bot answered something, but the customer still needs help. Most vendor dashboards blur this line. Ask 2 questions: what percentage of conversations end with the issue fully resolved, and how do you define "resolved"? Intercom counts a resolution when the customer confirms or stops replying after 72 hours. Zendesk uses a similar inactivity window. A customer who gives up and calls your phone line still counts as "resolved" in many systems. ## How Do the Top 6 AI Chatbots Compare on Price? Pricing models vary wildly - per resolution, per conversation, per seat, or annual contracts. The sticker price almost never matches what you actually pay. Here is how 6 options break down. 1. **Intercom Fin.** $0.99/resolution on top of plans starting at $29/month. Average resolution rate around 51-66% across its customer base. Transparent, but costs scale linearly with volume. 2. **Zendesk AI.** $1.50/resolution (committed) or $2.00 (pay-as-you-go), on top of Suite plans from $115/agent/month. Best for existing Zendesk shops. 3. **Ada.** No public pricing. Starts around $30,000/year per their [Salesforce AppExchange listing](https://appexchange.salesforce.com/appxListingDetail?listingId=a0N3A00000FYg6tUAD), but median buyers pay roughly $70,000/year per Vendr data. 4. **Forethought.** Quote-based. Blends platform fees with per-deflection costs. Best for teams wanting AI triage alongside resolution. 5. **Tidio (Lyro AI).** $29/month plus $39/month for AI add-on. Built for SMBs and Shopify - not designed for complex backend workflows. 6. **Lorikeet.** Purpose-built for multi-step CX workflows. Connects to backend systems to execute refunds, account changes, and policy lookups autonomously. ## Which AI Chatbot Handles Complex Tickets Best? Complex tickets require reading from and writing to backend systems - order management, billing, CRM. Most chatbots fail here because they were built for FAQ lookups, not workflow execution. Intercom Fin and Zendesk AI handle moderate complexity well - knowledge base answers, product questions, basic account lookups. But when a ticket needs checking refund policy, verifying eligibility, processing the return, and confirming - that multi-step chain is where general-purpose tools escalate. [Lorikeet](https://www.lorikeetcx.ai) is built for this layer. It connects to your systems, follows your business logic, and executes full workflows. For CX teams in fintech or e-commerce where tickets touch 3-5 systems, that architectural difference determines whether AI handles 30% or 70% of volume. ## What Resolution Rates Should You Realistically Expect? Vendor marketing shows 80%+ resolution rates from cherry-picked case studies. Real-world averages are lower. Here is what the data shows without the outliers. Intercom reports Fin's average resolution at 51-66% across 6,000+ customers, with the top 20% above 80%. Your results depend on knowledge source quality and ticket complexity. Tidio's Lyro works for scripted FAQ flows but drops off on anything requiring backend access. [McKinsey research](https://www.mckinsey.com/capabilities/operations/our-insights/beyond-the-bot-building-empathetic-customer-experiences-with-agentic-ai) shows generative AI improves agent productivity by 14% in issue resolution per hour - but that depends on whether the AI resolves or merely assists. ## How Do You Avoid Overpaying for AI That Underdelivers? The biggest cost trap is paying per resolution for tickets that would have self-resolved anyway. If your bot "resolves" a password reset your help center already covers, you are paying $0.99-2.00 for zero incremental value. Run a 30-day audit first. Categorize tickets into 3 buckets: simple FAQ, moderate (data lookup), and complex (multi-step actions). If 60%+ is simple, a better help center might be the right move. If complex tickets dominate, you need workflow execution - and pricing that reflects genuine resolution, not FAQ redirection. ## What Questions Should You Ask in Every Vendor Demo? Go beyond the polished script. These questions separate tools that resolve from tools that deflect - regardless of platform. - Show me a ticket that required 3+ system lookups and an action - not just an answer - What is your median resolution rate across all customers, not just top performers? - How do you define "resolution" - customer confirmation, timeout, or something else? - What happens when the AI is not confident - clean escalation or guessing? Any vendor that dodges these is selling deflection. Tools built for complex tickets - like [AI agents for customer service](/articles/ai-agent-for-customer-service) - welcome the scrutiny. ## Key Takeaways - Only 14% of service issues are fully resolved by self-service today - most chatbots deflect, not resolve (Gartner) - Intercom Fin costs $0.99/resolution at 51-66% average; Zendesk AI charges $1.50-2.00/resolution - Complex multi-step tickets require backend system access most general-purpose chatbots lack - Audit your ticket mix before buying - 60%+ simple volume may not need an AI agent - Ask for median resolution rates across all customers, not cherry-picked case studies ## Frequently Asked Questions ### What is the difference between an AI chatbot and an AI agent? A chatbot matches input to scripted responses or knowledge base articles. An [AI agent](/articles/ai-agent-vs-chatbot) reasons through problems, accesses backend systems, and takes actions like processing refunds. The distinction matters for complex tickets requiring multi-step workflow execution, not just information retrieval. ### How much does an AI chatbot for customer service cost? From $29/month for basic tools like Tidio to $70,000+/year for enterprise platforms like Ada. Per-resolution models like Intercom ($0.99) and Zendesk ($1.50-2.00) scale with volume. Always factor in base platform fees, agent seats, and AI add-ons when comparing total cost. ### Can AI chatbots handle refunds and account changes automatically? Most cannot. Standard chatbots answer questions but do not write to backend systems. AI agents with backend integrations can process refunds, update accounts, and execute policy-based decisions autonomously. The capability depends on integration architecture and workflow engine depth. The AI chatbot market is noisy and full of inflated metrics. If your team handles straightforward queries, a lighter tool or well-configured help center does the job. If tickets require multi-step actions across billing, orders, and account systems, you need workflow depth. Evaluate vendors against your hardest tickets, not your easiest. Run a 30-day audit, know your complexity mix, and demand median resolution data - not best-case showcases. **See how Lorikeet resolves complex multi-step tickets across backend systems. [Explore Lorikeet's AI agent](https://www.lorikeetcx.ai) built for CX teams that need resolution, not deflection.** --- # AI for Customer Support: What Works, What Doesn't, and Where It's Headed URL: https://www.lorikeetcx.ai/articles/ai-for-customer-support AI for customer support is the use of artificial intelligence - from FAQ bots and agent copilots to fully autonomous AI agents - to handle, route, and resolve customer service requests. [Gartner predicts](https://www.gartner.com/en/newsroom/press-releases/2025-03-05-gartner-predicts-agentic-ai-will-autonomously-resolve-80-percent-of-common-customer-service-issues-without-human-intervention-by-20290) agentic AI will autonomously resolve 80% of common service issues by 2029, up from under 5% today. - AI in support spans four categories: knowledge base automation, copilots, autonomous agents, and QA - Most "AI" support tools still deflect rather than resolve - the distinction matters for ROI - Production-ready autonomous agents now handle multi-step workflows across backend systems The term "AI for customer support" gets thrown around loosely. A keyword-matching FAQ bot and a system that processes refunds, updates billing, and writes personalized follow-ups both qualify. That ambiguity is the problem. CX leaders end up buying deflection engines dressed as automation. The reality in 2026 is that AI support tools sit on a maturity spectrum, and understanding where each category falls is the difference between cutting costs and cutting corners. ## What Are the Four Categories of AI in Customer Support? AI in customer support falls into four categories: knowledge base automation, agent assist and copilots, autonomous agents, and quality assurance. Each solves a different problem, and most mature CX operations deploy two or more. **Knowledge base automation** uses AI to generate, update, and surface help articles - reducing ticket volume through better self-service. **Agent assist and copilots** sit alongside human agents, suggesting responses and drafting replies. **Autonomous agents** handle tickets end-to-end without human involvement - reasoning through problems, accessing systems, and taking actions. **Quality assurance** AI scores agent performance and flags policy violations across 100% of interactions. ## What Is Production-Ready vs. Hype in AI Support? Knowledge base automation and copilots are mature and widely deployed. Autonomous agents are production-ready but require careful scoping - they work best on defined workflow types. Fully general "handle anything" AI remains aspirational. If a vendor claims 90% resolution across all ticket types on day one, they're measuring deflection or overpromising. Teams seeing real results start with 5-10 high-volume workflows, prove resolution rates, and expand. [McKinsey research](https://www.mckinsey.com/capabilities/operations/our-insights/beyond-the-bot-building-empathetic-customer-experiences-with-agentic-ai) shows generative AI improves customer satisfaction by 5-10% in customer care - meaningful, but not the overnight transformation some vendors promise. ## How Do Autonomous Agents Differ from Copilots? Copilots assist human agents by surfacing information and drafting responses. Autonomous agents replace the human in the loop entirely for specific workflows - they read context, make decisions, execute actions in backend systems, and confirm resolution with the customer. A copilot still requires a human to review and send, so staffing costs stay roughly the same. An [autonomous AI agent](/articles/ai-agent-vs-chatbot) eliminates the human step for qualifying tickets, dropping per-ticket costs dramatically. The trade-off is trust - you need guardrails, policy enforcement, and QA systems to ensure correctness without human review on every response. ## Where Do FAQ Bots Still Make Sense? FAQ bots work for simple, static queries where the answer rarely changes - store hours, return windows, shipping cutoffs. They're cheap, fast to maintain, and don't need backend integrations. If most of your volume is single-turn lookups, a bot handles it. The problem arises when teams stretch bots beyond their design. Once a customer needs account data, order history, or a policy judgment, the bot either hallucinates or escalates - creating a worse experience than no bot at all. ## How Should You Evaluate AI Support Tools? Evaluate on resolution rate - the percentage of tickets the AI fully resolves without human involvement. Deflection rate and "conversations handled" are vanity metrics if the customer still ends up waiting for a human agent. 1. **Measure resolution, not deflection.** A tool that "handles" 80% of conversations but resolves 25% is an expensive FAQ page. Demand end-to-end resolution data. 2. **Check integration depth.** Can the tool read and write to your order management, billing, and CRM? Platforms like [Lorikeet](https://www.lorikeetcx.ai) connect to backend systems to execute workflows, not just surface information. 3. **Require policy guardrails.** Autonomous systems need enforceable rules - refund limits, escalation triggers, compliance constraints. 4. **Audit at 100%, not 2%.** AI-powered QA tools like [Lorikeet Coach](https://www.lorikeetcx.ai/product/coach) score every interaction for accuracy, tone, and policy compliance. ## What Metrics Should You Track After Deployment? Track five metrics in the first 90 days: autonomous resolution rate, average handle time, CSAT on AI-handled vs. human-handled tickets, escalation rate, and cost per resolved ticket. These tell you whether the AI is working or creating the appearance of efficiency. Resolution rates for well-scoped autonomous agents land between 40-60% of qualifying volume within the first quarter. Handle time drops from 8-12 minutes to 2-4 minutes on AI-resolved tickets. [McKinsey's data](https://www.mckinsey.com/capabilities/operations/our-insights/beyond-the-bot-building-empathetic-customer-experiences-with-agentic-ai) suggests 5-10% CSAT improvement is a reasonable benchmark. Cost per resolved ticket typically drops 50-70% compared to fully human resolution. ## Why Does QA Matter More with AI in the Loop? When humans handle tickets, QA catches training gaps. When AI handles tickets, QA catches system-level failures that can affect thousands of customers simultaneously. A single misconfigured policy can repeat at scale before anyone notices. Traditional QA reviews 1-3% of conversations. With AI generating responses at volume, you need 100% coverage. AI-powered QA flags policy violations and incorrect actions in real time, giving CX leaders visibility into every interaction. This is where an [AI agent platform](/articles/ai-agent-platform) with built-in quality scoring becomes essential. ## Key Takeaways - AI support spans four categories - knowledge base, copilots, autonomous agents, and QA - each with different ROI profiles - Autonomous agents resolve 40-60% of qualifying tickets in the first 90 days when scoped to specific workflows - McKinsey data shows 5-10% CSAT improvement from generative AI in customer care - Gartner projects 80% autonomous resolution of common issues by 2029 - AI-powered QA auditing 100% of conversations is critical when AI generates responses at scale ## Frequently Asked Questions ### Is AI for customer support just chatbots with better marketing? No. Modern AI support tools include copilots that draft responses for human agents, autonomous agents that resolve tickets end-to-end by taking actions in backend systems, and QA tools that audit every conversation. Chatbots are one category on a much broader spectrum. ### How long does it take to deploy AI customer support? FAQ bots deploy in days. Copilots take 1-3 weeks. Autonomous agents need 3-6 weeks for system integrations, policy configuration, and testing. The timeline depends on how many backend systems need connecting and how many workflow types you're automating. ### Can AI handle sensitive or regulated customer interactions? Yes, with proper guardrails. AI agents enforce compliance rules more consistently than humans because policy logic is codified, not memorized. Choose a platform with auditable decision logs, configurable escalation rules, and QA that reviews 100% of interactions. AI for customer support has moved past the chatbot era into something more useful - and more complex. The teams getting results aren't chasing a single tool that does everything. They're layering knowledge automation for self-service, copilots for productivity, autonomous agents for high-volume workflows, and AI-powered QA to keep it all honest. The maturity spectrum matters. Start with the category that matches your biggest pain point - ticket volume, handle time, quality consistency, or staffing costs. Scope tightly, measure resolution, and expand from there. **See how Lorikeet combines autonomous AI agents with built-in quality assurance to resolve customer issues end-to-end. [Explore the platform](https://www.lorikeetcx.ai).** --- # AI Chatbots for Customer Service: From Script Followers to Autonomous Agents URL: https://www.lorikeetcx.ai/articles/ai-chatbots-for-customer-service AI chatbots for customer service are automated systems that handle support conversations - from simple FAQ lookups to complex issue resolution. [IBM reports](https://www.ibm.com/think/topics/chatbots) chatbots can reduce customer support costs by up to 30%. But what qualifies as a "chatbot" in 2026 looks nothing like what it meant five years ago. - Rule-based chatbots follow scripts and decision trees - they deflect, not resolve - NLU-based bots improved intent matching but still could not take action - LLM-powered AI agents reason through problems and execute multi-step workflows - The term "AI chatbot" now covers everything from basic FAQ bots to autonomous agents Search "AI chatbots for customer service" today and you will find vendors selling three very different products under the same label. Some are glorified FAQ pages. Some match intent to canned responses. And some connect to your backend systems and resolve issues without a human. The confusion is not accidental - it benefits companies selling outdated technology under a modern name. Here is what has actually changed and how to tell the difference. ## What Are the Three Generations of AI Chatbots? Customer service chatbots have gone through three distinct phases. Each generation addressed limitations of the last, but the jumps between them are architectural, not incremental. Understanding which generation a vendor is selling you matters more than any feature checklist. The first generation used rule-based decision trees. If a customer went off-script, the bot broke. The second generation added natural language understanding - bots recognized intent even when phrasing varied, but still pulled from a fixed response library. The third generation - LLM-powered AI agents - can reason, maintain context, access backend systems, and execute tasks autonomously. ## Why Did Rule-Based Chatbots Stop Working? Rule-based chatbots stopped working because customer expectations outgrew them. Customers describe problems in their own words, provide partial information, and expect the bot to figure out the rest. Decision trees cannot handle that variability. Maintaining these bots became unsustainable at scale too. Every new product, policy change, or edge case required manual branch updates. Teams spent more time maintaining the bot than the bot saved in agent hours. ## What Can LLM-Powered Chatbots Do That Older Ones Could Not? LLM-powered chatbots - more accurately called AI agents - take actions, not just provide answers. They process refunds, modify orders, update account details, and troubleshoot technical issues across multiple backend systems in a single conversation. They also maintain context. A customer can say "actually, cancel the other one instead" and the agent understands what "the other one" refers to. Beyond that, AI agents reason through company policy - evaluating the customer's situation against your rules and making a decision, the same way a trained human agent would. ## Is There a Real Difference Between an AI Chatbot and an AI Agent? Yes. An [AI agent](/articles/ai-agent-vs-chatbot) is autonomous - it understands a request, decides what to do, and executes. A chatbot responds to input with pre-mapped output. The industry uses both terms loosely, which creates confusion for buyers evaluating tools. The practical test is simple: can the tool take action in your systems? If it can process a return, adjust billing, or escalate with full context - that is an agent. If it links to a help article and says "was this helpful?" - that is a chatbot wearing a new label. ## What Should You Measure When Evaluating AI Chatbots? Measure automated resolution rate - the percentage of customer issues fully resolved by AI without human intervention. Not containment rate, not deflection rate. Resolution is the only metric that maps to cost savings and customer satisfaction. [Gartner predicts](https://www.gartner.com/en/newsroom/press-releases/2025-03-05-gartner-predicts-agentic-ai-will-autonomously-resolve-80-percent-of-common-customer-service-issues-without-human-intervention-by-20290) agentic AI will autonomously resolve 80% of common customer service issues by 2029. Companies deploying third-generation AI agents see handle time reductions of 40-60% compared to chatbot-only setups. CSAT improves not because the conversation feels better, but because the problem actually gets solved. ## What Mistakes Do Companies Make When Deploying AI Chatbots? The most common mistake is deploying a second-generation chatbot and expecting third-generation results. Intent-matching bots are marketed as "AI-powered," which is technically true but practically misleading. 1. **Optimizing for containment.** Containment measures how many conversations the bot kept away from humans - not whether the problem was solved. High containment with low resolution means you are frustrating people faster. 2. **Skipping backend integrations.** A chatbot without access to order management, CRM, or billing is limited to answering questions. Resolution requires read and write access to customer systems. 3. **Ignoring guardrails.** LLM-powered agents need policy boundaries. Look for platforms with built-in [quality assurance and coaching](https://www.lorikeetcx.ai/product/coach) that audits every interaction. 4. **Treating all tickets the same.** Simple queries - store hours, password resets - are fine for basic automation. Route complexity to the agent, simplicity to the bot. ## How Do You Choose the Right AI Chatbot for Your Team? Start by mapping your ticket types. If most of your volume involves multi-step actions - returns, billing disputes, account changes - you need an AI agent, not a chatbot. Most companies find their mix skews toward complexity. Ask vendors for resolution data, not engagement metrics. Ask whether the tool connects to your existing systems. The [best AI chatbot for customer service](/articles/best-ai-chatbot-for-customer-service) in 2026 is one that resolves the issues your customers bring - and proves it with data. ## Key Takeaways - AI chatbots have evolved through three generations - only LLM-powered agents can take autonomous action - IBM reports chatbots reduce support costs by up to 30%; actual savings depend on resolution rate, not deflection - Gartner predicts 80% autonomous resolution of common issues by 2029 - but only with agentic AI, not traditional bots - Measure automated resolution rate, not containment - deflection is not resolution ## FAQ ### Are AI chatbots replacing human customer service agents? Not replacing - redistributing work. AI chatbots handle repetitive, high-volume issues so human agents focus on complex cases requiring judgment. The goal is fewer agents doing low-value work, not zero agents. Teams that deploy AI agents well typically reassign staff rather than reduce headcount. ### How much do AI chatbots for customer service cost? Basic chatbots cost $50-500 per month. LLM-powered AI agents typically charge per resolved conversation - usually $0.50-3.00 each. The per-resolution model often costs less at scale because you only pay when the AI solves the problem, not when it deflects. ### How long does it take to set up an AI chatbot? Rule-based chatbots deploy in days with template scripts. AI agents take 2-6 weeks including system integrations, policy configuration, and testing. The longer setup reflects the difference - you are connecting AI to backend systems, not uploading FAQ content. ### What is the difference between a chatbot and an AI agent? A chatbot follows scripts and returns pre-mapped responses. An AI agent reasons through problems, accesses backend systems, and takes actions to resolve issues end-to-end. When evaluating tools, ask whether it can execute tasks, not just answer questions. ## Conclusion The phrase "AI chatbots for customer service" covers a wide range of tools - from basic script-followers to autonomous agents that resolve issues end-to-end. The label has not kept up with the technology. What matters is not what a vendor calls their product, but what it can actually do when a customer needs help. If your current chatbot deflects more than it resolves, the issue is not with AI - it is with which generation of AI you are running. The move from intent-matching to autonomous resolution is the biggest shift in customer service automation since chatbots first appeared. And it is happening now. **[Lorikeet](https://www.lorikeetcx.ai) is an AI agent built for complex CX - it connects to your systems, reasons through your policies, and resolves issues autonomously. See what third-generation AI customer service looks like.** --- # AI Customer Service Agent: How Autonomous Resolution Works URL: https://www.lorikeetcx.ai/articles/ai-customer-service-agent An AI customer service agent is software that autonomously resolves support tickets by reasoning through customer requests, accessing backend systems, and executing actions - like refunds, subscription changes, and order edits - without human involvement. [Gartner predicts](https://www.gartner.com/en/newsroom/press-releases/2025-03-05-gartner-predicts-agentic-ai-will-autonomously-resolve-80-percent-of-common-customer-service-issues-without-human-intervention-by-20290) agentic AI will autonomously resolve 80% of common customer service issues by 2029. - AI customer service agents resolve tickets end-to-end by taking actions, not just surfacing answers - They use a 3-layer architecture: understanding (LLM), reasoning (policy engine), and execution (API actions) - Per-resolution pricing ($0.99-2.00/ticket) typically costs less than one full-time human agent - They complement human agents by absorbing routine volume, not replacing judgment on complex cases Most support teams know the pattern. The same 10-15 ticket types fill the queue day after day - refund requests, order status checks, subscription downgrades, address changes. Human agents handle them accurately but slowly, and the repetition drives burnout. AI customer service agents exist to absorb that routine volume. Not by pointing customers at an FAQ page, but by doing the actual work: pulling up the account, checking policy, processing the action, confirming with the customer. The difference from a chatbot isn't cosmetic. It's architectural. ## How Does an AI Customer Service Agent Work Technically? An AI customer service agent operates on a 3-layer architecture. The understanding layer uses a large language model to parse customer intent from natural language. The reasoning layer applies your business policies to determine the correct action. The execution layer connects to your systems via APIs and carries out the resolution. Here's what that looks like in practice. A customer writes, "I was charged twice for my last order." The understanding layer identifies the intent: billing dispute, likely duplicate charge. The reasoning layer pulls billing records from Stripe, cross-references the order in your OMS, checks your refund policy for duplicate charges, and determines the customer qualifies for a refund. The execution layer processes the refund through your payment system and sends the customer a confirmation with the credit timeline. No human agent touches the ticket. The entire chain runs in seconds. ## How Is an AI Customer Service Agent Different from a Chatbot? A chatbot matches keywords to scripted responses using decision trees. An AI customer service agent reasons through problems, accesses live data, and executes multi-step workflows. One follows a script. The other makes decisions. For a deeper breakdown, see [AI agent vs chatbot](/articles/ai-agent-vs-chatbot). The practical gap is stark. A chatbot receiving "cancel my subscription" sends a link to a cancellation form or routes to a human. An AI customer service agent checks the subscription status, calculates any prorated refund, processes the cancellation in your billing system, triggers a confirmation email, and optionally presents a retention offer - all within the same conversation. Chatbots deflect. Agents resolve. ## What Can an AI Customer Service Agent Handle? AI customer service agents handle structured, policy-driven issues with clear resolution paths. They work best on tickets where the decision criteria are defined and the actions involve system interactions - not subjective judgment calls. 1. **Refund processing.** The agent verifies the purchase, checks eligibility against your refund policy, processes the credit through your payment system, and confirms the timeline with the customer. 2. **Subscription management.** Upgrades, downgrades, cancellations, and plan switches - including proration calculations, billing adjustments, and confirmation emails. 3. **Order modifications.** Address changes, item swaps, quantity adjustments, and cancellations before fulfillment. The agent connects to your OMS and shipping provider to determine what's still possible. 4. **Account updates.** Contact information, payment methods, passwords, and preferences. High-volume, low-complexity - ideal for full automation. 5. **Billing inquiries.** Charge explanations, invoice requests, payment failure troubleshooting, and duplicate charge resolution. ## How Does an AI Customer Service Agent Differ from a Copilot? A copilot assists a human agent by suggesting responses and pulling relevant data. An AI customer service agent works independently - it handles the full ticket lifecycle from intake to resolution without a human in the loop. Copilots reduce handle time. Agents reduce ticket volume. The distinction matters for staffing and cost models. A copilot still requires a human agent on every ticket, so your headcount stays the same. An AI customer service agent absorbs tickets entirely, freeing human agents for escalations, edge cases, and relationship-building conversations. Many teams start with a copilot and graduate to autonomous agents as confidence grows. For more on [AI for customer support](/articles/ai-for-customer-support), see our breakdown of deployment models. ## What Does an AI Customer Service Agent Cost? Most AI agent platforms charge $0.99-2.00 per resolved ticket. Some offer monthly seat-based plans at $100-500/month. Either model typically costs less than the fully loaded expense of one human agent handling the same volume. Consider a team processing 4,000 tickets per month. At 50% AI containment, that's 2,000 AI-resolved tickets. At $1.50 per resolution, the monthly cost is $3,000 - roughly a quarter of the fully loaded cost of one US-based support agent. The economics improve as containment climbs. Cost per resolution on AI-handled tickets runs $1-3, compared to $5-12 for human-resolved tickets across most industries. ## What Results Should You Expect After Deployment? Expect measurable movement within 30-60 days. The first metric to watch is containment rate - the percentage of tickets resolved without human involvement. Most teams see steady gains as the agent handles more ticket categories. Containment rates typically reach 40-60% within the first quarter for teams deploying across 3-5 ticket categories. Average handle time for AI-resolved tickets runs 1-3 minutes versus 8-12 minutes with human agents. [McKinsey research](https://www.mckinsey.com/capabilities/operations/our-insights/beyond-the-bot-building-empathetic-customer-experiences-with-agentic-ai) indicates AI in customer care can improve satisfaction scores by 5-10% when resolution quality is maintained. The compounding effect matters: as routine tickets shift to the AI, human agents handle fewer but more complex cases, improving their response times and job satisfaction. ## Key Takeaways - AI customer service agents resolve tickets autonomously using a 3-layer architecture: understanding, reasoning, execution - Expect 40-60% containment rates within 90 days and cost per resolution under $3 - Gartner projects agentic AI will resolve 80% of common service issues by 2029 - Agents complement human teams by absorbing routine volume - not replacing judgment on complex cases ## Frequently Asked Questions ### How long does it take to deploy an AI customer service agent? Initial deployment on a single ticket category takes 2-6 weeks, including system integrations, policy configuration, and testing. Each additional ticket category adds 1-2 weeks. Most teams reach full deployment across all target categories within 2-4 months. ### Will an AI customer service agent replace my human support team? No. AI customer service agents handle routine, policy-driven tickets - freeing human agents for complex escalations, sensitive situations, and high-value conversations. Most teams redeploy rather than reduce headcount, shifting agents toward work that requires empathy and judgment. ### How do you ensure an AI customer service agent doesn't make mistakes? Policy guardrails define what the agent can and cannot do - refund limits, escalation triggers, restricted actions. Platforms like [Lorikeet](https://www.lorikeetcx.ai) include [continuous quality assurance](https://www.lorikeetcx.ai/product/coach) that audits 100% of AI interactions, catching errors before they compound. ### Can an AI customer service agent work across chat, email, and voice? Yes. Most modern AI agent platforms operate across channels, maintaining context when a customer switches from chat to email or follows up via phone. The underlying reasoning and system integrations remain the same regardless of channel. AI customer service agents are not speculative technology. They are in production today, resolving millions of tickets across e-commerce, fintech, healthcare, and SaaS. The architecture is proven: LLM-based understanding, policy-driven reasoning, API-connected execution. The question for most CX teams is no longer whether AI agents work - it's how quickly you can deploy one on your highest-volume ticket types and start measuring results. Start with one category. Prove containment. Expand from there. The teams that wait will spend the next two years paying human agents to do work that software handles in seconds. **See how [Lorikeet's AI customer service agents](https://www.lorikeetcx.ai) resolve billing, order, and subscription tickets end-to-end. Built as an AI-native agent platform for complex workflows - not FAQ deflection.** --- # AI Chatbot for E-commerce: Automating Order Tracking, Returns, and Refunds URL: https://www.lorikeetcx.ai/articles/ai-chatbot-for-ecommerce An AI chatbot for e-commerce is a support automation system that handles order inquiries, returns, refunds, and shipping questions by connecting directly to your store's backend. Industry data consistently shows that 40-50% of e-commerce support tickets are "where is my order?" (WISMO) requests - making them the single largest automation opportunity for online retailers. - WISMO tickets account for 40-50% of e-commerce support volume and are the easiest to automate - AI chatbots pull real-time tracking data from shipping APIs instead of asking customers to wait - Returns, refunds, and payment issues follow policy-driven logic that AI handles consistently - Integrations with Shopify, WooCommerce, and Stripe are table stakes for any serious deployment Most e-commerce brands hit the same wall around 500 orders per day. Support queues fill up with repetitive questions - tracking updates, return labels, refund statuses - and hiring more agents doesn't scale. AI chatbots built for e-commerce don't just deflect these tickets to a help center article. They connect to your order management system, pull the real data, apply your policies, and resolve the issue. The difference between a useful AI chatbot and a frustrating one is whether it can actually do things - not just talk about them. ## What Should You Automate First with an E-commerce AI Chatbot? Start with WISMO - "where is my order?" - tickets. They're your highest volume, lowest risk, and most repetitive ticket type. Automating WISMO alone can cut support volume by 30-40% depending on your product category and shipping complexity. The priority order after WISMO: returns and exchanges, refund processing, shipping address modifications, product questions, then payment issues. Start with the highest volume and lowest stakes, then work toward higher complexity. Each category you automate compounds the effect, freeing human agents for cases that require judgment - damaged goods disputes, custom orders, or multi-issue escalations. ## How Does an AI Chatbot Handle "Where Is My Order?" Requests? The chatbot receives the customer message, identifies the order via email or order number, queries the shipping provider's API for real-time tracking data, and returns the current status with an estimated delivery date. No human involvement required. A customer messages "where is my order?" at 11 PM on a Sunday. The AI chatbot matches their email to an order in [Shopify](https://www.shopify.com/blog/ecommerce-chatbots), pulls the tracking number, queries the carrier API, and responds: "Your order #4829 shipped via FedEx and is in transit. Estimated delivery is Tuesday, Feb 19." Total time: under 10 seconds. No agent woke up. No ticket sat in a queue until Monday morning. ## Which E-commerce Ticket Types Are Best Suited for AI? Ticket types with clear policies, structured data, and binary outcomes are ideal. In e-commerce, that covers order tracking, returns within policy windows, standard refund processing, shipping modifications before fulfillment, and product availability questions. 1. **Order tracking (WISMO).** Connects to shipping providers, returns real-time status. Resolution rate above 90% because the answer is purely data-driven. 2. **Returns and exchanges.** Checks return window eligibility, generates return labels, initiates exchanges. Works well when your return policy has clear rules. 3. **Refund processing.** Validates eligibility, processes the refund through Stripe or your payment processor, confirms the timeline to the customer. 4. **Shipping modifications.** Address changes and delivery holds before fulfillment. The AI checks fulfillment status and makes the change if the order hasn't shipped. 5. **Product questions.** Stock availability, size guides, compatibility checks. Works best when the AI accesses your product catalog, not just FAQ content. ## What Integrations Does an E-commerce AI Chatbot Need? At minimum: your e-commerce platform (Shopify, WooCommerce, BigCommerce), your payment processor (Stripe, PayPal), and your shipping providers (FedEx, UPS, USPS, DHL). Without these connections, the chatbot is just a search bar over your help center. Integration depth is the difference between [a chatbot and an AI agent](/articles/ai-agent-vs-chatbot). A chatbot that only reads data will tell a customer their return is eligible. An AI agent connected to your systems will check the return window, generate the label, email it, and update the order status - all in one conversation. [Lorikeet](https://www.lorikeetcx.ai) connects to e-commerce backends to execute full workflows - processing refunds, modifying orders, and initiating returns without routing to a human queue. ## What Results Can You Expect from AI Chatbot Automation? Expect containment rates of 35-55% within the first 90 days when deploying across 3-4 ticket categories. Cost per resolution typically drops from $5-12 with human agents to $1-3 with AI. Handle time on automated tickets falls to under 2 minutes compared to 8-12 minutes with live agents. The volume reduction is the headline number, but downstream effects matter more. When AI absorbs 40% of tickets, human agents handle fewer but more complex cases. [Gartner predicts](https://www.gartner.com/en/newsroom/press-releases/2025-03-05-gartner-predicts-agentic-ai-will-autonomously-resolve-80-percent-of-common-customer-service-issues-without-human-intervention-by-20290) agentic AI will resolve 80% of common customer service issues without human intervention by 2029. E-commerce teams that start automating now build the operational muscle to hit those numbers first. ## How Do You Avoid Common E-commerce Chatbot Failures? Most chatbot failures come from three sources: no backend integration (the bot can't access real data), no policy guardrails (the bot makes up answers), and no quality monitoring (errors go undetected). Solve these three and failure rates drop significantly. The biggest mistake is deploying a generic chatbot trained only on FAQ pages. Customers contact support because their specific order has a specific problem. Generic chatbots loop them through irrelevant articles and then offer to "connect you with an agent." Instead, deploy an [AI agent built for customer service](/articles/ai-agent-for-customer-service) that accesses order data, applies your policies, and takes action. Then use [quality assurance tooling](https://www.lorikeetcx.ai/product/coach) to audit every AI interaction. ## Key Takeaways - Start with WISMO automation - it covers 40-50% of e-commerce tickets and has the highest resolution rate - Expect 35-55% containment within 90 days, with cost per resolution dropping to $1-3 - Backend integrations with Shopify, Stripe, and shipping providers are non-negotiable for real automation - Expand one ticket category at a time: WISMO, then returns, refunds, shipping changes, product questions ## Frequently Asked Questions ### How long does it take to deploy an AI chatbot for e-commerce? Initial deployment on a single ticket category - typically WISMO - takes 2-4 weeks including integration setup and testing. Each additional category adds 1-2 weeks. A full deployment across 4-5 categories runs 2-3 months for most mid-market e-commerce brands. ### Can an AI chatbot handle returns and refunds without human approval? Yes, when configured with clear policy guardrails. You set the rules - return window length, refund limits, exception triggers - and the AI applies them consistently. High-value or edge-case refunds can route to a human for approval while standard cases process automatically. ### Does an AI chatbot work with Shopify and WooCommerce? Most e-commerce AI platforms integrate natively with Shopify and WooCommerce via API. The integration lets the AI read order data, check fulfillment status, process returns, and update customer records. Verify that the platform supports write access, not just read - otherwise it can only report information, not resolve issues. E-commerce support doesn't need to be a staffing problem. The top 5 ticket types in most online stores - order tracking, returns, refunds, shipping changes, and product questions - follow predictable patterns and clear policies. AI chatbots built for e-commerce can absorb 35-55% of that volume within the first quarter, reducing cost per resolution and freeing human agents for conversations that actually need them. The key is starting with WISMO, insisting on real backend integrations, and expanding methodically. Skip the FAQ-only bots. Deploy AI that takes action. **See how [Lorikeet](https://www.lorikeetcx.ai) automates e-commerce support workflows end-to-end - from order tracking to returns to refunds. Built for complex e-commerce operations, not just FAQ deflection.** --- # AI Call Center Agent: Voice AI for Phone Support in 2026 URL: https://www.lorikeetcx.ai/articles/ai-call-center-agent An AI call center agent is software that handles inbound and outbound phone calls using speech-to-text, large language model reasoning, and text-to-speech - all in real time. Phone still accounts for 60-70% of contact center volume, according to [ContactBabel](https://www.contactbabel.com/), with average cost per call running $5-8. - Voice AI replaces legacy IVR trees with natural-language call handling - Real-time processing requires sub-500ms latency across the full speech pipeline - Top use cases include call routing, simple inquiries, payments, and appointment scheduling - Voice is harder than chat - accents, background noise, and emotion make it a different problem The phone is not going away. Despite the growth of chat and email, most customers still pick up the phone when something matters. The problem is cost. A human agent handles 8-12 calls per hour. The math on staffing a 24/7 call center is brutal - and it gets worse with attrition rates north of 30% annually. AI call center agents target this gap by automating calls that follow predictable patterns: balance checks, appointment confirmations, order status, payment processing. Not the calls where a customer is upset about a misdiagnosis on their insurance claim. ## How Does an AI Call Center Agent Process a Phone Call? An AI call center agent runs a three-stage pipeline in real time. Speech-to-text converts the caller's voice into text. An LLM interprets intent, checks policy, and generates a response. Text-to-speech converts that response back into natural-sounding audio. The entire loop needs to complete in under 500 milliseconds to feel conversational. Anything slower and the caller notices dead air - which triggers hang-ups. In chat, a 2-second response feels fast. On a phone call, 2 seconds of silence feels broken. Modern platforms use streaming architectures where speech-to-text begins processing before the caller finishes speaking. ## What Can Voice AI Actually Handle Today? Voice AI handles structured, short-turn interactions reliably. These are calls with predictable inputs, clear resolution paths, and limited back-and-forth. Open-ended or emotionally charged calls remain better suited for human agents. 1. **IVR replacement.** Instead of "press 1 for billing," callers state their issue in plain language. The AI routes them or resolves the issue without a transfer. 2. **Simple inquiries.** Account balances, order tracking, store hours, policy lookups. The AI pulls data from backend systems and reads it back in seconds. 3. **Appointment scheduling.** Booking, rescheduling, and cancellation workflows where the AI checks availability and confirms in real time. 4. **Payment processing.** Bill payments and payment arrangement setups with PCI-compliant voice capture for card details. 5. **Outbound notifications.** Appointment reminders, delivery confirmations, and payment-due alerts that free agents from manual dial-outs. ## Why Is Voice AI Harder Than Chat AI? Voice AI deals with signal problems that text channels never face. Accents, dialects, background noise, crosstalk, and variable audio quality all degrade speech-to-text accuracy. Chat AI receives clean text input. Voice AI has to earn its input. A model trained primarily on American English struggles with Indian, Australian, or Nigerian English - and errors compound downstream. If the AI mishears "refund" as "building," the interaction derails. Emotional detection also matters - a caller's tone carries urgency that text does not. The best voice AI systems adjust pacing based on detected sentiment, but this remains early-stage. ## How Do Customers Feel About Talking to Voice AI? Customer perception is mixed but improving. Most callers tolerate AI for simple tasks - checking a balance, confirming an appointment. Tolerance drops sharply when the issue is complex or emotional. The biggest driver of negative perception is the failure mode. When voice AI misunderstands a request and loops the caller three times, trust evaporates. [Forrester research](https://www.forrester.com/research/customer-experience/) shows that effort - not channel - determines satisfaction. If the AI resolves quickly, most customers do not care whether they spoke to a person. Deploy voice AI where it can resolve confidently and route everything else to a human fast. ## When Does Voice AI Make Sense vs. Keeping Humans on Phones? Voice AI makes sense when call patterns are high-volume, low-complexity, and policy-driven. Keep humans on calls involving disputes, complaints, retention, or any scenario where empathy and judgment drive the outcome. A useful rule: if the call follows a decision tree with fewer than 5 branches, voice AI can handle it. Most contact centers find that 30-40% of inbound call volume fits the AI-eligible profile. That is enough to reduce staffing pressure and hold times without forcing callers into AI interactions they resent. ## What Performance Metrics Matter for AI Call Center Agents? Containment rate, speech recognition accuracy, average handle time, and caller hang-up rate are the four to track from day one. Containment rate - calls resolved without human transfer - should reach 25-40% in the first 90 days, scaling to 40-55% as the system improves. Speech recognition accuracy needs to hold above 90% across your caller demographics. Handle time for AI-resolved calls runs 1-3 minutes versus 6-10 minutes with human agents. Cost per call drops from $5-8 (human) to $1-2 (AI-resolved). If callers abandon the AI mid-call above 15%, the experience needs work. ## Key Takeaways - Voice AI requires sub-500ms latency across the speech-to-text, LLM, and text-to-speech pipeline - 30-40% of inbound call volume is typically AI-eligible, at $1-2 per resolved call vs. $5-8 with humans - Speech recognition accuracy above 90% is the baseline - below that, caller frustration spikes - Multi-channel AI with shared context outperforms voice-only tools by covering chat, email, and phone ## Frequently Asked Questions ### Can AI call center agents handle calls in multiple languages? Yes, but quality varies. English, Spanish, and French have strong speech-to-text accuracy. Less common languages or heavy regional dialects see lower accuracy. Test with real caller audio from your population before committing to multilingual deployment. ### How long does it take to deploy a voice AI agent? Expect 4-8 weeks for the initial call type, including telephony integration, speech model tuning, and policy configuration. Each additional call type adds 1-2 weeks. Pilot with your highest-volume, simplest call category first. ### What happens when the voice AI cannot resolve a call? The AI transfers to a human agent with full call context - what the caller said, what the AI attempted, and why resolution failed. Good platforms avoid cold transfers entirely. The human picks up with complete history so the caller never repeats themselves. AI call center agents are not replacing phone support. They are absorbing the predictable, repetitive portion - the calls that follow a script anyway. For contact centers running legacy IVR and staffing for peak volume, voice AI offers lower costs and shorter hold times without degrading the experience on calls that need a person. The catch: voice AI is not a standalone solution. Customers move between channels. The best results come from deploying [AI agents](/articles/ai-agent-for-customer-service) across chat, email, and phone with shared context - not bolting a voice bot onto a fragmented stack. [Understanding when AI handles the call vs. when a human should](/articles/ai-vs-human-customer-service) is the real deployment decision. **See how [Lorikeet's AI agents](https://www.lorikeetcx.ai) handle customer interactions across chat, email, and voice from a single platform. Built for real resolution with [continuous quality assurance](https://www.lorikeetcx.ai/product/coach) - not IVR trees with a language model on top.** --- # AI Agent vs Chatbot: What Actually Resolves Customer Issues? URL: https://www.lorikeetcx.ai/articles/ai-agent-vs-chatbot An AI agent is an autonomous system that understands context, makes decisions, and takes actions to resolve customer issues end-to-end. A chatbot follows predefined scripts and decision trees. The distinction matters - [Gartner predicts](https://www.gartner.com/en/newsroom/press-releases/2025-03-05-gartner-predicts-agentic-ai-will-autonomously-resolve-80-percent-of-common-customer-service-issues-without-human-intervention-by-20290) that by 2029, agentic AI will autonomously resolve 80% of common customer service issues, while traditional chatbots typically deflect rather than resolve. - AI agents take actions (process refunds, update accounts) while chatbots route to human agents - AI agents resolve complex issues autonomously; chatbots handle routine FAQ lookups but escalate the rest - AI agents improve over time through learning; chatbots need manual script updates - The right choice depends on ticket complexity, not company size Every CX leader has heard the pitch: "Our chatbot will handle your tickets." Then reality hits. The bot deflects to a human 70% of the time, customers get frustrated repeating themselves, and your team spends more time babysitting the bot than it saves. The problem isn't automation itself. It's the difference between a tool that follows a script and one that actually thinks. That gap is the AI agent vs chatbot divide. ## What Is the Difference Between an AI Agent and a Chatbot? A chatbot matches user input to predefined responses using keyword triggers or decision trees. An AI agent uses large language models to understand intent, access backend systems, and execute multi-step workflows autonomously. One reads from a script. The other reasons through a problem. The technical difference comes down to architecture. Chatbots operate on if-then logic - if a customer says "refund," route them to the refund FAQ. AI agents operate on reasoning chains - they read the order history, check the refund policy, determine eligibility, process the refund, and confirm with the customer. No human in the loop. A [Forrester TEI study](https://tei.forrester.com/go/zendesk/advancedaisupport/?lang=en-us) found one e-commerce company cut average handle time by 42% after deploying AI-powered customer service tools. ## How Do AI Agents Handle Complex Customer Issues? AI agents connect to your backend systems - order management, billing, CRM - and execute actions across them in a single conversation. They handle multi-step workflows like processing a return, issuing a replacement, and adjusting loyalty points without escalating to a human agent. ### Multi-System Orchestration Where chatbots hit a wall at "let me transfer you to a specialist," AI agents pull data from 3-5 systems simultaneously. A customer asking about a delayed order gets real-time tracking data, shipping carrier status, and proactive compensation - all in one response. ### Context Retention Across Channels AI agents maintain conversation context across chat, email, and voice. A customer who starts on chat and follows up via email doesn't repeat themselves. Chatbots typically lose context the moment a session ends. ## When Should You Use a Chatbot Instead of an AI Agent? Chatbots still make sense for simple, high-volume interactions where the answer never changes - store hours, shipping policies, password resets with a single step. If your queries are repetitive and the answer fits in one sentence, a chatbot is cheaper and faster to deploy. The decision framework is straightforward. Map your ticket types by complexity. If 70%+ of your volume is simple FAQ lookups, a chatbot handles it. If your tickets require accessing customer data, making decisions based on policy, or taking actions in backend systems, you need an [AI agent for customer service](/articles/ai-agent-for-customer-service). Most mature CX operations find the split is closer to 40% simple / 60% complex - which means chatbots alone leave the majority of tickets untouched. ## What Results Can You Expect from Each Approach? The performance gap between chatbots and AI agents is measurable across every core CX metric. Companies that switch from chatbot-only to AI agent deployments see improvements within the first 90 days of going live. Average handle time drops from 8-12 minutes to 3-5 minutes per resolved interaction. [McKinsey research](https://www.mckinsey.com/capabilities/operations/our-insights/beyond-the-bot-building-empathetic-customer-experiences-with-agentic-ai) indicates generative AI can improve customer satisfaction by 5-10% in general customer care operations. First-contact resolution improves measurably - one banking implementation improved FCR from 50% to 70% with AI, per [industry benchmark data](https://www.fullview.io/blog/first-call-resolution-rate-industry-standards). These numbers compound. Higher containment means fewer tickets hitting your human team, which means lower staffing costs and faster response times for the complex issues that genuinely need a person. ## What Should You Look for When Choosing Between Them? Evaluate based on 3 factors: ticket complexity, system integrations needed, and resolution expectations. If you need the tool to take actions - not just answer questions - you need an agent, not a bot. 1. **Resolution vs. deflection.** Ask vendors for containment rates, not engagement rates. A chatbot that "handles" 90% of conversations but resolves 20% is just an expensive FAQ page. Demand resolution metrics. 2. **Backend integration depth.** Check whether the tool can read and write to your systems. AI agents like [Lorikeet](https://www.lorikeetcx.ai) connect to order management, billing, and CRM to actually execute workflows - not just surface information. 3. **Guardrails and accuracy.** AI agents need policy enforcement to prevent incorrect actions. Look for platforms with built-in [quality assurance](https://www.lorikeetcx.ai/product/coach) that audits 100% of AI interactions, not just random samples. ## Key Takeaways - Gartner projects AI agents will resolve 80% of common issues autonomously by 2029 - Switch to AI agents when tickets require multi-step actions across backend systems - Expect 30-40% handle time reduction and 5-10% CSAT improvement with AI agents - Evaluate on resolution rate, not engagement rate - deflection is not resolution ## Frequently Asked Questions ### How much do AI agents cost compared to chatbots? Chatbots range from $50-500/month for basic plans. AI agent platforms typically cost $0.50-3.00 per resolved conversation, which often works out cheaper at scale because they actually resolve issues instead of escalating them to your paid human team. ### How long does it take to deploy an AI agent vs a chatbot? Basic chatbots deploy in days with template scripts. AI agents take 2-6 weeks for full deployment including system integrations, policy configuration, and testing. The longer setup pays back through higher resolution rates from day one. ### Can AI agents completely replace chatbots? Yes, in most cases. AI agents can handle everything a chatbot does plus complex multi-step tasks. The main reason to keep a chatbot is if your volume is 90%+ simple FAQ queries and you don't need backend system actions. The chatbot era solved one problem - making support available 24/7. But availability without resolution is just a faster way to frustrate customers. AI agents close that gap by combining always-on availability with the ability to actually fix problems, access systems, and take action. For CX teams handling complex tickets across billing, orders, and account management, the choice is clear. The question isn't whether to move beyond chatbots - it's how quickly your team can make the switch. **See how AI agents handle real customer issues end-to-end. Explore [Lorikeet's autonomous AI agent](https://www.lorikeetcx.ai) built for complex CX workflows.** --- ## How to Improve CSAT Score: 8 Strategies That Work in 2026 URL: https://www.lorikeetcx.ai/articles/how-to-improve-csat-score Updated: 2026-02-26 Keywords: how to improve CSAT score, CSAT improvement, customer satisfaction score Improving your CSAT score means fixing the specific interactions that are failing, not surveying customers more frequently. Teams that score above 80% consistently combine 3 things: fast first response (under 4 hours), high first-contact resolution, and systematic quality review. In 2026, Lorikeet and other AI-native support platforms make it possible to address all 3 simultaneously without scaling headcount. - First-contact resolution is the top CSAT driver - every repeat contact drops satisfaction by 15-25 points, per SQM Group research. - 46% of customers expect a reply within 4 hours, making response speed the most-cited driver of low CSAT in B2C support. - AI that resolves issues (not deflects) consistently outperforms traditional chatbots on CSAT - routing to unhelpful FAQs actively lowers your score. - Teams reviewing 100% of tickets via AI-assisted QA consistently catch coaching opportunities that 2-5% manual sampling misses entirely - issues only visible at scale. A falling CSAT score is rarely a mystery. Customers tell you exactly what went wrong through repeat contacts, escalations, and comment fields. The challenge for most support teams is connecting those signals to the specific workflows causing them. You know the score dropped. You don't know which ticket type, which agent, or which process step is responsible. This guide breaks down the strategies that consistently move CSAT from the 70-75% range into 80-85%, with specific actions behind each one. ## What Drives CSAT Scores Up or Down? CSAT rises when customers feel heard, get accurate answers quickly, and don't need to repeat themselves. It falls when resolution requires multiple contacts, agents lack context, or responses are slow. The 3 strongest predictors are first-contact resolution rate, response speed, and answer accuracy. SQM Group research shows that every additional contact required to resolve an issue drops CSAT by 15-25 percentage points. Customers who resolve at first contact rate their experience 20-30% higher than those needing a follow-up. This means deflecting tickets to a bot that can't resolve them doesn't just frustrate customers - it actively damages your CSAT even when it technically reduces ticket volume. ## How Does AI Improve CSAT Scores? AI improves CSAT when it resolves issues - querying order data, processing refunds, answering policy questions accurately. It hurts CSAT when it routes customers to knowledge base articles that don't address their problem. Resolution versus deflection is the most important factor in whether AI helps or harms your score. ### Resolution-focused agents vs. script-based chatbots AI agents built for resolution take actions on backend systems - checking account status, issuing credits, updating subscriptions - rather than surfacing FAQs and hoping customers self-serve. Teams deploying resolution-focused agents see first-contact resolution rates above 50-60%, which correlates consistently with CSAT above 80%. Traditional chatbots relying on deflection rarely move CSAT because they don't fix the actual reason customers contacted support. ### Context preservation across interactions One of the most common CSAT complaints is "having to repeat myself." AI agents that maintain full context across sessions and channels eliminate this friction. When a customer moves from email to live chat, the agent already knows the history. No re-explaining required - this is one of the fastest and least expensive CSAT improvements available to most teams. ## What Are the Highest-Impact CSAT Improvement Strategies? The strategies that move CSAT most are those that attack root causes - slow responses, low resolution quality, and inconsistent agent performance. Start with whichever maps most closely to your current drivers of low scores. - Track first-contact resolution separately from CSAT. FCR and CSAT are strongly correlated, but tracking them separately reveals root causes faster. When FCR drops and CSAT follows, the problem is in the resolution process - not in how friendly agents are. - Review 100% of tickets, not 2-5%. Manual QA samples a tiny fraction of interactions. AI-powered QA tools review every ticket, identifying patterns random sampling would never catch - like one ticket type with consistently poor resolution, or an agent who handles refusals differently from the rest of the team. - Set and publish response time SLAs. Customers tolerate wait times better when they know what to expect. Teams that commit to specific response windows (e.g., "4-hour email response") and meet them report CSAT improvements even when handle time stays constant. - Send CSAT surveys within 10 minutes of resolution. Surveys sent immediately post-resolution capture the actual interaction experience. Surveys sent hours later get lower response rates and noisier data that's harder to act on. - Route complex issues to specialists on first contact. Misrouting is a silent CSAT killer. When a billing dispute or technical issue lands in a general queue, handle time increases and resolution quality drops. Routing logic that identifies issue complexity upfront cuts misrouting and improves FCR in parallel. ## What Results Can You Expect From CSAT Improvement Programs? CSAT improvements compound over time. Early changes to routing and response time show results within weeks; systemic changes to resolution quality take longer but have larger effects. Teams that reduce email response time from 12+ hours to under 4 hours typically see CSAT improve by 5-8 points within 30-60 days. Moving from 2-5% manual QA sampling to 100% AI-assisted review reduces policy inconsistency complaints and cuts time-to-identify-coaching-opportunity by roughly 75%. Improving first-contact resolution from 55% to 70% - achievable through better routing and resolution-focused AI - is typically the biggest single lever, worth 10-15 CSAT points over a quarter for most teams. Programs combining all 3 systematically tend to move scores from the low 70s into the 80-85% range within 90 days. The gains also reduce operating cost. Higher FCR means fewer repeat contacts, which lowers cost per ticket while pushing CSAT up - a combination few other CX investments deliver simultaneously. ## Lorikeet's Take on CSAT Improvement At Lorikeet, we've seen teams move from CSAT in the low 70s to consistently above 82% - not by adding agents or running more surveys, but by raising resolution quality. Most CX vendors frame CSAT improvement as a training problem or a culture problem. In practice, it's almost always a resolution problem. Customers rate their experience based on whether their issue was fixed, not on how politely the interaction went. Lorikeet builds AI agents that resolve issues end-to-end, paired with continuous QA that catches quality drift before it compounds into score drops. If you're serious about moving your CSAT number, see how Lorikeet approaches quality review and resolution tracking. ## Key Takeaways - First-contact resolution is the #1 CSAT driver - every repeat contact drops scores 15-25 points, per SQM Group data. - AI agents that resolve (not deflect) consistently correlate with CSAT above 80% when FCR rates exceed 50-60%. - 100% AI-assisted QA review surfaces coaching opportunities that random 2-5% manual sampling structurally misses - scale is what changes the outcome. - Programs combining faster response, higher FCR, and systematic QA can move scores from the low 70s to 80-85% within 90 days. ## Frequently Asked Questions ### How quickly can you improve CSAT scores after making changes? Quick operational changes - improved routing or response time SLAs - show CSAT improvements within 2-4 weeks. Deeper changes to resolution quality and QA processes typically show measurable impact within 60-90 days. Tracking CSAT at the ticket-category --- ## How to Reduce Customer Service Costs Without Cutting Quality URL: https://www.lorikeetcx.ai/articles/how-to-reduce-customer-service-costs Updated: 2026-02-26 Keywords: how to reduce customer service costs, support cost reduction, cost per ticket Reducing customer service costs starts with understanding what's actually driving them. The biggest cost in most support operations isn't headcount - it's repeat contacts. Every issue that requires 2 or more contacts to resolve doubles the cost of that interaction. In 2026, the most effective cost reduction strategies focus on improving resolution quality, not just reducing ticket volume. Lorikeet data shows that teams improving first-contact resolution from 55% to 70% typically cut support operating costs by 20-30% without reducing headcount. - Repeat contacts are the largest cost multiplier in support - a ticket requiring 3 contacts costs 3x more to serve than one resolved at first contact. - Industry benchmarks show cost per ticket ranges from $2.70 for retail e-commerce to $30-$60 for B2B support, per LiveChatAI industry data. - AI agents that resolve tickets end-to-end reduce cost per resolution by 40-60% compared to agent-assisted interactions for eligible ticket types. - Self-service done well costs $0.50-$2.37 per resolution - but only when it genuinely resolves, not when it deflects to content that doesn't help. Most cost reduction initiatives make the same mistake: they focus on keeping contacts out of the queue rather than resolving the ones that arrive efficiently. Deflecting a customer to an FAQ that doesn't answer their question doesn't reduce cost - it generates a repeat contact, an escalation, or an unhappy customer who churns. The real lever is resolution rate, not contact volume. This guide covers the cost drivers that actually matter, the benchmarks worth tracking, and the strategies that reduce cost per ticket without sacrificing CSAT. ## What Are the Main Drivers of Customer Service Costs? Customer service costs are driven by 4 main factors: contact volume, handle time per contact, cost per agent hour, and repeat contact rate. Of these, repeat contact rate has the largest leverage because it multiplies across all other cost drivers - a contact that generates a repeat doubles volume and handle time simultaneously. Most cost discussions focus on headcount and handle time. But first-contact resolution rate is the stronger predictor of total support unit costs than agent efficiency metrics in isolation. A team with 70% FCR at moderate handle time will almost always cost less to operate than a team with 90% efficiency but 50% FCR - because the latter is handling far more contacts per customer issue, multiplying every cost it incurs. ## What Does Customer Service Cost Per Ticket Actually Look Like? Cost per ticket varies significantly by industry, channel, and ticket complexity. Understanding your benchmark is the first step to knowing where you have room to improve. ### Industry benchmarks According to LiveChatAI's 2025 industry analysis, typical cost per ticket ranges are: retail and e-commerce at $2.70-$5.60, SaaS support at $18-$35, B2B support at $30-$60, and telecom and utilities at $20-$30. The global baseline sits around $6-$7 per contact. Self-service resolution costs $0.50-$2.37 per issue when the self-service content actually resolves the problem - not just deflects the contact. ### The hidden cost: repeat contacts Most cost benchmarks measure cost per contact, not cost per issue. These are different numbers. If an issue requires 2.3 contacts on average to resolve - a common figure for complex product issues - your real cost per issue is 2.3x your cost per contact. Tracking contact-per-issue rate alongside cost per ticket reveals the true operating cost, and usually identifies the most expensive ticket categories to fix. ## How Do You Reduce Customer Service Costs With AI? AI reduces support costs when it takes ownership of complete resolution, not when it routes customers to knowledge base articles. The distinction matters: a chatbot that deflects tickets generates the same repeat contact pattern as a poorly trained agent. An AI agent that resolves tickets - checking account data, issuing refunds, updating records - removes those contacts from the queue entirely. - Deploy AI on high-volume, well-defined ticket types. Start with the 20% of ticket types that generate 60-70% of volume - password resets, order status, refund requests, subscription changes. These are fully automatable at high resolution quality, with cost-per-resolution dropping to near self-service levels when AI handles them end-to-end. - Eliminate misrouting to reduce handle time. Intelligent triage that routes tickets to the right agent or AI workflow on first contact cuts average handle time significantly. Every misrouted ticket requires re-queuing, agent context-switching, and often a follow-up contact from the customer. - Build a quality assurance loop. Cost creep in support often comes from knowledge gaps - agents giving inconsistent answers that generate clarification contacts. AI-assisted QA reviewing 100% of interactions catches knowledge inconsistencies before they compound into repeat contact patterns. - Measure cost per issue, not just cost per contact. Until you're tracking how many contacts the average issue requires, you're optimising for the wrong metric. Teams that switch to cost-per-issue tracking identify their highest-cost ticket categories within weeks and can prioritise automation and training investment accordingly. ## What Cost Reduction Results Should You Expect? Cost reduction in customer service compounds when multiple levers are pulled together. Individual improvements are meaningful; combined programs are significantly more powerful. Teams that deploy AI agents on high-volume, eligible ticket types typically reduce cost per resolution for those categories by 40-60% within the first quarter. Improving first-contact resolution from 55% to 70% - by combining better routing with resolution-focused AI - reduces repeat contact volume by 25-30%, which flows through to total cost reduction directly. Teams that implement both AI resolution and 100% QA review together report operating cost reductions of 20-30% within 6 months, while maintaining or improving CSAT scores. The combined effect compounds: lower repeat contacts mean lower volume, which improves agent availability, which improves response time, which improves CSAT further. The cost and quality improvements are not in tension. Resolution-focused AI raises CSAT and lowers cost simultaneously - because the thing customers most want (their issue resolved) is also the thing that eliminates repeat contacts and handle time. ## Lorikeet's Take on Reducing Support Costs At Lorikeet, we see the same mistake repeatedly: teams optimise for deflection and wonder why costs don't drop. Deflecting a customer to a knowledge base article that doesn't solve their problem doesn't save money - it creates a repeat contact, an escalation, and a CSAT score that tanks. Lorikeet's approach to cost reduction is resolution-first: AI agents that own the full interaction, not the first 2 exchanges before passing to a human. The teams we work with that achieve the largest cost reductions are the ones who measure contact-per-issue rather than just cost-per-contact - because that's where the actual cost is hiding. See how Lorikeet approaches resolution-driven cost reduction. ## Key Takeaways - Repeat contacts are the largest cost driver in support - a 2.3 contact-per-issue rate means real cost per issue is 2.3x your cost-per-contact benchmark. - AI agents that resolve end-to-end reduce cost per resolution by 40-60% for eligible ticket types vs. agent-assisted handling. - Improving first-contact resolution from 55% to 70% typically cuts repeat contact volume by 25-30%, flowing directly to cost reduction. - Cost per ticket ranges from $2.70 (retail e-commerce) to $30-$60 (B2B support) - industry benchmarks help identify where automation ROI is highest. ## Frequently Asked Questions ### How much does AI reduce --- ## What Are AI Guardrails for Customer Service? A Practical Guide URL: https://www.lorikeetcx.ai/articles/what-are-ai-guardrails-for-customer-service Updated: 2026-02-26 Keywords: AI guardrails customer service, AI safety rules, AI guardrails definition AI guardrails for customer service are the rules, constraints, and safety controls that define what an AI agent can and cannot do in a customer interaction. They determine when AI escalates to a human, which topics it can address, what commitments it can make, and how it handles sensitive data. In 2026, configuring effective guardrails is the difference between an AI agent that builds trust and one that creates liability. Lorikeet builds guardrail configuration into every deployment to ensure AI agents operate safely within their intended scope. - Guardrails operate at 3 levels: input filtering (what the AI receives), output control (what it says), and action limits (what it can do in backend systems). - Without output guardrails, AI agents can make unauthorised commitments - like promising refunds or service levels that violate policy - creating real liability. - Escalation guardrails are the most business-critical: poorly defined escalation triggers are the leading cause of AI-related complaints in regulated industries. - Well-configured guardrails don't reduce resolution capability - they improve trust and allow AI to operate in higher-stakes interaction types safely. Most discussions about AI guardrails focus on the technical implementation - safety classifiers, content filters, NeMo frameworks. That's useful for engineering teams. But for support leaders deploying AI agents, the relevant questions are different: what should my AI be allowed to do, what happens when it's uncertain, and how do I know when guardrails are being triggered? This guide answers those questions from a customer service operations perspective. ## What Are AI Guardrails and Why Do They Matter? AI guardrails are controls that keep an AI agent operating within its intended scope. They prevent the agent from producing inaccurate outputs, making unauthorised commitments, mishandling sensitive information, or taking actions it isn't authorised to take. For customer service, guardrails are what make it possible to trust AI with real customer interactions. McKinsey's overview of AI guardrails describes them as "the safeguards that ensure AI systems operate safely, responsibly, and within defined boundaries." In customer service specifically, these boundaries matter because AI agents interact with real customers, handle sensitive account data, and make operational decisions that affect business outcomes. An AI without guardrails isn't just a safety risk - it's an operational liability. ## What Are the Main Types of AI Guardrails? Customer service AI guardrails fall into 4 categories, each addressing a different failure mode. Effective deployments configure all 4 - most problems with AI agents trace back to gaps in one of these areas. ### Input guardrails Input guardrails filter and classify what comes into the AI before it processes a request. They detect sensitive content (PII, financial data, health information), identify the topic and intent of an inquiry, and flag inputs that fall outside the AI's permitted scope. Input filtering prevents the AI from processing requests it shouldn't handle and routes them appropriately before a problematic response is generated. ### Output guardrails Output guardrails control what the AI says. They prevent the agent from making unauthorised commitments (offering refunds beyond policy limits, guaranteeing service levels), generating factually incorrect responses, or producing content that could cause harm. IBM's AI guardrail documentation describes output controls as the primary mechanism for maintaining "professional tone, preventing policy breaches, and avoiding misinformation." For customer service, output guardrails are the last line of defence before a response reaches a customer. ### Action guardrails Action guardrails define what the AI can do in connected backend systems. Can it issue a refund? Up to what amount? Can it update account settings without secondary confirmation? Action limits are the most critical guardrails for AI agents that have system access - and the most commonly under-configured. Most deployments start with conservative action guardrails and expand them as confidence in resolution accuracy grows. ### Escalation guardrails Escalation guardrails define when the AI hands off to a human agent. Triggers typically include: customer sentiment below a threshold, topic type (legal, medical, regulatory), unresolved intent after N turns, or explicit customer request for a human. Well-defined escalation guardrails are the most business-critical configuration decision - because poor escalation logic is what generates the most serious AI-related complaints in regulated industries. ## How Should You Configure AI Guardrails for Customer Service? Guardrail configuration is not a one-time setup - it's an iterative process that tightens as you learn how the AI behaves in production. Start conservative and expand; don't start permissive and try to restrict after problems surface. - Map your escalation triggers first. Before configuring anything else, define explicitly what situations require human involvement: transaction size thresholds, regulated topics, frustrated customer signals, and requests the AI can't reliably resolve. Escalation guardrails protect both customers and the business, and they need to be grounded in real interaction data, not assumptions. - Define action limits by category and confidence threshold. Not all actions carry the same risk. Password resets and address updates can operate with minimal guardrails. Refunds, cancellations, and account closures should have amount limits, confirmation steps, and audit logging. Configure each action category separately rather than using a single global permission level. - Build topic scope controls. Define which topics the AI is authorised to handle. For most support deployments, the permitted scope starts with your top 20-30 ticket types by volume and expands as performance data validates capability. Topics outside scope should route to humans cleanly, not generate speculative AI responses. - Monitor guardrail trigger rates, not just resolution rates. If escalation guardrails are triggering 40% of the time, the AI's scope is too broad or the guardrails are misconfigured. If they're triggering 0.5% of the time in a high-complexity environment, guardrails may be too narrow. Trigger rate data is the primary signal for guardrail calibration. ## What Happens Without Proper AI Guardrails? Without properly configured guardrails, AI agents create problems that are expensive to reverse. Unauthorised commitments - an AI promising a refund beyond policy, or confirming a service level that doesn't exist - become contractual obligations once delivered to a customer. Topic drift - an AI engaging with legal, medical, or regulatory questions it has no business addressing - creates liability. Hallucinated information - factually wrong responses delivered confidently - damages trust in ways that are hard to recover from. Most AI customer service failures are guardrail failures. The underlying model performs well in its intended scope; problems arise when inputs, outputs, or actions fall outside that scope without appropriate controls. ## Lorikeet's Take on AI Guardrails At Lorikeet, we treat guardrail configuration as core to deployment, not an afterthought. Most vendors treat safety controls as a technical checkbox. In practice, guardrail design is a business strategy decision - it determines where AI can operate independently, where it escalates, and what level of trust your team places in automated decisions. Lorikeet's guardrail framework is built around action confidence thresholds: the AI takes actions it can resolve reliably, escalates when it can't, and never speculates outside its configured scope. The goal isn't maximum restriction - it's maximum resolution within a trust boundary that expands over time. See how Lorikeet approaches AI safety in production customer service deployments. ## Key Takeaways - AI guardrails operate at 4 levels: input filtering, output control, action limits, and escalation triggers - all 4 must --- ## Cost Per Support Ticket: Benchmarks and How to Reduce It URL: https://www.lorikeetcx.ai/articles/customer-service-cost-per-ticket Updated: 2026-02-26 Keywords: cost per support ticket, customer service cost per ticket, support cost benchmarks Customer service cost per ticket ranges from $2.70 for simple retail interactions to $60 or more for complex B2B support cases. The wide range reflects differences in handle time, channel, issue complexity, and agent cost. Understanding your cost per ticket - and what's driving it - is the starting point for any serious support cost reduction effort. In 2026, Lorikeet and other AI-native platforms have made it possible to reduce cost per ticket by 40-60% for eligible ticket types without degrading resolution quality. - Industry benchmarks: retail e-commerce $2.70-$5.60 per ticket, SaaS support $18-$35, B2B support $30-$60, per LiveChatAI's 2025 industry analysis. - Self-service resolution costs $0.50-$2.37 per issue - but only when content genuinely resolves, not when it deflects without solving. - The biggest hidden cost multiplier is repeat contacts: a 2.3 contact-per-issue rate means your real cost per issue is 2.3x your cost-per-contact benchmark. - First-contact resolution rate is a stronger predictor of total support cost than handle time - because FCR determines how many contacts each issue generates. Cost per ticket is one of the most widely tracked metrics in customer support - and one of the most commonly misread. The number most teams see is cost per contact. The number that matters for budget planning is cost per issue resolved, which accounts for how many contacts each problem requires. Until you track both, you're optimising for the wrong unit. This guide covers what drives cost per ticket, how your numbers compare to industry benchmarks, and the levers with the highest impact on reduction. ## What Is Cost Per Ticket in Customer Service? Cost per ticket is the total cost of handling one customer support interaction, including agent time, overhead, tooling, and management. It's calculated by dividing total support operating costs by total ticket volume over a period. For most teams, it's the primary unit cost metric - but it only tells part of the story when analysed in isolation from resolution quality. The calculation includes direct costs (agent salaries, benefits, tooling licences) and indirect costs (management overhead, training, quality review). BMC's analysis of service desk cost structures shows that agent labour typically represents 70-80% of total cost per ticket, with tooling and overhead making up the remainder. This means cost reduction strategies that don't address labour efficiency - by either automating eligible ticket types or improving resolution speed - have limited impact on the headline number. ## What Are the Industry Benchmarks for Cost Per Ticket? Cost per ticket varies substantially by industry, ticket complexity, and channel. Knowing your benchmark category helps you understand whether your costs are typical or whether there's significant room to improve. ### By industry According to LiveChatAI's 2025 cross-industry analysis, benchmarks by sector are: retail and e-commerce at $2.70-$5.60 per ticket, SaaS and software support at $18-$35, high-tech product support at $28-$35, B2B enterprise support at $30-$60, and telecom and utilities at $20-$30. The global baseline across all industries sits around $6-$7 per contact. B2B costs are high because of longer handle times, more complex issues, and higher agent specialisation requirements. ### By channel Channel has a major impact on cost per ticket. Phone support is consistently the most expensive channel, with costs 3-5x higher than email or chat for the same issue type, because handle time is longer and agents can only manage 1 call at a time. Live chat agents can handle 3-5 concurrent sessions, which reduces effective cost per ticket significantly. AI-handled tickets cost $0.50-$2.37 per resolution when AI takes full ownership - comparable to self-service, but with resolution quality more like an experienced agent. ## What Drives Cost Per Ticket Higher Than Benchmarks? When cost per ticket exceeds benchmark for your industry, the causes are almost always found in one of 4 areas: high repeat contact rates, misrouting, channel mix, or resolution quality issues that generate escalations. - Repeat contacts. If 30% of your issues require 2 or more contacts to resolve, your effective cost per issue is significantly higher than your cost per contact. A ticket averaging 2.3 contacts costs 2.3x your stated cost-per-ticket figure. Tracking contact-per-issue rate reveals this multiplier - and targeting it is the highest-leverage cost reduction available to most teams. - Channel misrouting. When a simple password reset lands in a queue with a 2-hour SLA behind more complex cases, you're paying for queue position and agent overhead that aren't necessary. Intelligent routing that matches ticket type to the right resolution path - AI, self-service, or specialist agent - eliminates this overhead. - Insufficient AI resolution rate. AI that deflects rather than resolves doesn't reduce cost - it moves contacts around. AI that takes actions end-to-end reduces cost per resolution for eligible ticket types by 40-60% by eliminating agent involvement entirely for those interactions. - Escalation rate from AI to agents. If AI is escalating 50%+ of tickets to human agents, the combined AI + agent cost per issue often exceeds pure agent handling. Monitor escalation rates as a cost efficiency signal - high escalation usually means AI scope needs tightening, not that AI can't help. ## What Results Can You Expect From Cost Per Ticket Reduction? The size of the improvement depends on where you're starting from and which levers you pull. Teams closest to benchmark have smaller room to move; teams significantly above benchmark often find multiple overlapping causes that, fixed together, produce compounding savings. Automating the top 20% of ticket types by volume (typically the simplest, most repeatable issues) reduces cost for those categories from $18-$35 per ticket to $0.50-$2.37 per resolution - a reduction of 85-95% for that segment. Improving FCR from 55% to 70% eliminates 25-30% of repeat contacts, reducing total contact volume without any change to ticket eligibility for automation. Teams implementing AI-assisted QA to review 100% of tickets identify knowledge inconsistencies that generate repeat contacts - typically finding 5-10% of ticket categories where agent answers are inconsistent, generating unnecessary follow-up contacts. Combined, these 3 levers typically reduce total support operating cost by 20-35% within 6-12 months. The key insight: cost reduction through resolution improvement is durable. Cost reduction through deflection erodes as customers find alternative contact paths or churn. ## Lorikeet's Take on Cost Per Ticket At Lorikeet, we see cost-per-ticket discussions miss the repeat contact problem constantly. Teams celebrate a 20% reduction in cost-per-contact while contact-per-issue quietly climbs, leaving total support cost flat or higher. Lorikeet's approach focuses on cost-per-issue-resolved as the primary unit, because that's the metric that actually reflects operational efficiency. The teams we work with that achieve the largest cost reductions do so by raising FCR and resolution quality - not by deflecting contacts or cutting agent capacity. See how Lorikeet approaches resolution-first cost reduction in practice. ## Key Takeaways - Cost per ticket ranges from $2.70 (retail) to $60 (B2B support) - benchmark first to know where you stand relative to your industry. - Repeat contact rate is the largest hidden cost multiplier - a 2.3 contact-per-issue average means real cost per issue is 2.3x your cost-per-contact figure. - AI handling eligible ticket types end-to-end reduces cost per resolution by 85-95% for those categories vs. agent-assisted handling. - Improving FCR from 55% to 70% eliminates 25-30% of repeat contact volume --- ## How to Offer 24/7 Customer Support: 5 Practical Approaches URL: https://www.lorikeetcx.ai/articles/how-to-offer-24-7-customer-support Updated: 2026-02-26 Keywords: 24/7 customer support, round-the-clock support, 24 7 customer service Offering 24/7 customer support in 2026 doesn't require 24/7 staffing. AI agents that resolve issues autonomously, async workflows that acknowledge and triage contacts outside business hours, and smart coverage models that match human staffing to peak demand hours make round-the-clock support achievable for teams of any size. The critical distinction: 24/7 availability without 24/7 resolution quality is not an improvement. A chatbot that deflects enquiries at 2am is available 24/7 - but it doesn't resolve anything. Lorikeet builds AI agents designed for after-hours resolution, not after-hours deflection. - More than half of customers globally expect 24/7 support availability, with the figure rising to 75% among younger demographics, per customer experience surveys from Document360 and Zoom - and unresolved after-hours contacts have higher churn correlation than daytime contacts. - AI agents handling eligible ticket types end-to-end deliver true 24/7 resolution - not just 24/7 acknowledgment - for the most common contact types. - After-hours contact patterns typically skew toward simpler, higher-volume issues: account access, order status, billing queries - all strong candidates for AI resolution. - Async-first workflows (AI triage + priority escalation) let human teams handle after-hours escalations in under 2 hours without staffing a night shift. The pressure for 24/7 support comes from customers who contact outside business hours and find nothing but an auto-reply. For B2C companies in e-commerce, fintech, or subscription businesses, these contacts are often high-intent and time-sensitive: a failed payment, an order that needs cancelling, an account that needs unlocking. Leaving them unresolved until morning creates churn risk. This guide covers the practical approaches to 24/7 coverage, how to choose the right model for your team, and what to measure to know it's working. ## What Does 24/7 Customer Support Actually Mean? True 24/7 customer support means customers get their issue resolved at any hour, not just acknowledged. Availability and resolution are different things. Many companies offer 24/7 "availability" through chatbots that surface FAQ articles and send auto-reply emails - but customers who contact at 2am with a real problem are still waiting until the next morning for resolution. The standard worth targeting is 24/7 resolution for eligible issue types, with clear expectations about human response windows for issues that require one. Understanding which issues customers bring outside business hours is the first step. For most B2C and SaaS businesses, after-hours contacts break down as: account access issues (passwords, MFA, locked accounts), order and subscription management (cancellations, updates, status checks), billing questions, and basic product troubleshooting. These are also the ticket types most amenable to AI resolution - well-defined, data-driven, and actionable without human judgement. ## What Are the Main Approaches to 24/7 Support? There are 4 practical approaches to 24/7 coverage. The right model depends on contact volume, after-hours issue complexity, and budget. Most teams end up combining 2 or 3 of these, not relying on a single approach exclusively. ### AI agents for eligible ticket types For the ticket types most common in after-hours queues - account access, order management, billing - AI agents that resolve end-to-end provide genuine 24/7 resolution without human involvement. The key requirement is that the AI is connected to the systems it needs (CRM, billing, order management) and has the action permissions to complete resolutions, not just diagnose them. An AI that can identify a billing issue but can't issue the credit doesn't provide 24/7 resolution - it provides 24/7 triage with a morning queue. ### Async-first workflows for complex issues Not every issue is AI-eligible. For complex or high-stakes contacts outside business hours, an async-first workflow offers a realistic alternative to human staffing: the AI triages and classifies the contact, sets accurate expectations with the customer (including a specific callback time), and flags high-priority issues to an on-call escalation path. Customers tolerate waiting for human resolution when they receive accurate timelines - the frustration comes from uncertainty, not from wait times per se. ### Offshore or shift-based human coverage For teams where after-hours volume justifies human staffing, distributed teams across time zones provide genuine 24/7 human coverage without night shifts. The model works when volume is sufficient to keep agents productive across shifts. Below a certain volume threshold, it's more cost-effective to combine AI resolution with async escalation than to staff shifts for low-density after-hours queues. ### Self-service for low-complexity queries A well-built knowledge base and self-service portal covers the simplest after-hours contacts - FAQs, how-to guides, status pages for outages. Self-service works when the content genuinely answers the question; it fails when it routes customers to articles that don't match their problem. Track self-service deflection quality rate (contacts that reach a human after attempted self-service) to understand whether your self-service is resolving or just delaying. ## How Do You Set Up 24/7 Support With AI? Setting up AI for 24/7 support requires 3 things: scope definition (which ticket types AI handles), system connectivity (what data and actions AI can access), and escalation logic (how AI hands off to humans when needed). - Start with your most common after-hours ticket types. Pull after-hours contact data to identify the top 5-10 ticket types by volume. These are your AI candidates. Prioritise ticket types that are well-defined, require no human judgement, and have clear resolution actions in your systems. - Give AI the system access it needs to resolve. An AI that can only read data but not write provides triage, not resolution. For 24/7 resolution to work, AI needs to query AND update the systems required to complete the most common resolution actions - account unlocks, status checks, subscription changes, refund processing within policy limits. - Define escalation triggers for after-hours routing. Decide which contacts warrant waking someone up vs. waiting until morning with a set expectation. High-value customers, active service failures, and urgent safety or compliance matters typically warrant escalation. Standard billing or account queries typically don't - they can wait with an accurate timeline set by the AI. - Set honest customer expectations about resolution windows. For contacts that AI can't resolve after hours, the single most important thing is setting an accurate expected resolution time. Customers who receive a specific callback commitment report higher satisfaction than those receiving a generic "we'll be in touch" auto-reply - even when the wait time is the same. ## What Results Should You Expect From 24/7 AI Coverage? The impact of 24/7 AI support shows up in 3 metrics: after-hours CSAT, churn rate for customers who contact outside business hours, and first-contact resolution rate across all hours. Teams deploying AI for after-hours coverage typically see after-hours CSAT improve significantly within the first quarter - primarily because customers who previously received no resolution now receive genuine issue resolution. After-hours contacts have higher churn correlation than daytime contacts, so resolving them effectively has an outsized retention impact relative to their volume. First-contact resolution rate across total contact volume typically improves by 10-15 percentage points when AI handles eligible after-hours ticket types, because these ticket types are often simpler and AI resolution rates for them are high. The cost picture is also favourable. AI-handled after-hours resolution costs $0.50-$2.37 per ticket vs. $20-$60 for equivalent human handling. For teams currently staffing night shifts to handle low-volume, high-routine contact queues, AI replacement of those --- ## How to Reduce First Response Time in Customer Service (2026) URL: https://www.lorikeetcx.ai/articles/how-to-reduce-first-response-time Updated: 2026-02-26 Keywords: how to reduce first response time, FRT reduction, customer service response time Reducing first response time (FRT) in customer service requires addressing the bottlenecks before and during ticket handling - not just speeding up individual agent responses. In 2026, the most effective FRT reductions come from AI agents that eliminate queuing for eligible ticket types, intelligent triage that routes contacts to the right handler immediately, and async workflows that acknowledge contacts within seconds. Teams that focus only on canned responses and macros improve FRT marginally; teams that fix routing and deploy resolution-focused AI see FRT drop by 60-80% for eligible ticket categories. Lorikeet's data shows that AI resolution of the top 20% of ticket types by volume removes those contacts from the FRT queue entirely. - 82% of customers expect a response within 10 minutes for live chat; email expectations are typically 4 hours, per Zendesk benchmarking data. - The average email first response time is approximately 12 hours, per SuperOffice research across 1,000 companies - far above the 4-hour window 46% of customers expect. - AI agents that resolve eligible ticket types have a near-zero FRT for those categories, since they respond instantly with no queue dependency. - Intelligent routing that eliminates misrouting reduces FRT by removing the queue re-entry that mis-triaged tickets require. First response time is often treated as an agent performance metric - something to improve through faster typing, better macros, and smaller queues. In reality, FRT is mostly a queue and routing problem. The time a ticket spends waiting for the first response is almost always longer than the time an agent takes to write it. Fixing the wait is more impactful than fixing the write. This guide covers where FRT time actually gets lost and the strategies with the highest leverage for reducing it. ## What Is First Response Time and Why Does It Matter? First response time is the time between a customer submitting a support request and receiving an initial substantive reply from a support agent or AI. It's one of the most visible customer service metrics because customers experience it directly. Long FRT signals to customers that their issue isn't a priority - and in channels like live chat where expectations are sub-2-minutes, any significant wait actively damages customer trust. Research from SuperOffice found that the average first response time for customer service emails is 12 hours - with only 36% of companies responding within the 4-hour window that 46% of customers expect. That gap represents a significant and addressable satisfaction problem for most support teams. ## Where Does FRT Time Actually Get Lost? Understanding where FRT delays occur is the first step to fixing them. Most FRT problems are not agent speed problems - they're queue, routing, and prioritisation problems. ### Queue wait time The majority of FRT in email and ticket-based support is queue wait time - the period between ticket creation and an agent opening the ticket. In high-volume environments during peak periods, queue wait can be 4-8 hours for non-urgent tickets. Reducing queue depth through AI resolution of eligible tickets is the highest-leverage FRT improvement available: when AI handles the top 20-30% of tickets by volume, queue depth drops significantly, and the remaining human-handled tickets wait less. ### Misrouting and re-queuing Tickets that arrive in the wrong queue, or are triaged to the wrong team, must be re-queued before getting a first response. A billing ticket that lands in the technical support queue can wait through the initial queue, be identified as misrouted, and then wait again in the billing queue. Each re-queue adds the full wait time again. Intelligent initial triage that routes tickets correctly on the first attempt eliminates this compounding delay. ### After-hours gaps Contacts submitted outside business hours accumulate in the queue until agents return. For B2C businesses with global customer bases, after-hours contacts can represent 30-50% of total volume. AI agents that resolve eligible ticket types 24/7 eliminate FRT for those categories regardless of when the contact arrives. For non-AI-eligible contacts, async workflows that send accurate acknowledgments with specific resolution timelines improve customer experience even when FRT on human resolution is unchanged. ## What Are the Highest-Impact Strategies to Reduce First Response Time? The strategies below are ordered by leverage - start with the ones that address queue and routing before optimising individual agent speed, as the former has 5-10x the impact of the latter. - Deploy AI agents for the highest-volume, best-defined ticket types. AI that resolves tickets end-to-end has an effective FRT of seconds - the system responds instantly when a ticket is submitted. For the ticket types AI handles, FRT becomes a non-issue. Focus AI deployment on the 20-30% of ticket types that generate 60%+ of volume, and FRT for those categories drops to near-zero immediately. - Fix triage and routing before optimising agent speed. A ticket in the right queue with the right agent context gets a first response faster than a ticket that gets re-queued twice. Implement routing logic that classifies ticket type and intent accurately at submission, and monitors misrouting rate (tickets re-assigned after initial routing) as a health metric. - Set up immediate automated acknowledgment for non-AI tickets. For tickets that will wait in a human queue, an immediate acknowledgment with an accurate expected response window sets expectations and reduces anxiety-driven follow-up contacts. Automated acknowledgments do not reduce FRT by definition - but they improve customer experience during the wait and reduce the repeat contacts that FRT delays generate. - Segment your queue by SLA priority, not arrival order. Treating every ticket with the same queue priority means high-value customers and urgent issues wait behind low-complexity routine queries. Priority segmentation based on customer tier, issue type, and urgency signals ensures your FRT improvement effort goes where impact is highest. - Reduce handle time through better knowledge access. Agent-side improvements - faster knowledge base search, better canned response libraries, and contextual CRM data surfaced automatically - reduce the time agents spend researching responses. This doesn't reduce queue wait time, but it does reduce the time from "agent opens ticket" to "first response sent," which matters particularly for synchronous channels like live chat. ## What FRT Improvements Should You Expect? FRT improvement depends heavily on which approach you take. Queue and routing fixes produce the largest absolute improvements; agent-side optimisation produces smaller but still meaningful gains. AI resolution of the top 20% of ticket types by volume reduces those tickets' FRT to near-zero and reduces overall queue depth - improving FRT for the remaining human-handled tickets as a secondary effect. Teams implementing this typically see overall average FRT drop by 40-60% within the first quarter after deployment. Fixing routing logic that eliminates re-queuing reduces FRT by 20-30% for misrouted ticket types - a significant improvement with relatively low implementation complexity. Priority queue segmentation reduces FRT for high-value and urgent tickets by 50-70%, at the cost of slightly increased FRT for low-priority contacts. For most businesses, this trade-off is strongly positive on customer satisfaction and retention impact. Combining AI resolution, routing improvement, and priority segmentation typically moves average FRT from 7-10 hours to under 2 hours for email within a quarter - pushing most teams above the 4-hour threshold that 46% of customers consider acceptable. ## Lorikeet's Take on First --- ## First Response Time Benchmarks for Customer Service in 2026 URL: https://www.lorikeetcx.ai/articles/first-response-time-benchmark-customer-service Updated: 2026-02-26 Keywords: first response time benchmark, FRT benchmark, email response time benchmark First response time (FRT) benchmarks vary significantly by channel: under 40 seconds for live chat, under 1 hour for social media, and under 4 hours for email is where top-performing support teams operate in 2026. Industry averages fall well short of these targets - the average email first response time across industries is 7-10 hours, despite 46% of customers expecting a reply within 4 hours. Understanding channel-specific benchmarks is the starting point for setting realistic FRT targets and identifying where your team's performance stands relative to industry standards. Lorikeet's AI-native resolution approach eliminates FRT entirely for eligible ticket types. - Live chat benchmark: under 2 minutes average, with 40 seconds considered strong performance by Zendesk research. - Email benchmark: top performers achieve under 4 hours; industry average is 7-10 hours across all industries. - Social media benchmark: under 60 minutes recommended; customers on social expect faster responses than email. - 82% of customers expect a response within 10 minutes, per customer expectations surveys - significantly faster than most team averages. FRT benchmarks matter because customer expectations vary by channel, business type, and issue urgency. A 4-hour email FRT is acceptable for most SaaS support contexts but would be catastrophic for an e-commerce chat during a purchase flow. Getting benchmarks right means knowing which channel you're measuring and which customer expectation you're targeting, not applying a single standard across all channels. This guide covers FRT benchmarks by channel, industry-specific context, and how to interpret your numbers relative to what customers actually expect. ## What Is First Response Time and How Is It Measured? First response time is the elapsed time between when a customer submits a support request and when they receive the first substantive reply from a support agent or AI. For ticket-based channels (email, help desk), it's measured from ticket creation to first non-automated response. For live chat, it's measured from when the customer sends their first message to when an agent or AI replies. FRT does not include automated acknowledgment messages - it measures the first response that addresses the customer's actual issue. Most modern support platforms (Zendesk, Intercom, Freshdesk) track FRT automatically at the ticket level. Reporting at the median rather than the mean is more useful, since a small number of extremely delayed tickets can inflate average FRT significantly, masking the typical customer experience. Track both: median FRT (typical experience) and 90th percentile FRT (worst-case experience for 10% of customers). ## What Are the FRT Benchmarks by Channel? Benchmarks differ substantially across channels because customer expectations - and technical norms - differ across channels. Measuring against the right benchmark for each channel is essential for accurate performance assessment. ### Live chat Live chat has the most demanding FRT expectations of any channel. Zendesk identifies 40 seconds as a strong first response benchmark for live chat, with the industry average sitting around 2 minutes. Customers who initiate a chat session expect near-immediate acknowledgment - wait times above 3-5 minutes in live chat correlate with significantly higher abandonment rates. AI-powered chat agents that respond instantly to eligible queries have an effective FRT of under 5 seconds, which is why they're increasingly used as the first layer in live chat workflows. ### Email support Email has the widest gap between benchmark and actual performance of any channel. SuperOffice research found the industry average email first response time is 12 hours, with only 36% of companies responding within 4 hours. The customer expectation, meanwhile, is 4 hours for 46% of customers and under 24 hours for the vast majority. Top-performing support teams target under 4 hours as a floor; premium tiers often commit to under 2 hours or under 1 hour for enterprise customers. ### Social media Social media FRT expectations fall between live chat and email. The recommended benchmark for social responses is under 60 minutes, with many brands publishing a 1-hour commitment publicly. Response times above 4 hours on social media correlate with public escalation - customers who don't receive timely social responses are more likely to post publicly about the failure, which multiplies the reputational impact beyond the original interaction. ### Phone support Phone FRT (time-to-answer or hold time) benchmark is 20-80 seconds, with most centres targeting under 30 seconds for standard priority queues. Unlike other channels, phone FRT has immediate abandonment pressure - customers who wait more than 2-3 minutes on hold abandon at high rates. Call abandonment rate is the key secondary metric to track alongside phone FRT. ## What Are FRT Benchmarks by Industry? Industry context matters because support complexity, customer expectations, and ticket types differ significantly across sectors. - E-commerce and retail. FRT expectations are high because many contacts are time-sensitive (order cancellations, payment failures, delivery issues). Email FRT target: under 2 hours for standard; under 30 minutes for priority or premium tiers. Chat FRT: under 1 minute, since purchase-flow contacts often require near-instant response to prevent cart abandonment. - SaaS and software support. FRT expectations vary by customer tier. Enterprise customers typically expect under 1 hour; standard plan customers expect under 4 hours. Chat and in-app support FRT is typically held to a higher standard than email. B2B SaaS with service-level agreements typically have contractually defined FRT targets by severity level. - Financial services. Strict regulatory and compliance context means FRT for urgent matters (fraud alerts, failed transactions) needs to be near-real-time. General enquiries typically operate on 4-hour email standards. Phone FRT for urgent matters is typically under 2 minutes with direct escalation paths. - Healthcare and insurance. High-sensitivity ticket types (claims, coverage questions, urgent medical matters) often have regulatory FRT requirements in addition to customer expectations. Standard enquiry FRT benchmarks are similar to financial services - under 4 hours for email, under 2 minutes for phone. ## How Do Top-Performing Teams Hit FRT Benchmarks? Teams that consistently beat FRT benchmarks typically use a combination of AI resolution, intelligent routing, and queue prioritisation - not just agent training and macros. AI agents that resolve the top 20-30% of ticket types by volume (account access, order status, billing queries) remove those tickets from the FRT queue entirely. Their effective FRT is seconds - no queue wait, no agent involvement. This reduces total queue depth, which improves FRT for all remaining human-handled tickets. Teams that combine AI resolution with priority queue segmentation - routing high-value customers and urgent issues to the front of the human queue - see email FRT improve from 7-10 hours to under 2 hours for priority segments within a quarter. Detailed FRT reduction strategies cover the routing and triage changes that produce the largest improvements. The teams that struggle with FRT are typically treating it as an agent efficiency problem when it's a queue design problem. Queue depth and routing quality explain 80% of FRT variation; individual agent response speed explains the rest. ## Lorikeet's Take on FRT Benchmarks At Lorikeet, FRT benchmarks are a starting point, not the end goal. A team that achieves 4-hour email FRT by sending low-quality first responses that generate 3 follow-up contacts hasn't improved customer experience - it's just moved the bottleneck. Lorikeet's view is that FRT and first-contact resolution rate need to be tracked together: a fast first response that resolves the issue --- ## How Do AI Guardrails Work? Types, Triggers, and Configuration URL: https://www.lorikeetcx.ai/articles/how-do-ai-guardrails-work Updated: 2026-02-26 Keywords: how do AI guardrails work, AI guardrail types, guardrail configuration AI guardrails work by intercepting and evaluating both inputs (what customers say) and outputs (what the AI responds) against a set of defined rules before any customer-facing action is taken. When a guardrail is triggered, it either blocks the action, modifies the response, or escalates to a human agent depending on configuration. In 2026, well-designed guardrail systems allow Lorikeet and other AI platforms to operate in regulated, high-stakes environments by ensuring AI actions stay within defined policy boundaries at every step. - Guardrails operate inline in the AI pipeline - they evaluate every input and output before the customer interaction continues, with no perceptible latency in most implementations. - Input guardrails detect intent, topic type, sentiment, and sensitive data before the AI generates a response - preventing problems before they start. - Output guardrails validate AI responses for policy compliance, factual accuracy constraints, and unauthorised commitments before delivery to the customer. - Escalation guardrails are triggered by sentiment, topic flags, or resolution failure signals - routing contacts to human agents at precisely the right moment. Understanding how guardrails work mechanically helps support leaders configure them correctly and interpret guardrail trigger data accurately. A guardrail that fires too rarely signals misconfiguration - not a perfectly behaving AI. A guardrail that fires constantly signals scope creep or an AI operating outside its intended boundaries. This guide explains the operational mechanics of AI guardrails in customer service contexts - how they're triggered, what they do, and how to calibrate them for production deployment. ## How Do AI Guardrails Process Each Customer Interaction? Guardrails operate as a series of inline checks throughout the conversation flow - before the AI generates a response (input guardrails), after generation but before delivery (output guardrails), and after resolution actions are proposed (action guardrails). The pipeline runs on every message exchange, typically in under 100 milliseconds, making it invisible to customers in most implementations. IBM's AI guardrail documentation describes the core function as "identifying and mitigating unsafe interactions in real time" - the key word being real-time. Unlike post-hoc content moderation that reviews interactions after they've happened, inline guardrails intercept problems before the customer sees them. This is what makes them operationally viable for live customer support rather than just audit tools. ## How Do Input Guardrails Work? Input guardrails are the first line of evaluation in the AI pipeline. They process the customer's message before the AI generates any response, classifying the input across several dimensions to determine how it should be handled. ### Intent classification Intent classification identifies what the customer is trying to accomplish - whether the request falls within the AI's permitted scope and what resolution path is appropriate. A customer message asking "can I get a refund?" gets classified as a refund intent and routed to the refund handling workflow. A message asking for "legal advice about my contract" gets classified as out-of-scope and escalated to a human before the AI attempts a response it shouldn't generate. ### Sensitive data detection Input guardrails scan for personally identifiable information (PII), financial data, health information, and other regulated content types. When sensitive data is detected, guardrails can redact it from logs, flag it for compliance review, or modify how the AI handles the interaction. PII detection prevents sensitive data from being stored inappropriately or processed by AI models that aren't authorised to handle it. ### Sentiment and escalation signals Sentiment analysis within input guardrails detects frustration, distress, and urgency in customer messages - and triggers escalation before the AI attempts a response that might worsen the situation. A customer expressing anger after a billing error gets routed to a human agent rather than generating an AI-crafted apology that doesn't address the root cause. ## How Do Output Guardrails Work? Output guardrails evaluate the AI's generated response before it's delivered to the customer. They're the last checkpoint before a response becomes a customer-facing commitment, making them the most critical safety layer for business risk management. - Policy compliance checks. Output guardrails verify that AI responses don't contain commitments that violate policy - refund amounts above thresholds, service level promises that don't exist, or pricing information that's outdated. When a policy violation is detected, the guardrail blocks the response and either generates a compliant alternative or escalates to a human. McKinsey identifies unauthorised commitments as one of the primary business risks addressed by AI guardrails in customer-facing applications. - Factual accuracy constraints. For domains where factual accuracy is verifiable (order status, account balance, policy terms), output guardrails can validate AI responses against live data sources before delivery. If the AI's response contains information that doesn't match current system data, the guardrail intercepts it and regenerates with current data. - Tone and content standards. Output guardrails enforce brand voice, professional tone standards, and content policies. Responses that are too informal, too technical for the audience, or contain content that violates community standards are modified or blocked before delivery. - Scope enforcement. Responses that address topics outside the AI's permitted scope - legal, medical, regulatory advice, or topics explicitly excluded from the AI's brief - are blocked and replaced with an appropriate escalation message, regardless of whether the AI's attempted answer seems reasonable. ## How Do Action Guardrails Work? Action guardrails apply when AI agents take actions in connected backend systems - issuing refunds, updating account settings, processing cancellations. They're distinct from input and output guardrails because they operate on operations that affect real-world state, not just what's said in the conversation. Action guardrails evaluate each proposed action against defined parameters: amount limits for financial transactions, confirmation requirements for irreversible actions, audit logging requirements for compliance, and confidence thresholds that determine whether the AI has sufficient certainty to take the action unilaterally or should confirm with a human. Taking actions in backend systems safely requires action guardrails that match each action type to its risk level - password resets operate differently from account closures, which operate differently from refund processing. Confidence thresholds are a particularly important action guardrail mechanism: if the AI is below a defined confidence level for a given resolution (say, 85%), the action guardrail routes the interaction to human review rather than proceeding autonomously. This provides a risk-calibrated approach to automation that expands AI authority as confidence is validated by production data. ## How Do You Calibrate Guardrails After Deployment? Guardrail calibration is an ongoing process, not a one-time setup. Production data reveals edge cases, misconfigured triggers, and scope gaps that aren't visible during pre-launch testing. Monitor 3 metrics during calibration: guardrail trigger rate (what percentage of interactions trigger each guardrail), false positive rate (guardrails blocking or escalating interactions that should have been handled normally), and false negative rate (interactions that passed guardrails but contained problems identified in QA review). A trigger rate that's unexpectedly high usually means guardrail scope is too broad or thresholds are misconfigured. A false positive rate above 5% means guardrails are blocking legitimate resolutions, hurting CSAT and increasing escalation cost. False negatives identified in QA review of AI interactions are the most important signal - they reveal where guardrails need to be tightened. ## Lorikeet's Take on How AI Guardrails Should Work At Lorikeet, we design guardrails around confidence thresholds and action --- ## How FinTech Companies Are Using AI to Handle Complex Complaints URL: https://www.lorikeetcx.ai/articles/ai-customer-support-fintech Updated: 2026-03-11 Keywords: AI customer support fintech, fintech complaints AI, AI compliance fintech FinTech support teams spend $6-8 per human-handled complaint while AI resolves the same issues for under $1 - yet most still route 80% of tickets to agents. AI customer support in fintech uses artificial intelligence to manage, triage, and resolve customer service interactions for financial technology companies - including complaints, account disputes, and compliance-sensitive inquiries. The AI in fintech market reached USD 30 billion in 2025, growing at 22.6% CAGR according to Mordor Intelligence. Key points: AI handles up to 98% of queries at leading neobanks like WeBank. FinTech teams report 40-75% of queries resolved instantly. Compliance costs consume 15-25% of operating expenses. Governance and guardrails are the critical differentiator. According to IBM's Institute for Business Value, AI-powered customer service delivers $3.50 in ROI per $1 invested. Gartner projects conversational AI will reduce contact centre labour costs by $80 billion by 2026. Per-ticket costs drop from $6-8 to under $1 for automated cases. Lorikeet is an AI customer support platform that resolves fintech tickets end-to-end with configurable compliance guardrails and full audit trails. --- ## KYC Automation in 2026: How Automated KYC Verification Works URL: https://www.lorikeetcx.ai/articles/kyc-automation-fintech Updated: 2026-03-11 Keywords: KYC automation, KYC automation fintech, automated KYC verification The average bank spends $1,500-3,000 to complete a single client's KYC review - while automated systems do it in under 30 seconds. KYC automation uses AI and digital verification to streamline Know Your Customer checks. The global KYC market is projected to grow from $6.73 billion (2025) to $16.31 billion by 2031, per Mordor Intelligence. 54% of banks spend $1,500-3,000 per manual KYC review (Fenergo). Automated KYC reduces costs by up to 70% and cuts verification time by 78% (AU10TIX). 70% of financial firms lost clients due to slow onboarding. AI adoption in KYC jumped from 42% to 82% in one year, yet only 4% have fully automated. McKinsey reports agentic AI delivers 200-2,000% productivity uplifts in KYC. Total KYC/KYB spend projected at $35.5 billion in 2026 (Juniper Research). Lorikeet handles the customer-facing side of KYC workflows - guiding applicants through document submission and status updates across chat, email, and voice. --- ## Customer Onboarding Automation: How Fintech Teams Stop the 68% Dropout URL: https://www.lorikeetcx.ai/articles/customer-onboarding-automation-fintech Updated: 2026-03-11 Keywords: customer onboarding automation, fintech onboarding, digital onboarding financial services 68% of consumers have abandoned a financial onboarding process at least once - and the average dropout happens at just 18 minutes and 53 seconds. Customer onboarding automation uses AI and digital workflows to bring new customers from application to active account. 70% of financial firms lost clients due to slow onboarding (Fenergo 2025). Fully digital onboarding drives 20% higher acquisitions and 15% cost reduction (McKinsey). BCG reports 60% efficiency gains and 40% cost reductions with AI-driven onboarding. 63% of customers more likely to stay with banks offering smooth digital onboarding (Incode/The Financial Brand). Top dropout reasons at 21% each: process too long, too much info required, changed mind. 38% lacked documents. Lorikeet manages the customer communication layer during onboarding - answering questions, guiding document resubmission, and providing proactive status updates. --- ## AI in Financial Services: Where It Works and What It Costs in 2026 URL: https://www.lorikeetcx.ai/articles/ai-in-financial-services Updated: 2026-03-11 Keywords: AI in financial services, AI banking customer service, artificial intelligence finance Bank of America's AI assistant Erica has handled over 3 billion customer interactions - saving the equivalent of 11,000 employees' daily workload. AI in financial services automates customer interactions, detects fraud, assesses risk, and streamlines compliance. 59% of finance leaders use AI (Gartner 2025), up from 37% in 2023. McKinsey projects 15-20% net cost reductions for banks. 54% of US bank customer interactions are fully automated (CoinLaw). BofA Erica: 3 billion interactions, 50 million users, 58 million interactions/month. Customer service chatbots account for 32.5% of the AI agents market (Precedence Research). GenAI fraud losses projected at $40 billion by 2027 (Deloitte). AI agents market valued at $1.79 billion (2025), projected $6.54 billion by 2035. Gartner: conversational AI to reduce contact centre costs by $80 billion by 2026. Lorikeet resolves financial services support tickets end-to-end with configurable guardrails ensuring every AI action stays within regulatory boundaries. --- ## What Is a Good NPS Score? Benchmarks by Industry in 2026 URL: https://www.lorikeetcx.ai/articles/what-is-a-good-nps-score Updated: 2026-03-12 Keywords: what is a good nps score, nps benchmarks, net promoter score by industry NPS fell in 20 out of 39 industry-country combinations last year, yet most benchmark guides still treat it as a universal scorecard. A good Net Promoter Score (NPS) in 2026 depends entirely on your industry, business model, and channel mix. Any score above 0 means more promoters than detractors. A score above 30 is strong. Above 50 is excellent. Above 70 puts you in the top tier globally. But a "good" score in SaaS (median 36) would be below average in manufacturing (median 65). - B2C companies average an NPS of 49 while B2B companies average 38, an 11-point gap driven by relationship complexity - Technology and services industries lead with median NPS scores of 60-66, while B2B software sits at 36-41, according to Retently's 2025 benchmarks - Only 23% of enterprise CX leaders still use NPS as a primary performance metric, per a TELUS Digital and Statista survey - NPS explains roughly 1% of variance in customer spending, according to Marketing Science Institute research covering 250,000+ ratings Last updated: March 2026 NPS remains one of the most widely tracked customer experience metrics, even as its limitations become clearer. The question "what is a good NPS score?" has no single answer because benchmarks vary by 40+ points across industries. What matters more than the number itself is understanding what drives it. Support quality, resolution speed, and effort required all feed directly into whether a customer becomes a promoter or detractor. The companies seeing the biggest NPS gains are not running better survey programs. They are fixing the operational inputs that NPS measures: response times, first-contact resolution, and how much effort customers expend to get help. ## What Is a Good NPS Score in 2026? A good NPS score in 2026 falls between 30 and 50 for most industries. Scores above 50 are excellent, and scores above 70 are world-class. However, "good" is relative. A SaaS company with an NPS of 40 is outperforming its industry median. A healthcare company with the same score may be significantly below its peers. NPS measures the likelihood that customers will recommend your company. Respondents score from 0 to 10. Scores of 9-10 are promoters, 7-8 are passives, and 0-6 are detractors. Your NPS equals the percentage of promoters minus the percentage of detractors, yielding a score from -100 to +100. Net Promoter Score (NPS): A customer loyalty metric calculated by subtracting the percentage of detractors (0-6 ratings) from the percentage of promoters (9-10 ratings), producing a score from -100 to +100. Lorikeet is an AI customer support platform that resolves tickets end-to-end, processing refunds, updating accounts, and handling complex multi-step workflows across chat, email, and voice. Teams using Lorikeet consistently report NPS improvements because customers receive instant resolution instead of queue-based responses. ## How Do NPS Benchmarks Vary by Industry? NPS benchmarks vary by more than 40 points between the highest and lowest-performing industries. Manufacturing leads at a median of 65. Software and SaaS sit near the bottom at 30-36. This gap exists because industries with tangible products and simple interactions generate higher satisfaction than those with complex, ongoing service relationships. ### Industry Breakdown According to Retently's 2025 benchmark data, technology and professional services lead with median scores of 60-66. Healthcare ranges from 53-80 depending on segment and methodology. Retail averages around 50. Banking and hospitality sit at 41-44. B2B software and SaaS averages 36-41. Insurance shows the widest variance of any industry, ranging from 23 to 80 depending on segment. ### B2B vs B2C Split B2C companies average an NPS of 49 versus 38 for B2B. Healthcare shows the largest gap at 32 points (B2C: 70, B2B: 38). The difference comes down to transaction simplicity. B2C interactions tend to be shorter, more transactional, and easier to resolve. B2B involves longer sales cycles, multiple stakeholders, and complex support needs. > "NPS is a relationship metric. It measures loyalty attitudes, not customer experiences."— Qualtrics XM Institute ## What Drives NPS Up or Down Operationally? NPS moves based on three operational inputs: resolution speed, first-contact resolution rate, and customer effort. Companies that improve these three metrics see NPS gains of 15-25 points within 90 days. Companies that only send more surveys see no change. - First-contact resolution rate. Every ticket that requires a callback or escalation creates a potential detractor. Teams with fast first response times and high FCR rates (above 70%) consistently score 10-20 NPS points higher than teams with FCR below 50%. - Customer effort. Gartner research shows 96% of high-effort customers become disloyal versus just 9% of low-effort customers. Reducing the steps to get help directly lifts NPS. - Resolution quality. Deflecting a ticket to a help article is not the same as resolving it. The distinction between deflection and resolution shows up directly in NPS. Customers whose issues are fully resolved score 20-30 points higher than those who were merely deflected. ## What NPS Results Can You Expect from AI-Native Support? Teams that deploy AI agents capable of end-to-end resolution, not just FAQ deflection, consistently see measurable NPS improvements. The gains come from eliminating wait times, reducing effort, and resolving issues in a single interaction. According to McKinsey, advanced AI deployments reduce service interactions by 40-50%. First-contact resolution typically moves from 20-30% with traditional chatbots to 55-70% with AI agents that access backend systems. CSAT scores improve by 15-25 points when customers receive instant resolution. Average handle time drops from 8-12 minutes to under 3 minutes for routine requests. Cost per resolution falls from $8-12 to $1-3 when AI handles the full workflow. These improvements feed directly into NPS because each resolved interaction removes a potential detractor and creates a potential promoter. Teams using AI-native support see 3x faster resolution and 40-50% fewer repeat contacts. See how Lorikeet handles end-to-end ticket resolution. ## Should You Still Use NPS as Your Primary CX Metric? NPS remains useful as a directional indicator of customer loyalty, but it should not be your only CX metric. The Marketing Science Institute found that NPS explains roughly 1% of variance in share of wallet across 250,000+ consumer ratings. Use it alongside operational metrics that you can actually act on. According to Forrester's 2025 Global NPS Rankings, NPS fell in 20 out of 39 industry-country combinations, rose in only 3, and held steady in 16. The broader trend is clear: NPS is declining industry-wide, and the companies maintaining or growing their scores are those investing in resolution quality, not survey frequency. Pair NPS with CSAT, customer effort score, and first-contact resolution rate. Together, these four metrics give you a complete picture of customer experience. NPS tells you the outcome. The operational metrics tell you why. ## Lorikeet's Take on NPS At Lorikeet, we have seen teams lift NPS by 15-25 points when they shift focus from survey optimization to resolution quality. Most CX vendors will tell you the path to better NPS is better survey design or faster follow-up on detractor feedback. That helps at the margins. The real lever is resolving customer issues completely in the first interaction, without transfers, callbacks, or "check our help center" deflections. Lorikeet's approach to NPS is operational: give AI agents the ability to take action in backend systems, and the score takes... --- ## What Is Customer Effort Score (CES)? Benchmarks and How to Improve URL: https://www.lorikeetcx.ai/articles/what-is-customer-effort-score Updated: 2026-03-12 Keywords: what is customer effort score, CES benchmarks, customer effort score calculation Your customers are not leaving because your product is bad - they are leaving because getting help is too hard. Customer Effort Score (CES) is a service metric that quantifies how easy or difficult customers find it to resolve an issue, measured on a 1-5 or 1-7 Likert scale. In 2026, with 96% of high-effort customers becoming disloyal according to Gartner, CES has become one of the most actionable indicators of support quality. - CES measures ease of resolution, not satisfaction or sentiment - scored on a 1-7 Likert scale - Low-effort experiences drive 94% repurchase rates and 88% increase spending, per Gartner - CES below 70% signals improvement needed; above 90% indicates strong support operations - CES predicts purchasing behavior more accurately than CSAT or NPS Last updated: March 2026 Most support teams obsess over satisfaction scores while ignoring the friction that actually drives churn. CES flips the script. Instead of asking "How happy are you?" it asks "How easy was that?" The difference matters. Gartner found that 96% of customers who experience high effort become disloyal, compared to just 9% of low-effort customers. This guide covers what CES is, what good looks like, and how to reduce effort across every support interaction. ## What Is Customer Effort Score and How Is It Calculated? Customer Effort Score measures how much work a customer has to do to get their issue resolved. It is calculated by asking a single question - typically "How easy was it to handle your issue?" - on a Likert scale, then averaging responses or calculating the percentage of positive ratings. The most common approach uses a 1-7 scale, where 1 means "very difficult" and 7 means "very easy." To calculate CES as a percentage, divide the number of positive responses (typically 5, 6, or 7) by total responses, then multiply by 100. For example, if 340 out of 400 respondents rate their experience a 5 or higher, your CES is 85%. Some teams use a simpler 1-5 scale, but the principle stays the same: higher scores mean lower effort. Customer Effort Score (CES): A post-interaction metric that measures how easy or difficult a customer found it to resolve their issue, typically on a 1-7 Likert scale from "very difficult" to "very easy." Lorikeet is an AI customer support platform that resolves tickets end-to-end - processing refunds, updating accounts, and handling complex multi-step workflows across chat, email, and voice. By handling complex workflows autonomously, Lorikeet reduces the effort customers spend reaching a resolution, directly improving CES. ## What Is a Good Customer Effort Score in 2026? A good Customer Effort Score sits above 80% on a percentage scale, meaning 4 out of 5 customers found it easy to get help. According to Gartner benchmarks, CES below 70% signals a need for improvement, while scores above 90% indicate strong, low-effort support operations. Context matters significantly. Industries with straightforward transactions like e-commerce tend to score higher (82-90%), while complex environments like B2B enterprise support often score lower (65-75%). A CES of 75% might be excellent for a technical SaaS platform but mediocre for a retail brand. Rather than chasing a universal number, benchmark against your own historical performance and your industry peers. The trend over time matters more than any single reading. ## How Does CES Compare to CSAT and NPS? CES measures ease of experience, CSAT measures satisfaction with a specific interaction, and NPS measures overall brand loyalty. CES is the strongest predictor of repeat purchase behavior because it captures the friction that actually drives customers away. As the Qualtrics XM Institute puts it: "NPS is a relationship metric. It measures loyalty attitudes, not customer experiences." CES fills that gap by focusing on what happened, not how someone feels about your brand in general. The smartest support teams track all 3 metrics together. Use CES to identify friction points, CSAT to measure interaction quality, and NPS to track long-term loyalty trends. Each answers a different question. ## What Causes High Customer Effort? The biggest drivers of high customer effort are channel switching, repeat contacts, transfers between agents, and having to re-explain issues. Gartner's research shows that 96% of customers who experience these friction points become disloyal - making effort reduction the single highest-leverage retention strategy. Repeat contacts are especially damaging. When a customer reaches out 2 or 3 times for the same issue, effort compounds and frustration escalates. This also inflates your cost per ticket while destroying loyalty. Other common causes include confusing self-service options, slow escalation paths, and agents who lack the context or authority to resolve issues on first contact. Each handoff and each delay adds effort. ## What Results Can You Expect from Reducing Customer Effort? Reducing customer effort produces measurable improvements in loyalty, spending, and support costs. The gains compound because low-effort customers stay longer, buy more, and generate fewer repeat tickets. According to Gartner, 94% of low-effort customers repurchase compared to just 4% of high-effort customers. Low-effort customers are also 88% more likely to increase their spending. McKinsey found that advanced AI deployments reduce service interactions by 40-50%, directly cutting the number of touchpoints customers need to reach resolution. Teams deploying AI agents that resolve issues end-to-end see first response times drop while repeat contact rates fall by 30-50%. Teams using AI-native support cut repeat contacts by 30-50% and reduce total customer effort. See how Lorikeet resolves tickets end-to-end. ## How Can AI Reduce Customer Effort Score? AI reduces customer effort by resolving issues faster, eliminating transfers, and providing instant responses across channels. But not all AI reduces effort equally. Basic chatbots that deflect without resolving often increase effort because customers end up repeating themselves to a human agent. Lorikeet's Resolution Loop takes a different approach - resolving complex workflows autonomously so customers never need to follow up, switch channels, or re-explain their problem. When AI can actually process a refund or update an account mid-conversation, the effort drops to near zero. If your CSAT is dropping, high effort is likely a contributing factor. AI that resolves rather than deflects addresses both metrics simultaneously. ## Lorikeet's Take on Customer Effort Score At Lorikeet, we believe effort reduction should be the primary design principle for every support operation. Most CX vendors tell you to measure CES and follow up with detractors. That helps at the margins. The real lever is eliminating the effort in the first place - no transfers, no callbacks, no "check our help center" deflections. The data supports this approach. Gartner shows 88% of low-effort customers increase their spending, making effort reduction a revenue strategy, not just a support optimization. Lorikeet is built around this principle: give AI agents the ability to take action in backend systems so customers get resolution in a single interaction. If effort reduction matters to your team, see how Lorikeet drives it through resolution. ## Key Takeaways - CES measures ease of resolution on a Likert scale - aim for above 80%, with 90%+ being strong per Gartner - 96% of high-effort customers become disloyal vs just 9% of low-effort customers - Low-effort customers repurchase at 94% and 88% increase their spending - Track CES alongside CSAT and NPS - each metric answers a different... --- ## Customer Service Metrics That Actually Matter in 2026 URL: https://www.lorikeetcx.ai/articles/customer-service-metrics Updated: 2026-03-12 Keywords: customer service metrics, contact center KPIs, support metrics 2026 Most support teams track 15+ metrics and still can't explain why customers leave - because they're measuring activity, not outcomes. Customer service metrics are the quantitative measures teams use to evaluate support quality, efficiency, and cost. In 2026, the 8 metrics that matter most have shifted - with benchmarks like 85%+ CSAT targets and average first contact resolution hovering at 70% according to SQM Group - yet most dashboards still drown teams in vanity KPIs. - The 8 customer service metrics that drive retention and cost savings in 2026 - Current benchmarks for CSAT, NPS, AHT, FCR, CES, and AI-era KPIs - Why resolution rate matters more than deflection rate - How leading teams use AI to improve metrics without sacrificing quality Last updated: March 2026 Contact centers in 2026 operate in a different reality. Gartner projects that 80% of contact centers will use AI for routing or coaching by year's end. McKinsey reports that AI deployments reduce service interactions by 40-50%. The metrics your team tracked in 2023 may no longer tell you what you need to know. This guide narrows the field to 8 contact center KPIs that connect directly to customer retention, operational cost, and team performance. We include current benchmarks, explain what each metric actually reveals, and flag the AI-era metrics most competitors ignore. ## What Are the Core Customer Service Metrics Every Team Should Track? Every support team needs a foundation of 4 proven metrics: CSAT, NPS, First Contact Resolution, and Average Handle Time. These measure satisfaction, loyalty, effectiveness, and efficiency respectively. Together they give a baseline view of how your operation performs against industry standards. Lorikeet is an AI customer support platform that resolves complex tickets end-to-end. With tools like Coach and Resolution Loop, Lorikeet helps teams improve these metrics by automating resolution - not just deflection. CSAT (Customer Satisfaction Score): The industry target is 85%+. CSAT captures how a customer felt about a single interaction. It is fast to collect and easy to benchmark, but it only reflects the moment - not long-term loyalty. For a deeper breakdown, see our guide on what is a good CSAT score. NPS (Net Promoter Score): B2C averages sit at 49 and B2B at 38. Forrester found NPS fell in 20 out of 39 industry-country combinations in 2025, suggesting that customer expectations are outpacing service improvements. NPS is useful for tracking brand-level loyalty over quarters, not individual interactions. First Contact Resolution (FCR): SQM Group puts the average FCR benchmark at 70%, with top performers reaching 85%. FCR is the single strongest predictor of customer satisfaction. Every percentage point improvement directly reduces repeat contacts and cost. Read more on first response time benchmarks. Average Handle Time (AHT): Industry benchmarks range from 4-7 minutes for voice. AHT should never be optimized in isolation - pushing agents to rush calls tanks FCR and CSAT. The goal is lower AHT as a byproduct of better tools and training, not as a target. ## Which Efficiency Metrics Separate High-Performing Contact Centers? Customer Effort Score and call abandonment rate separate average teams from high performers. CES measures how hard customers work to get help. Abandonment rate reveals operational bottlenecks. Gartner found that 96% of high-effort customers become disloyal - making CES one of the most predictive metrics available. Customer Effort Score (CES): CES asks customers to rate how easy it was to resolve their issue. Low-effort experiences drive retention far more reliably than "delightful" ones. Teams using Lorikeet's Resolution Loop reduce customer effort by resolving issues fully on first contact, without transfers or callbacks. Call Abandonment Rate: A healthy rate sits between 2-5%. Anything above 8% signals serious staffing or routing problems. Research shows 60% of customers feel that holding for even 1 minute is too long. This metric is a leading indicator - it spikes before CSAT drops. ## What AI-Era Metrics Are Most Teams Missing? Resolution rate and cost per resolution are the 2 metrics that matter most in AI-augmented support, yet most dashboards still track deflection rate instead. Deflection measures tickets avoided. Resolution rate measures tickets actually solved. The difference determines whether your AI is helping customers or just hiding them. Resolution Rate vs Deflection Rate: Deflection counts a ticket as "handled" if the customer stops contacting you. Resolution confirms the issue was actually fixed. A chatbot that frustrates customers into giving up shows great deflection numbers and terrible retention. Lorikeet focuses on verified resolution - confirming the outcome, not just closing the ticket. Cost Per Resolution: Gartner benchmarks median cost per contact at $1.84 for self-service versus $13.50 for assisted channels. But cost per contact is incomplete - it does not account for repeat contacts from unresolved issues. Cost per resolution divides total support spend by tickets genuinely resolved. For a full cost breakdown, see our guide on customer service cost per ticket. Teams that track resolution rate and cost per resolution see where their real efficiency gains are. See how Lorikeet measures verified resolution. ## How Should You Benchmark Your Customer Service Metrics? Benchmark against your own trendlines first, industry averages second. A CSAT of 82% is strong for a complex B2B product but weak for a simple e-commerce operation. Context matters more than raw numbers. Use industry data as a starting point, then track month-over-month improvement relative to your own baseline. Key benchmarks for 2026: - CSAT: 85%+ target - NPS: 49 B2C average, 38 B2B average - FCR: 70% average, 85% top performers - AHT: 4-7 minutes (voice) - Abandonment: 2-5% healthy range - Cost per contact: $1.84 self-service, $13.50 assisted (Gartner) Benchmarks shift when AI handles a meaningful share of volume. Teams using Lorikeet typically see AHT and cost per resolution drop as the AI resolves straightforward tickets, freeing agents for complex work. Tracking these metrics with and without AI-handled tickets gives you the clearest picture. ## How Do You Connect Customer Service Metrics to Business Outcomes? Link every metric to either retention, revenue, or cost. CSAT and CES predict retention. FCR and resolution rate predict repeat contact volume. Cost per resolution predicts operational margin. If a metric does not connect to one of these 3 outcomes, question whether it belongs on your dashboard. The trap is reporting metrics without acting on them. A weekly QA process that reviews scores alongside agent conversations turns numbers into coaching opportunities. Lorikeet's Coach automates this loop, flagging where AI and human performance diverge so managers spend time on improvements, not spreadsheets. ## Lorikeet's Take on Customer Service Metrics At Lorikeet, we believe the shift from deflection to resolution is the most important measurement change in support this decade. Deflection-focused metrics reward AI that pushes customers away. Resolution-focused metrics reward AI that solves problems. Every Lorikeet deployment is measured on verified resolution rate, not ticket avoidance. We also see cost per resolution replacing cost per contact as the standard efficiency metric. When AI resolves a $13.50 assisted-channel ticket for under $2, the savings are real and measurable. When AI deflects that ticket and the customer calls back twice, the "savings" are an illusion. Read more about our approach on the Lorikeet blog. ## Key Takeaways - Focus on 8 metrics, not 15+: CSAT, NPS, FCR, AHT, CES, abandonment rate,... --- ## Contact Center Benchmarks in 2026: The Numbers You Need URL: https://www.lorikeetcx.ai/articles/contact-center-benchmarks Updated: 2026-03-12 Keywords: contact center benchmarks, call center benchmarks 2026, contact center KPIs Most contact centers are still measuring themselves against benchmarks that assume every interaction is handled by a human - and that assumption is already outdated. Contact center benchmarks are standardized performance metrics - like average handle time, first contact resolution, and cost per contact - that organizations use to evaluate operational efficiency. In 2026, the benchmark landscape has split: traditional human-only metrics now sit alongside AI-augmented performance targets, with 80% of contact centers expected to use AI this year according to Gartner and McKinsey reporting that AI deployments reduce interaction volume by 40-50%. - Median cost per contact: $1.84 self-service vs $13.50 assisted (Gartner) - Average handle time for voice: 4-7 minutes - First contact resolution target: 70-85% - CSAT benchmark: 85%+ - Call abandonment healthy range: 2-5% - Global average speed of answer: 28 seconds Last updated: March 2026. Sources include Gartner, McKinsey, Natterbox Contact Center Benchmarks 2026 report, Nextiva, Dialpad, and CloudTalk benchmark data. The gap between top-performing contact centers and the rest is widening. Teams that blend AI resolution with human expertise are posting numbers that would have seemed impossible 2 years ago. Below, we break down every major benchmark for 2026 - split by human-only and AI-augmented performance - so you can see exactly where your operation stands. ## What Are the Key Contact Center Benchmarks for 2026? The key contact center benchmarks for 2026 are average handle time (4-7 minutes voice), first contact resolution (70-85%), cost per contact ($1.84 self-service, $13.50 assisted), CSAT (85%+), call abandonment (2-5%), and average speed of answer (28 seconds globally). AI-augmented centers are shifting every one of these numbers. These benchmarks come from aggregated data across Gartner, the Natterbox Contact Center Benchmarks 2026 report, and operational data from platforms like Nextiva, Dialpad, and CloudTalk. The critical shift this year is that benchmarks now need context: are you measuring a fully human queue, or an AI-augmented operation? Lorikeet is an AI customer support platform that resolves tickets end-to-end across chat, email, and voice. It uses structured workflows and real-time quality scoring through Coach to hit the AI-augmented benchmarks outlined in this article. Schedule adherence remains a human-side benchmark at 85-92%, and voice occupancy holds steady at 75-85%. These numbers have not changed much because they are fundamentally about workforce management. What has changed is how much volume reaches human agents in the first place. ## How Has Cost Per Contact Changed With AI? Cost per contact is where AI has created the most dramatic benchmark split. Gartner puts the median at $1.84 for self-service channels and $13.50 for human-assisted contacts. AI-augmented centers are resolving 40-50% of interactions before they reach a human agent, which is compressing blended cost per contact significantly. The math is straightforward. If your blended cost per contact was $10 and you shift 40% of volume to AI resolution at under $2 per interaction, your new blended cost drops below $7. That is not a marginal improvement. It is a structural change in unit economics. For a deeper breakdown, see our guide on customer service cost per ticket. Teams using Lorikeet's Resolution Loop are seeing this play out in real time. The platform identifies which tickets can be fully resolved by AI and routes accordingly, which directly lowers blended cost per contact without sacrificing resolution quality. ## What Should Your First Contact Resolution Rate Be? The 2026 benchmark for first contact resolution (FCR) is 70-85% across most call centers. Top-performing human teams hit the upper end of that range. AI-augmented operations are pushing FCR above 85% on eligible ticket types because AI does not forget to check a knowledge base or skip a troubleshooting step. FCR is arguably the single most important benchmark because it correlates directly with CSAT and cost. Every repeat contact adds cost and erodes satisfaction. The challenge has always been consistency. Human agents have good days and bad days. AI does not. That said, FCR only counts if the resolution is actually correct. This is where quality assurance matters. Lorikeet's Coach product scores every AI response against your policies in real time, catching errors before they reach the customer. This keeps FCR high without inflating it through premature ticket closures. For detailed benchmarks on the speed side of first contact, read our breakdown of first response time benchmarks. See how your benchmarks compare. Lorikeet shows you AI resolution rate, cost per contact, and quality scores in a single dashboard. Get started and benchmark your operation against these 2026 numbers. ## What Are the AI-Specific Benchmarks Contact Centers Should Track? In 2026, contact centers should track AI resolution rate, AI-assisted handle time, routing accuracy, and AI quality score alongside traditional metrics. Natterbox reports a 54% drop in "Hunting Time" from AI-powered call routing alone. These metrics did not exist 3 years ago. Now they are essential. AI resolution rate - the percentage of tickets fully resolved without human involvement - is becoming the defining metric for AI-augmented centers. There is no universal benchmark yet, but leading operations report 30-50% AI resolution across all ticket types, with some categories exceeding 80%. AI-assisted handle time measures how long a human agent takes when AI has already gathered context, pre-populated fields, or drafted a response. Early data from Dialpad and CloudTalk suggests this cuts AHT by 20-35% compared to unassisted interactions. Routing accuracy matters because a misrouted ticket is a wasted interaction. The 54% reduction in hunting time reported by Natterbox shows how much waste exists in traditional routing. AI-powered routing does not just speed things up. It fundamentally changes which agent - or which AI workflow - handles each contact. ## How Do You Benchmark a Blended Human-AI Operation? Benchmarking a blended operation requires tracking metrics at 3 levels: AI-only performance, human-only performance, and blended totals. A healthy 2026 contact center should see AI handling 30-50% of volume, human agents maintaining FCR above 75%, and blended CSAT at 85% or higher. The mistake most teams make is averaging everything together. If your AI resolves 40% of tickets with a 92% CSAT and your human agents handle the remaining 60% at 80% CSAT, your blended number is 85%. That looks fine. But it hides the fact that your human-handled tickets need attention. Lorikeet surfaces these splits automatically. You see AI and human performance side by side, so you know exactly where to invest. Learn more about reducing customer service costs through this kind of targeted optimization. ## Lorikeet's Take on 2026 Benchmarks At Lorikeet, the biggest shift in 2026 is not any single benchmark moving. It is that the benchmark framework itself has split. Human-only benchmarks still matter for the work humans do. But measuring your entire operation against human-only standards misses the full picture. We see teams struggle most when they try to retrofit old KPIs onto AI-augmented operations. Cost per contact drops but nobody tracks AI resolution quality. AHT improves but nobody asks whether AI is just closing tickets faster without actually resolving them. The contact centers posting the strongest numbers in 2026 are the ones tracking both sides. They measure AI resolution rate and AI quality score alongside traditional AHT, FCR, and CSAT. They use tools like Lorikeet's Coach to ensure AI quality stays high.... --- ## First Contact Resolution Rate: Benchmarks and How to Improve FCR URL: https://www.lorikeetcx.ai/articles/first-contact-resolution-rate Updated: 2026-03-12 Keywords: first contact resolution rate, FCR benchmarks, how to improve FCR Your team resolves most tickets on the first try - but "most" still means 1 in 4 customers need to come back, costing you revenue and trust. First contact resolution (FCR) is the percentage of customer issues resolved during the initial interaction, without requiring a follow-up. In 2026, the cross-industry average sits at 70%, with top performers reaching 80-85% according to SQM Group. - Every 1% improvement in FCR correlates with a 1% improvement in CSAT (SQM Group) - 96% of customers who experience high-effort interactions become disloyal (Gartner) - AI deployments reduce total service interactions by 40-50% (McKinsey) - AI-native platforms now achieve 55-70% autonomous first contact resolution Last updated: March 2026 First contact resolution rate has been the gold standard for measuring support efficiency for decades. But the definition is shifting. Traditional FCR tracked whether a human agent solved the problem in a single call or chat. Today, AI-native platforms like Lorikeet resolve issues autonomously - meaning the customer never reaches a human at all. That changes what "first contact" means and how you should benchmark it. This guide breaks down current FCR benchmarks by industry, explains how AI is redefining the metric, and gives you 5 concrete ways to improve your rate. ## What Is First Contact Resolution Rate and Why Does It Matter? First contact resolution rate measures the percentage of customer issues fully resolved during the first interaction, with no callbacks, transfers, or follow-up tickets required. It matters because it directly impacts customer satisfaction, operational cost, and loyalty - making it one of the most reliable predictors of support quality. The math is simple: FCR = (issues resolved on first contact / total issues) x 100. A team handling 1,000 tickets that resolves 720 on the first touch has a 72% FCR. First Contact Resolution (FCR): The percentage of customer issues fully resolved in the first interaction without callbacks, transfers, or follow-up tickets. The financial impact is significant. According to Supportbench, improving FCR reduces repeat contacts, lowers cost per ticket, and increases customer lifetime value. SQM Group's research shows that every 1% gain in FCR produces a corresponding 1% gain in CSAT. Lorikeet is an AI customer support platform purpose-built for complex service environments. Unlike traditional chatbots that deflect tickets, Lorikeet's Resolution Loop resolves issues end-to-end - checking orders, processing returns, updating accounts - autonomously and on first contact. ## What Are Good FCR Benchmarks by Industry in 2026? Good FCR benchmarks range from 65% to 85% depending on industry, with the cross-industry average at 70% according to SQM Group. E-commerce leads at 75-85%, while complex regulated industries like healthcare sit closer to 71%. Top-performing teams across all sectors consistently exceed 80%. - E-commerce / Retail: 75-85%. Higher rates driven by standardized issues like order tracking and returns. - Healthcare: approximately 71%. Regulatory complexity and multi-system workflows drag down resolution speed. - Financial Services: 65-75%. Compliance requirements often force escalations even for routine requests. - SaaS / Technology: 65-75%. Technical troubleshooting varies widely in complexity. - General cross-industry: 70% average, with top performers at 80-85%. Context matters more than the number. A 72% FCR in healthcare may represent stronger performance than an 80% in retail, given the relative complexity of each interaction. ## How Does AI Change the Way We Measure FCR? AI changes FCR measurement by introducing autonomous resolution - issues solved without any human involvement. Traditional FCR only counted whether a human agent resolved the issue in one interaction. AI-native platforms now resolve 55-70% of contacts autonomously, creating a new tier of performance that legacy metrics were never designed to capture. This distinction matters. Traditional chatbots achieve just 10-25% resolution rates because they deflect rather than resolve. They route customers to FAQ pages or force them to rephrase until they give up. That is not resolution - it is abandonment mislabeled. Platforms like Lorikeet take a different approach. The Resolution Loop executes multi-step workflows - verifying identity, checking order status, initiating refunds - without handing off to a human. McKinsey's data showing AI deployments reduce service interactions by 40-50% reflects this shift. The new framework requires 2 FCR metrics: human-assisted FCR (traditional) and autonomous FCR (AI-resolved without escalation). Tracking both gives you an accurate picture of total support performance. Want to see what autonomous first contact resolution looks like? Get started with Lorikeet and see how AI-native resolution compares to your current FCR benchmarks. ## What Are 5 Proven Ways to Improve First Contact Resolution? The 5 most effective ways to improve FCR are: deploying AI for routine resolution, improving knowledge base quality, reducing transfers, empowering agents with better tools, and analyzing repeat contact patterns. Teams that combine these approaches typically see 10-15 percentage point improvements within 6 months. - Deploy AI that resolves, not just deflects. Traditional chatbots push customers to self-service. AI-native tools like Lorikeet resolve issues end-to-end. The difference between 10-25% chatbot resolution and 55-70% autonomous resolution is the difference between deflection and actual FCR improvement. - Fix your knowledge base. Agents and AI systems both rely on accurate, current documentation. Audit your articles monthly. Use Lorikeet's Coach to identify knowledge gaps based on real ticket patterns. - Reduce unnecessary transfers. Every transfer drops FCR. Map your escalation paths and eliminate handoffs that exist because of policy, not necessity. Give frontline agents (and AI) the permissions to resolve more issue types directly. - Improve agent tooling and access. Agents who need to check 4 systems to answer 1 question will struggle with FCR. Consolidated dashboards and integrated workflows reduce resolution time and improve first response time. - Analyze repeat contact reasons. Pull your top 10 reasons customers contact you a second time. Each one is a specific, fixable failure. Address them in order of volume and you will see measurable FCR gains within weeks. ## How Does FCR Connect to CSAT, Cost, and Loyalty? FCR directly drives 3 core business metrics: customer satisfaction rises 1% for every 1% FCR improvement (SQM Group), cost per ticket drops as repeat contacts decrease, and loyalty improves because 96% of high-effort customers become disloyal (Gartner). FCR is the single metric that moves all 3 simultaneously. The cost impact compounds. A repeat contact does not just double the ticket cost - it also increases handle time on the second interaction because the customer is frustrated and the agent needs to review prior context. Improving FCR from 70% to 80% on 10,000 monthly tickets eliminates 1,000 repeat contacts. If your average cost per ticket is $15, that is $15,000 in monthly savings from a 10-point FCR improvement. The CSAT uplift and reduced churn add further value that is harder to quantify but often larger. ## Lorikeet's Take on First Contact Resolution At Lorikeet, we believe the real question is not "did the agent resolve it on the first call?" but "did the customer have to do anything more than ask once?" That reframe is important. A customer who messages once and gets an automated resolution in 30 seconds has a better experience than one who waits 4 minutes for a human to do the same thing. Traditional chatbots muddied FCR by claiming high "containment" rates that were really abandonment. Lorikeet's Resolution Loop measures actual resolution -... --- ## AI Customer Service Statistics: 30 Data Points for 2026 URL: https://www.lorikeetcx.ai/articles/ai-customer-service-statistics Updated: 2026-03-12 Keywords: ai customer service statistics, AI support data 2026, AI customer service market size Contact centers spent $80 billion on labor last year, yet only 14% of customer issues actually resolve through self-service - here is what the data says about closing that gap. AI customer service statistics are the measurable data points - market size, adoption rates, resolution percentages, cost benchmarks, and consumer sentiment scores - that track how artificial intelligence is reshaping support operations in 2026. The global AI customer service market is projected to reach $15.12 billion this year, with 88% of contact centers already using some form of AI. - 30 verified statistics across 6 categories with named sources - Clear distinction between AI-assisted and AI-resolved metrics - Cost benchmarks: $1.84 self-service vs $13.50 assisted per contact (Gartner) - AI-native platforms achieve 55-70% FCR at under $3 per resolution Last updated: March 2026 The gap between AI adoption and AI resolution is the most important story in customer service right now. Nearly 9 in 10 contact centers report using AI in some capacity. But only 25% have fully integrated automation into daily operations. That difference explains why so many teams feel stuck - they have the tools but not the outcomes. This article organizes 30 data points into categories that actually matter: market growth, adoption rates, cost savings, performance benchmarks, consumer preferences, and what AI-native platforms deliver differently. Where possible, we distinguish between AI-assisted metrics (AI helped a human) and AI-resolved metrics (AI handled it end to end). ## How Big Is the AI Customer Service Market in 2026? The global AI customer service market is projected to reach $15.12 billion in 2026, growing at a 25.8% compound annual growth rate. By 2034, that figure is expected to hit $117.87 billion. This makes AI-powered support one of the fastest-growing segments in enterprise software. Lorikeet is an AI customer support platform that resolves tickets end-to-end - not just deflects them. Its Resolution Loop handles tier-1 and tier-2 issues autonomously, while Coach assists human agents on complex cases. - $15.12 billion - projected global AI customer service market size in 2026 - $117.87 billion - projected market size by 2034 - 25.8% - compound annual growth rate from 2026 to 2034 - $80 billion - projected reduction in contact center labor costs enabled by AI by 2026 Gartner projected that 80% of customer service organizations would use generative AI by 2025. Current data suggests that timeline was roughly accurate, with adoption now moving from experimentation to integration. For a deeper look at how these tools work in practice, see our guide to AI in customer service. ## What Are the Real AI Adoption Rates Across Industries? 88% of contact centers report using some form of AI, but only 25% have fully integrated automation into daily operations. The difference between "using AI" and "deploying AI at scale" is where most organizations stall. Adoption varies significantly by vertical. - Telecom - 95% AI adoption rate, the highest of any vertical - Banking and finance - 92% adoption rate - 88% - overall percentage of contact centers using some form of AI - 25% - percentage with fully integrated AI automation - 95% - projected share of all customer interactions AI will handle by end of 2026 That last number - 95% of interactions - needs context. "Handle" does not mean "resolve." AI can route, triage, summarize, and assist without ever completing a resolution autonomously. This is the core distinction teams miss when evaluating their AI performance. Learn more about the difference in our comparison of AI agents vs chatbots. ## How Much Does AI Actually Reduce Customer Service Costs? Gartner benchmarks the median cost per contact at $1.84 for self-service versus $13.50 for agent-assisted interactions - a 7x difference. AI-native platforms like Lorikeet operate in the $1 to $3 per resolution range, delivering full outcomes at a fraction of traditional costs. - $1.84 - Gartner median cost per self-service contact - $13.50 - Gartner median cost per agent-assisted contact - $1-3 - cost per resolution on AI-native platforms - $80 billion - total contact center labor cost reduction projected by 2026 - 40-50% - interaction reduction reported by McKinsey for AI deployments The cost per ticket breakdown matters more than the headline numbers. A $1.84 self-service contact that fails to resolve still generates a $13.50 follow-up. For a full analysis of these economics, see our article on customer service cost per ticket. ## What Do AI Resolution and Performance Metrics Actually Look Like? Companies using AI for tier-1 support resolve 65% of issues without human intervention. AI-native platforms achieve 55-70% first contact resolution rates with average handle times under 3 minutes. But Gartner found only 14% of issues fully resolve through traditional self-service channels. - 65% - tier-1 issues resolved without human intervention using AI - 55-70% - first contact resolution rate on AI-native platforms - Under 3 minutes - average handle time on AI-native platforms - 14% - percentage of issues fully resolved through traditional self-service (Gartner) - 92% - businesses reporting improved CSAT after implementing AI The gap between 14% self-service resolution and 65% AI-native resolution tells the real story. Legacy chatbots deflect. Platforms like Lorikeet's Resolution Loop actually complete the task - process the refund, update the order, change the subscription. Average handle time drops because the AI does the work, not because it routes faster. Want to see what 55-70% FCR looks like for your team? Get started with Lorikeet and benchmark your resolution rate in weeks, not quarters. ## What Do Customers Actually Prefer - AI or Human Agents? 79% of Americans still prefer interacting with humans over AI for customer service. But 51% prefer bots when they want immediate service. Consumer sentiment is not anti-AI - it is anti-bad-AI. Speed and accuracy change the equation. - 79% - Americans who prefer human agents over AI - 51% - consumers who prefer bots when they want immediate service - 92% - businesses reporting improved customer satisfaction after AI implementation These numbers are not contradictory. Customers prefer humans when they expect the AI will fail. When AI resolves the issue quickly and completely, preference shifts. The 92% CSAT improvement stat reflects what happens when AI works well. Lorikeet's Voice product handles phone-based support with the same resolution-first approach. Customers calling in get outcomes, not hold music followed by a transfer. For a broader look at the tools available, see our roundup of AI tools for customer service automation. ## Lorikeet's Take on AI Customer Service Statistics At Lorikeet, the industry tracks AI adoption but underweights AI resolution. 88% adoption means nothing when only 14% of self-service interactions actually resolve. Lorikeet focuses on the metric that matters: did the customer's problem get solved without escalation? Three patterns stand out in this data. First, cost savings only materialize when AI resolves - not when it deflects to a cheaper channel that still fails. Second, average handle time under 3 minutes is achievable but only when the AI has access to backend systems to take action. Third, the 79% human preference stat will erode as AI-native platforms close the quality gap. Read more about the broader landscape in our overview of AI for customer support. ## Key Takeaways - The AI customer service... --- --- # Glossary ## Account Takeover Rate URL: https://www.lorikeetcx.ai/glossary/account-takeover-rate **Learn what account takeover rate measures in customer service, how to track it, and strategies to reduce unauthorized access to protect your customers.** **Account takeover rate** measures the frequency of unauthorized account access incidents, typically expressed as a percentage of total active accounts or as incidents per 10,000 accounts. Account takeovers (ATOs) are among the most damaging customer experiences in fintech. Unlike a simple service hiccup, an ATO means someone's financial identity has been compromised. The support interaction that follows—account recovery, transaction review, reimbursement—determines whether that customer ever trusts you again. Prevention is the primary defense: strong authentication, device fingerprinting, behavioral analytics, and transaction velocity limits. But CX teams own the detection and response layers. Train agents to recognize ATO indicators: password changes followed immediately by contact info changes, transactions from new locations, customers calling confused about activity they didn't initiate. Your ATO response playbook matters as much as your prevention tech. How fast can you lock an account? How quickly can you reverse fraudulent transactions? Do you have a dedicated escalation path for ATO cases? The best teams resolve ATO incidents in under 24 hours with full customer restoration. Track your ATO-to-resolution time as seriously as you track the rate itself. **Related terms:** [Chargeback rate](https://www.lorikeetcx.ai/glossary/chargeback-rate), [KYC/KYB completion rate](https://www.lorikeetcx.ai/glossary/kyc-kyb-completion-rate), [Security incident response](https://www.lorikeetcx.ai/glossary/security-incident-response) ## Agent assist URL: https://www.lorikeetcx.ai/glossary/agent-assist **Discover what agent assist is, how AI-powered agent assist tools work in real time, and why they help support teams resolve issues faster.** Agent assist is technology that supports human customer service agents in real-time during customer interactions. Rather than replacing the agent, it provides contextual suggestions — relevant knowledge articles, recommended responses, customer history summaries, or next-best-action prompts — to help agents resolve issues faster and more consistently. Agent assist tools typically work by monitoring the conversation in real-time, identifying the customer's intent, and surfacing relevant information from internal knowledge bases, CRM systems, or past interactions. More advanced implementations use large language models to draft responses that the agent can review and send. The value proposition is straightforward: reduce average handling time, improve consistency across agents, and flatten the learning curve for new hires. For regulated industries, agent assist can also surface compliance requirements and required disclosures based on the conversation context. However, agent assist represents a fundamentally different architecture than AI agents that handle conversations autonomously. It's an augmentation play, not an automation play. The economics are different — you still need the same headcount, you're just making each person more productive. For organizations processing thousands of tickets monthly, the math on full AI resolution (where the AI handles the conversation end-to-end) typically looks more compelling than the incremental productivity gains of agent assist. Many teams start with agent assist as a stepping stone toward full AI automation, using it to build confidence in AI-generated responses before transitioning to autonomous handling. **Related terms:** [AI agent](https://www.lorikeetcx.ai/glossary/ai-agent), [knowledge base](https://www.lorikeetcx.ai/glossary/knowledge-base), [average handling time](https://www.lorikeetcx.ai/glossary/average-handling-time) ## Agent Handle Time URL: https://www.lorikeetcx.ai/glossary/agent-handle-time **Understand agent handle time (AHT), how it differs from average handling time, and proven methods to optimize it without sacrificing service quality.** **Agent handle time (AHT)** measures the total time a human support agent spends actively working on a ticket, from first touch to resolution or escalation. In hybrid AI-human support environments, AHT becomes a more nuanced metric. When AI handles routine queries and escalates only the complex cases, human AHT naturally increases—agents are now dealing exclusively with harder problems. This isn't a failure; it's the intended outcome. The meaningful comparison shifts from "average AHT across all tickets" to "AHT by complexity tier." A well-tuned AI system should reduce overall AHT while increasing human AHT, because humans are only touching tickets that genuinely require human judgment. If your AI is escalating easy tickets, your human AHT will look artificially low—and you're wasting agent capacity. Smart teams track AHT alongside resolution rate and customer satisfaction. Handle time without context is dangerous: rushing agents to hit AHT targets often trades short-term efficiency for long-term churn. **Related terms:** [Warm Handoff](https://www.lorikeetcx.ai/glossary/warm-handoff), [Ticket Complexity](https://www.lorikeetcx.ai/glossary/ticket-complexity), [Credit Consumption](https://www.lorikeetcx.ai/glossary/credit-consumption) ## Agent Occupancy URL: https://www.lorikeetcx.ai/glossary/agent-occupancy **Learn what agent occupancy measures, how to calculate it, and why balancing occupancy rates is critical for agent productivity and burnout prevention.** **Agent occupancy** is the percentage of an agent's logged-in time spent actively handling customer interactions or performing directly productive work, versus waiting for contacts. Occupancy answers: "When agents are scheduled to work, how much of that time involves actual work?" An occupancy of 75% means agents spend 45 minutes of every hour on calls, chats, or tickets, and 15 minutes waiting. The remaining time is productive capacity you're paying for but not using. High occupancy (85%+) sounds efficient but causes burnout and quality degradation—agents need breathing room between intense interactions. Low occupancy (<65%) signals overstaffing or poor forecasting. The sweet spot for most contact centers is 75-85%, though this varies by channel and interaction complexity. Insurance claims agents handling emotionally difficult calls need lower occupancy targets than agents answering basic fintech account questions. Balance occupancy against quality scores and attrition rates to find your sustainable level. **Related terms:** [Ticket volume](https://www.lorikeetcx.ai/glossary/ticket-volume), [Average speed of answer](https://www.lorikeetcx.ai/glossary/average-speed-of-answer), [Agent quality score](https://www.lorikeetcx.ai/glossary/agent-quality-score) ## Agent Quality Score URL: https://www.lorikeetcx.ai/glossary/agent-quality-score **Discover what agent quality score is, how it is calculated across interactions, and how to use it to improve customer service performance.** **Agent quality score** is an individual agent's aggregated performance rating across QA evaluations, customer feedback, and operational metrics over a defined period. Where QA score measures interaction quality, agent quality score measures agent quality—rolling up multiple signals into a single view of performance. This typically combines QA audit results, CSAT/CES feedback tagged to the agent, adherence metrics, and sometimes peer or supervisor assessments. The weighting depends on what you're optimizing for. Use agent quality scores for coaching, not just ranking. A score without diagnosis is useless—break it down into components so agents know what to improve. For regulated industries, include compliance-specific criteria: did the agent follow required disclosures, verify identity properly, document the interaction correctly? Also watch for gaming. Agents optimizing for score might cherry-pick easy tickets, rush to close, or pressure customers for good ratings. Design the score to reflect behaviors you actually want, and audit for manipulation. **Related terms:** [Quality Assurance score](https://www.lorikeetcx.ai/glossary/quality-assurance-score), [Agent occupancy](https://www.lorikeetcx.ai/glossary/agent-occupancy), [Escalation rate](https://www.lorikeetcx.ai/glossary/escalation-rate) ## Agentic AI URL: https://www.lorikeetcx.ai/glossary/agentic-ai **Learn what agentic AI is, how autonomous AI agents make decisions and take actions, and why agentic AI is transforming customer experience.** Agentic AI refers to artificial intelligence systems that can autonomously plan, reason, and take actions to achieve goals — rather than simply responding to individual prompts. In customer service, agentic AI systems can decompose a complex customer request into steps, decide which actions to take (checking an account, issuing a refund, updating a record), execute those actions across multiple backend systems, and verify the outcome — all within a single interaction. This is a meaningful departure from earlier generations of AI in customer service, which could classify intent or suggest responses but couldn't actually do anything. An agentic system doesn't just understand that a customer wants to reschedule a shipment — it checks inventory, finds available dates, updates the order, and confirms the change. The distinction matters because it shifts AI from a deflection tool (intercepting simple queries before they reach humans) to a resolution tool (completing the actual work the customer needs done). Agentic systems can handle multi-step processes, maintain context across a conversation, and make decisions based on business rules. That said, agentic AI in customer service raises important questions about guardrails, auditability, and control. An autonomous system taking real actions on customer accounts needs clear boundaries around what it can and cannot do, full logging of its reasoning, and the ability for human teams to configure and override its behavior. The more capable the system, the more critical the governance. **Related terms:** [AI agent](https://www.lorikeetcx.ai/glossary/ai-agent), [AI guardrails](https://www.lorikeetcx.ai/glossary/ai-guardrails), [AI audit trail](https://www.lorikeetcx.ai/glossary/ai-audit-trail), [human-in-the-loop](https://www.lorikeetcx.ai/glossary/human-in-the-loop) ## AI agent URL: https://www.lorikeetcx.ai/glossary/ai-agent **Understand what an AI agent is, how AI agents automate customer service workflows, and the key capabilities that define modern AI-powered support.** An AI agent is a software system that can autonomously handle customer interactions from start to finish — understanding the customer's request, reasoning about the best course of action, taking actions in backend systems, and resolving the issue without human intervention. Unlike chatbots that follow scripted decision trees, AI agents use large language models to understand natural language and navigate complex, multi-turn conversations. In customer service, AI agents represent a shift from deflection (routing customers away from human agents) to resolution (actually completing the work). A well-built AI agent can authenticate a customer, look up their account, process a return, issue a credit, and send a confirmation — all in a single conversation. The capabilities that differentiate AI agents from earlier automation include: - **Multi-step reasoning:** Breaking a complex request into sequential actions - **System integration:** Reading from and writing to CRM, billing, inventory, and other backend systems - **Context maintenance:** Tracking the full conversation history and customer state across turns - **Policy adherence:** Following business rules and regulatory requirements while handling edge cases The critical question for any AI agent deployment is not whether it can handle the easy cases — most can — but how it handles the hard 20%: exceptions, edge cases, ambiguous requests, and situations requiring judgment. This is where the difference between AI agent vendors becomes apparent, and where organizations in complex or regulated industries need to evaluate carefully. Equally important is transparency. When an AI agent takes actions on customer accounts, the business needs to understand exactly why each decision was made and have full auditability of the interaction. **Related terms:** [Agentic AI](https://www.lorikeetcx.ai/glossary/agentic-ai), [resolution rate](https://www.lorikeetcx.ai/glossary/resolution-rate), [AI guardrails](https://www.lorikeetcx.ai/glossary/ai-guardrails), [AI audit trail](https://www.lorikeetcx.ai/glossary/ai-audit-trail) ## AI Agent Handoff URL: https://www.lorikeetcx.ai/glossary/ai-agent-handoff **Learn what AI agent handoff is, how AI transfers conversations to human agents with full context, and best practices for seamless escalation.** **AI agent handoff** is the process of transferring a customer conversation from an AI system to a human agent, including the passing of context, conversation history, and relevant customer information. Handoff is where many AI implementations fail the customer. Poor handoffs dump customers to agents without context, forcing repetition of information already provided. Good handoffs transfer full conversation history, customer identification, issue classification, and attempted resolution—everything the agent needs to continue seamlessly. Effective handoff design requires decisions about: when to handoff (confidence thresholds, explicit requests, escalation triggers, sensitive topics), where to route (skill-based routing, priority queuing, relationship preservation), and what to transfer (conversation transcript, classified intent, customer data, suggested resolution). The handoff itself should be invisible to the customer or explicitly acknowledged ("I'm connecting you with a specialist who can help further"). For support operations, handoff rate and handoff quality are critical metrics. High handoff rates suggest the AI can't handle common issues. Poor handoff quality (agents asking customers to repeat themselves) erodes trust in the overall system. The goal: AI resolves what it can, hands off smoothly when it can't, and sets agents up for success rather than creating frustration for customers and staff alike. **Related terms:** [Human-in-the-loop](https://www.lorikeetcx.ai/glossary/human-in-the-loop), [AI agent orchestration](https://www.lorikeetcx.ai/glossary/ai-agent-orchestration), [Intelligent Virtual Agent](https://www.lorikeetcx.ai/glossary/intelligent-virtual-agent) --- # CX Metrics & KPIs ## AI Agent Memory URL: https://www.lorikeetcx.ai/glossary/ai-agent-memory **Discover how AI agent memory works, why persistent context across conversations matters, and how memory improves personalized customer experiences.** **AI agent memory** is the capability of an AI system to retain and utilize information from past interactions—both within a single conversation and across multiple sessions—to provide contextually relevant, personalized responses. Memory transforms AI from a stateless responder to a relationship-aware agent. Within a conversation, memory means the AI recalls what was discussed earlier: "As I mentioned, your order shipped yesterday." Across conversations, memory means recognizing returning customers: "I see you contacted us about this last week—let me check the status." Implementing memory involves tradeoffs. Longer memory windows improve context but increase inference cost and latency. Cross-session memory improves personalization but raises privacy considerations—customers may not expect the AI to remember previous conversations. Memory scope decisions should be explicit and, ideally, customer-controllable. For complex support interactions, memory is essential. A multi-step troubleshooting flow breaks down if the AI forgets what the customer already tried. Customer expectations are shaped by human interactions where memory is assumed—AI without memory feels broken, not just limited. **Related terms:** [Context window](https://www.lorikeetcx.ai/glossary/context-window), [Multi-turn conversation](https://www.lorikeetcx.ai/glossary/multi-turn-conversation), [AI personalization](https://www.lorikeetcx.ai/glossary/ai-personalization) ## AI Agent Orchestration URL: https://www.lorikeetcx.ai/glossary/ai-agent-orchestration **Learn what AI agent orchestration is, how multiple AI agents coordinate to resolve complex queries, and why orchestration drives CX efficiency.** **AI agent orchestration** is the coordination layer that manages multiple AI capabilities, tools, and workflows to handle complex customer requests that require sequential or parallel operations across different systems. A customer asking "cancel my subscription, refund the last charge, and confirm my data will be deleted" triggers multiple distinct operations. Orchestration decomposes this into tasks (cancel, refund, data deletion), determines dependencies and sequencing, dispatches each to the appropriate system or capability, handles failures, and assembles a coherent response. Orchestration complexity increases with the scope of what AI can do. Simple chatbots don't need orchestration—they just respond. AI agents that access databases, execute transactions, and integrate with external systems require sophisticated orchestration to manage state, handle errors, and ensure reliability. For CX teams evaluating AI platforms, orchestration capability determines what's possible. Can the AI handle requests requiring multiple backend actions? Can it recover from partial failures? Can it maintain coherence when customer requests change mid-conversation? These aren't model capabilities—they're orchestration capabilities. **Related terms:** [AI workflow automation](https://www.lorikeetcx.ai/glossary/ai-workflow-automation), [AI agent handoff](https://www.lorikeetcx.ai/glossary/ai-agent-handoff), [Next-best action](https://www.lorikeetcx.ai/glossary/next-best-action) ## AI audit trail URL: https://www.lorikeetcx.ai/glossary/ai-audit-trail **Understand what an AI audit trail is, how it logs AI decisions and actions, and why auditability is essential for compliance and trust.** An AI audit trail is a complete, human-readable log of every decision an AI system makes during a customer interaction — what information it accessed, what reasoning it applied, what actions it took, and why. In regulated industries like financial services, healthcare, and insurance, audit trails are not optional: they're a compliance requirement. Traditional customer service channels produce audit trails naturally. A human agent's actions are logged in the ticketing system, their notes explain their reasoning, and their conversation is recorded. When AI handles conversations autonomously, that same level of traceability needs to exist — arguably more so, because AI decisions happen faster and at higher volume. A robust AI audit trail should include: - **Decision reasoning:** Not just what the AI did, but why — which policies it applied, how it interpreted the customer's request, what alternatives it considered - **Data accessed:** Which customer records, knowledge articles, and system data informed the decision - **Actions taken:** Every backend action (refund issued, account updated, ticket escalated) with timestamps - **Confidence signals:** Where the AI was certain vs. where it was uncertain and how it handled ambiguity This matters beyond compliance. Audit trails are how CX teams debug issues, improve AI performance, and build trust with internal stakeholders. Without them, AI is a black box — the team knows what happened but not why, making it impossible to improve systematically. Organizations evaluating AI for customer service should treat audit trail depth as a primary evaluation criterion, not a checkbox. **Related terms:** [AI compliance](https://www.lorikeetcx.ai/glossary/ai-compliance), [AI guardrails](https://www.lorikeetcx.ai/glossary/ai-guardrails), [AI observability](https://www.lorikeetcx.ai/glossary/ai-observability), [quality assurance in customer service](https://www.lorikeetcx.ai/glossary/quality-assurance-in-customer-service) ## AI compliance URL: https://www.lorikeetcx.ai/glossary/ai-compliance **Learn what AI compliance means for customer service, key regulations and frameworks, and how to ensure your AI systems meet compliance standards.** AI compliance refers to the set of practices, controls, and documentation that ensure AI systems meet regulatory requirements, industry standards, and internal governance policies. In customer service, this is especially critical for industries where interactions involve sensitive data, financial transactions, medical information, or insurance claims. Key compliance considerations for AI in customer service include: - **Data handling:** How customer data is processed, stored, and retained by AI systems. Regulations like GDPR, HIPAA, and SOC 2 impose specific requirements. - **Decision transparency:** Regulators increasingly require that automated decisions be explainable. When AI denies a claim, changes an account, or provides medical information, the reasoning must be auditable. - **Consent and disclosure:** Many jurisdictions require businesses to disclose when customers are interacting with AI rather than a human. - **Accuracy obligations:** In financial services and healthcare, providing incorrect information can create regulatory liability. AI systems need guardrails to prevent hallucination on compliance-sensitive topics. - **Record retention:** Customer interactions must be retained for prescribed periods, with full audit trails of AI decisions. A common mistake is treating compliance as a post-deployment checkbox. The more effective approach is building compliance into the AI system's architecture — deterministic guardrails for sensitive actions, mandatory escalation paths for regulated scenarios, and continuous quality assurance across 100% of interactions (not just a sample). Organizations in regulated industries should evaluate AI vendors on their compliance infrastructure, not just their automation rate. A high resolution rate is meaningless if it creates regulatory exposure. **Related terms:** [AI audit trail](https://www.lorikeetcx.ai/glossary/ai-audit-trail), [AI guardrails](https://www.lorikeetcx.ai/glossary/ai-guardrails), [AI hallucinations](https://www.lorikeetcx.ai/glossary/ai-hallucinations) ## AI concierge URL: https://www.lorikeetcx.ai/glossary/ai-concierge **Discover what an AI concierge is, how it delivers personalized, proactive service, and why AI concierge experiences drive customer loyalty.** An AI concierge is an AI system designed to serve customers across their entire relationship with a business — not just reactive support tickets, but acquisition, onboarding, growth, and retention interactions. The term distinguishes a full-lifecycle approach from narrower applications like support chatbots or FAQ deflection tools. Where a traditional AI support agent waits for inbound tickets, an AI concierge proactively reaches out: following up on abandoned carts, checking in after onboarding, surfacing relevant product features based on usage patterns, or intervening before a customer churns. It maintains memory of the full customer relationship and uses that context to personalize every interaction. The concierge model reflects how customers actually experience a business. Customers don't think in terms of "support" vs. "sales" vs. "success" — they have one relationship with the brand and expect it to be coherent. An AI concierge serves that entire relationship rather than operating in a single departmental silo. This approach is particularly valuable for businesses with complex customer lifecycles: financial services companies where customers hold multiple products, healthcare platforms managing ongoing patient relationships, or subscription businesses where retention depends on proactive engagement. The key architectural requirements for an AI concierge differ from a support-only AI agent. It needs cross-system memory (not just ticket history), multi-channel capability (not just chat), outbound as well as inbound interaction, and the ability to operate across different business functions with different rules and objectives. **Related terms:** [AI agent](https://www.lorikeetcx.ai/glossary/ai-agent), [customer journey mapping](https://www.lorikeetcx.ai/glossary/customer-journey-mapping), [customer lifetime value](https://www.lorikeetcx.ai/glossary/customer-lifetime-value), [proactive customer service](https://www.lorikeetcx.ai/glossary/proactive-customer-service) ## AI Grounding URL: https://www.lorikeetcx.ai/glossary/ai-grounding **Learn what AI grounding is, how grounding techniques anchor AI responses in factual data, and why it reduces hallucinations in customer service.** **AI grounding** is the practice of connecting AI-generated responses to verified source information—knowledge bases, policies, customer data—rather than relying on the model's training data alone. Grounding is the primary defense against hallucination. An ungrounded AI generates responses from its parametric knowledge—what it learned during training, which may be outdated, incomplete, or wrong for your specific business. A grounded AI generates responses based on retrieved, verified information that you control. The most common grounding architecture is retrieval-augmented generation (RAG), where relevant documents are retrieved and provided to the model as context. But grounding extends beyond document retrieval: grounding the AI in real-time customer data, in current policy versions, in live system states. Each grounding source adds accuracy for its domain. Grounding is necessary but not sufficient for accuracy. The AI can still misinterpret retrieved information, fail to surface the right documents, or confuse context when multiple sources conflict. Grounding reduces hallucination risk; monitoring and guardrails address the remaining risk. **Related terms:** [Retrieval augmented generation](https://www.lorikeetcx.ai/glossary/retrieval-augmented-generation), [Knowledge base](https://www.lorikeetcx.ai/glossary/knowledge-base), [Hallucination detection](https://www.lorikeetcx.ai/glossary/hallucination-detection) ## AI guardrails URL: https://www.lorikeetcx.ai/glossary/ai-guardrails **Understand what AI guardrails are, how they constrain AI behavior within safe boundaries, and best practices for implementing guardrails in CX.** AI guardrails are the constraints, rules, and safety mechanisms that govern what an AI system can and cannot do during customer interactions. They prevent AI from taking inappropriate actions, generating harmful content, or operating outside defined boundaries — particularly important when AI is handling real customer accounts in regulated industries. Guardrails operate at multiple levels: - **Action-level:** Defining which backend actions the AI is authorized to perform (e.g., it can issue refunds up to $50 but must escalate above that amount) - **Content-level:** Preventing the AI from generating responses on prohibited topics, making unauthorized promises, or providing inaccurate compliance-sensitive information - **Conversation-level:** Detecting when a conversation has moved outside the AI's competence and triggering escalation to a human - **Business logic:** Enforcing rules like "always verify identity before accessing account details" or "never modify a policy without confirmation" The most effective guardrails combine deterministic rules (hard constraints that cannot be overridden) with AI-powered judgment (detecting nuanced situations that require escalation). A purely rule-based approach breaks on edge cases; a purely AI-driven approach introduces unacceptable risk for sensitive operations. Guardrails should not be confused with limitations. A well-guardrailed AI agent can still handle complex, multi-step interactions — it just does so within clearly defined boundaries. The goal is confidence and control, not restriction. For CX leaders evaluating AI platforms, the sophistication of the guardrail system is often a better indicator of production readiness than headline automation rates. High automation without strong guardrails is a liability, not an achievement. **Related terms:** [AI compliance](https://www.lorikeetcx.ai/glossary/ai-compliance), [AI audit trail](https://www.lorikeetcx.ai/glossary/ai-audit-trail), [human-in-the-loop](https://www.lorikeetcx.ai/glossary/human-in-the-loop), [AI hallucinations](https://www.lorikeetcx.ai/glossary/ai-hallucinations) ## AI hallucinations URL: https://www.lorikeetcx.ai/glossary/ai-hallucinations **Learn what AI hallucinations are, why large language models generate false information, and proven strategies to detect and prevent them.** AI hallucinations occur when an AI system generates information that is factually incorrect, fabricated, or unsupported by its training data or knowledge base. In customer service, hallucinations can range from minor inaccuracies (citing a policy that doesn't exist) to serious errors (providing incorrect medical dosage information or misquoting financial terms). Hallucinations are a fundamental property of large language models, not a bug that can be fully patched. LLMs generate text by predicting probable next tokens — they don't have a truth-verification mechanism. This means any customer-facing AI system needs architectural safeguards against hallucination, not just better prompts. Effective hallucination mitigation strategies include: - **Retrieval-augmented generation (RAG):** Grounding AI responses in verified source documents rather than relying on the model's parametric knowledge - **Deterministic logic for critical paths:** Using rule-based systems for actions where accuracy is non-negotiable (financial calculations, dosage information, regulatory disclosures) - **Confidence thresholds:** Escalating to human agents when the AI's confidence in its response falls below a defined threshold - **Automated QA:** Reviewing 100% of AI-handled interactions for accuracy, not just a statistical sample - **Source citation:** Requiring the AI to cite the specific knowledge base article or policy it's referencing, making verification possible In regulated industries, the risk calculus around hallucination is different from general consumer applications. A hallucinated restaurant recommendation is an inconvenience; a hallucinated insurance coverage answer is a compliance violation. CX teams in these industries should evaluate AI vendors on their hallucination prevention architecture, not just their reported hallucination rate. **Related terms:** [retrieval-augmented generation](https://www.lorikeetcx.ai/glossary/retrieval-augmented-generation), [AI guardrails](https://www.lorikeetcx.ai/glossary/ai-guardrails), [AI compliance](https://www.lorikeetcx.ai/glossary/ai-compliance) ## AI observability URL: https://www.lorikeetcx.ai/glossary/ai-observability **Discover what AI observability means, how to monitor AI agent performance in production, and key metrics for maintaining reliable AI systems.** AI observability is the ability to monitor, understand, and debug an AI system's behavior in production. In customer service, this means having real-time visibility into how AI agents are performing: which conversations they're handling well, where they're struggling, what errors are occurring, and how their performance is trending over time. Observability goes beyond basic metrics like resolution rate or CSAT. It includes: - **Conversation-level inspection:** The ability to review any individual AI-handled interaction, see the AI's reasoning, and understand why it made specific decisions - **Pattern detection:** Identifying systematic issues — topics the AI consistently struggles with, customer segments that escalate more frequently, or knowledge gaps that cause failures - **Performance trending:** Tracking whether AI performance is improving, degrading, or plateauing over time, and correlating changes with model updates, knowledge base edits, or workflow modifications - **Anomaly detection:** Alerting when AI behavior deviates from expected patterns, such as sudden increases in escalation rate or unusual response patterns For CX teams managing AI agents, observability is what separates a system you can confidently operate from one you hope is working. Without it, teams discover problems only when customers complain — by which time the damage is done. The practical implication: the best AI customer service platforms treat observability as a core product feature, not an add-on. CX leaders should be able to understand exactly what their AI is doing, why, and how to improve it — without needing a data science team to interpret raw logs. **Related terms:** [AI audit trail](https://www.lorikeetcx.ai/glossary/ai-audit-trail), [quality assurance in customer service](https://www.lorikeetcx.ai/glossary/quality-assurance-in-customer-service), [resolution rate](https://www.lorikeetcx.ai/glossary/resolution-rate) ## AI personalization URL: https://www.lorikeetcx.ai/glossary/ai-personalization **Learn what AI personalization is, how AI tailors customer experiences in real time, and why personalization increases satisfaction and revenue.** AI personalization in customer service is the use of AI to tailor interactions based on the individual customer's history, preferences, behavior, and context — rather than providing the same generic experience to every customer. It means treating returning customers differently from first-time contacts, VIP customers differently from standard tier, and frustrated customers differently from satisfied ones. Personalization goes beyond using the customer's name. Meaningful personalization includes: - **Context awareness:** Knowing the customer's recent interactions, open issues, and account status before they explain anything - **Communication adaptation:** Adjusting tone, detail level, and channel preferences based on past interactions - **Proactive relevance:** Surfacing information or offers that are specifically relevant to this customer's situation - **Preference memory:** Remembering how the customer likes to interact (prefers email over phone, wants detailed explanations vs. quick answers) - **Segment-specific handling:** Applying different workflows or policies based on customer tier, lifecycle stage, or risk profile AI enables personalization at scale in a way that human agents struggle to match. A human agent handling 50 conversations per day can't remember each customer's history and preferences. An AI agent accesses the full customer profile for every interaction, every time. For regulated industries, personalization must operate within compliance boundaries. A financial services AI should personalize communication style but not vary the accuracy of product information. A healthcare AI should adapt tone but not deviate from clinical guidelines. The personalization layer sits on top of the compliance layer, not alongside it. **Related terms:** [AI concierge](https://www.lorikeetcx.ai/glossary/ai-concierge), [customer journey mapping](https://www.lorikeetcx.ai/glossary/customer-journey-mapping), [sentiment analysis](https://www.lorikeetcx.ai/glossary/sentiment-analysis) ## AI Voice Agent URL: https://www.lorikeetcx.ai/glossary/ai-voice-agent **Understand what an AI voice agent is, how voice AI handles phone-based customer support, and the technology behind natural voice interactions.** **AI voice agent** is a conversational AI system that interacts with customers through spoken language over phone or voice channels, handling calls from greeting through resolution without human intervention. Voice adds complexity that text channels don't have: speech recognition accuracy, natural-sounding synthesis, managing interruptions and crosstalk, handling background noise, and maintaining natural conversation rhythm. These technical challenges mean voice agents typically achieve lower automation rates than text-based AI, but the operational leverage is substantial—phone is often the most expensive support channel. The use cases best suited for voice AI are structured, predictable interactions: appointment scheduling, order status checks, account balance inquiries, basic troubleshooting. Complex, emotionally sensitive, or highly variable calls still benefit from human handling. For CX leaders, voice AI decisions involve unique considerations: caller expectations (many still expect human answer), regulatory requirements (some jurisdictions require AI disclosure), and quality thresholds (voice quality issues are more jarring than text typos). Start with contained use cases and expand based on customer acceptance and resolution rates. **Related terms:** [Conversational IVR](https://www.lorikeetcx.ai/glossary/conversational-ivr), [Conversational AI](https://www.lorikeetcx.ai/glossary/conversational-ai), [Human-in-the-loop](https://www.lorikeetcx.ai/glossary/human-in-the-loop) ## AI Workflow Automation URL: https://www.lorikeetcx.ai/glossary/ai-workflow-automation **Learn what AI workflow automation is, how AI automates repetitive support tasks end-to-end, and the ROI of intelligent workflow automation.** **AI workflow automation** is the use of artificial intelligence to execute multi-step business processes in customer service—from reading and understanding requests, to querying systems, applying business logic, taking actions, and generating responses—without human intervention. Traditional automation required developers to hard-code every step: if customer says X, check system Y, then do Z. AI workflow automation introduces flexibility. The AI interprets requests in natural language, reasons about which steps are needed, and executes them in sequence. This handles the variance and ambiguity that broke earlier rule-based systems. For CX operations, AI workflow automation means the difference between answering questions and actually resolving issues. A customer asking "where's my order and can you change the address?" triggers a workflow that checks order status, identifies whether the order has shipped, evaluates whether address change is possible, executes the change if so, and confirms—all without human touch. The implementation challenge is defining boundaries. Which workflows can AI fully automate? Which require human approval at certain steps? Which are too risky for AI at all? These decisions should be explicit, auditable, and revisable as you build confidence. **Related terms:** [AI agent orchestration](https://www.lorikeetcx.ai/glossary/ai-agent-orchestration), [Conversational AI](https://www.lorikeetcx.ai/glossary/conversational-ai), [Human-in-the-loop](https://www.lorikeetcx.ai/glossary/human-in-the-loop) ## Appointment Completion Rate URL: https://www.lorikeetcx.ai/glossary/appointment-completion-rate **Discover what appointment completion rate measures, how to calculate it, and strategies to reduce no-shows and improve scheduling outcomes.** **Appointment completion rate** is the percentage of scheduled telehealth appointments that result in a completed virtual visit between patient and provider. This metric matters because scheduled appointments that don't convert to completed visits represent lost revenue and wasted provider capacity. For telehealth operations, completion rates typically run 15-20% lower than in-person visits due to technical barriers, forgotten appointments, and the ease of simply not showing up when there's no physical commitment. Track this separately from no-show rate. A patient who shows up but can't complete the visit due to technical issues or provider delays counts differently than one who never appeared. The distinction points to different operational fixes—one is a patient engagement problem, the other is infrastructure or scheduling. High-performing telehealth operations target 85%+ completion rates. If you're below 75%, look at your reminder cadence, technical onboarding, and scheduling friction before blaming patient behavior. **Related terms:** [Virtual visit no-show rate](https://www.lorikeetcx.ai/glossary/virtual-visit-no-show-rate), [Virtual visit completion rate](https://www.lorikeetcx.ai/glossary/virtual-visit-completion-rate), [Technical connection success rate](https://www.lorikeetcx.ai/glossary/technical-connection-success-rate) ## AR Aging (120+ Days) URL: https://www.lorikeetcx.ai/glossary/ar-aging-120-days **Learn what AR aging over 120 days indicates, why aged receivables matter in healthcare, and how to reduce outstanding balances.** **AR aging (120+ days)** is the dollar value or percentage of accounts receivable that has remained uncollected for more than 120 days from the date of service. AR over 120 days old is generally considered problem AR—claims stuck in payer limbo, patient balances that likely won't be collected without aggressive intervention, or write-off candidates. Healthy telehealth organizations keep less than 15% of AR in the 120+ bucket; above 25% signals serious revenue cycle dysfunction. This metric is more actionable than total Days in AR because it focuses attention on the claims most likely to become bad debt. Work this bucket aggressively: identify patterns, escalate stuck claims, write off uncollectible balances, and improve upstream processes to prevent claims from aging into this bucket. Review AR aging weekly, not monthly. A claim that sits untouched from day 90 to day 120 crosses into problem territory silently. Catch these earlier. **Related terms:** [Days in Accounts Receivable](https://www.lorikeetcx.ai/glossary/days-in-accounts-receivable), [Net collection rate](https://www.lorikeetcx.ai/glossary/net-collection-rate), [Reimbursement cycle time](https://www.lorikeetcx.ai/glossary/reimbursement-cycle-time) ## Auto-tagging URL: https://www.lorikeetcx.ai/glossary/auto-tagging **Understand what auto-tagging is, how AI automatically categorizes support tickets, and why automated tagging improves routing and reporting.** Auto-tagging is the automated classification of customer service interactions by topic, intent, sentiment, or other metadata — replacing the manual process of agents categorizing tickets after handling them. AI-powered auto-tagging uses natural language understanding to analyze the content of a conversation and apply relevant tags in real-time. Accurate tagging matters because it's the foundation of CX analytics. Without consistent, reliable tags, teams can't identify trending issues, measure category-level resolution rates, or detect emerging problems. Manual tagging is notoriously inconsistent — agents tag the same issue differently, skip tagging when busy, and use categories inconsistently. Studies show manual tagging accuracy typically falls between 60-80%. Modern auto-tagging systems can: - Classify tickets across multiple taxonomies simultaneously (topic, product area, urgency, customer segment) - Apply tags in real-time as conversations progress, not just at close - Detect multiple intents within a single conversation - Identify emerging topics that don't fit existing categories - Maintain consistency across thousands of interactions per day The most valuable application of auto-tagging for CX leaders isn't operational routing (though that helps) — it's strategic visibility. When every interaction is tagged consistently, you can answer questions like: "What are the top 5 issues driving contact volume this week?" or "Which product change caused the spike in billing complaints?" These insights compound over time, creating a feedback loop between customer interactions and product decisions. **Related terms:** [intent detection](https://www.lorikeetcx.ai/glossary/intent-detection), [sentiment analysis](https://www.lorikeetcx.ai/glossary/sentiment-analysis), [ticketing system](https://www.lorikeetcx.ai/glossary/ticketing-system) ## Automated quality assurance (Auto QA) URL: https://www.lorikeetcx.ai/glossary/automated-quality-assurance-auto-qa **Learn what automated QA is in customer service, how AI scores 100% of interactions, and why Auto QA replaces manual review processes.** Automated quality assurance (Auto QA) uses AI to evaluate the quality of customer service interactions at scale — reviewing every conversation rather than the small sample (typically 2-5%) that manual QA processes can cover. Auto QA systems assess factors like accuracy, policy adherence, tone, resolution completeness, and customer effort. Traditional QA in customer service has a fundamental sampling problem. Even dedicated QA teams can only review a fraction of interactions, which means most quality issues go undetected until they surface as customer complaints or CSAT drops. Auto QA eliminates this gap by scoring 100% of conversations — whether handled by human agents or AI. Key capabilities of Auto QA systems include: - **Policy adherence checking:** Did the agent follow required procedures? Were mandatory disclosures made? - **Accuracy verification:** Was the information provided correct? Did the resolution match the customer's actual request? - **Tone and empathy scoring:** Was the interaction professional and appropriate for the situation? - **Compliance flagging:** Were regulatory requirements met? Were prohibited actions avoided? - **Trend analysis:** Surfacing systematic quality patterns across agents, teams, topics, or time periods Auto QA is particularly valuable in two scenarios: (1) for human agent teams, where it replaces subjective, sample-based review with comprehensive, consistent evaluation, and (2) for AI agent deployments, where it provides continuous monitoring of AI behavior in production. In the second case, Auto QA acts as a safety net — catching issues that guardrails missed and providing the feedback loop needed to improve AI performance over time. **Related terms:** [quality assurance in customer service](https://www.lorikeetcx.ai/glossary/quality-assurance-in-customer-service), [AI observability](https://www.lorikeetcx.ai/glossary/ai-observability), [customer satisfaction score](https://www.lorikeetcx.ai/glossary/customer-satisfaction-score) ## Automation Rate URL: https://www.lorikeetcx.ai/glossary/automation-rate **Discover what automation rate measures, how to calculate it for your support operation, and benchmarks for AI-driven customer service automation.** Automation Rate measures the percentage of support tickets where AI handled some or all of the interaction. It's the most commonly reported metric for AI support, and also the most easily gamed. Basic automation rate counts any ticket touched by AI: tickets where AI collected information before handoff, tickets where AI attempted resolution but failed, tickets where AI sent one message before the customer demanded a human. This inflates the number because it measures activity, not outcomes. Useful automation rate requires qualification: what counts as "automated"? A narrow definition only counts tickets resolved without human involvement. A broader definition includes tickets where AI handled first response but humans completed. The right definition depends on what you're trying to measure—cost savings (narrow) or AI utilization (broad). When evaluating AI vendors, always ask how they calculate automation rate. A vendor claiming 80% automation might be counting every ticket their bot touched, while a vendor claiming 50% might only count full resolutions. The 50% vendor may be delivering more actual value. **Related terms:** [Effective Automation Rate](https://www.lorikeetcx.ai/glossary/effective-automation-rate), [Containment Rate](https://www.lorikeetcx.ai/glossary/containment-rate), [Deflection Rate](https://www.lorikeetcx.ai/glossary/deflection-rate), [Match Rate](https://www.lorikeetcx.ai/glossary/match-rate) ## Average handling time (AHT) URL: https://www.lorikeetcx.ai/glossary/average-handling-time-aht **Learn what average handling time (AHT) is, how to calculate it, and proven strategies to reduce AHT while maintaining service quality.** Average handling time (AHT) is the mean duration of a customer service interaction, measured from the moment the interaction begins to its conclusion, including any hold time, transfer time, and post-interaction documentation. It is one of the most widely tracked operational metrics in customer service. **AHT = (Talk time + Hold time + After-call work) / Total interactions handled** AHT has historically been a primary efficiency metric for contact centers, used for workforce planning, cost modeling, and agent performance evaluation. A lower AHT means more interactions handled per agent per hour, which directly affects staffing costs. However, optimizing for AHT in isolation creates perverse incentives. Agents who rush through interactions to hit AHT targets may sacrifice resolution quality — transferring customers instead of resolving their issue, providing incomplete answers, or skipping verification steps. This is why modern CX organizations increasingly pair AHT with quality metrics like first contact resolution and customer satisfaction. In the context of AI-handled interactions, AHT takes on a different character. AI agents can handle multiple conversations simultaneously and don't experience fatigue, making raw AHT less meaningful as an efficiency metric. More relevant is whether the AI resolved the customer's issue completely in the interaction, regardless of duration. - That said, AHT remains useful for: - **Workforce planning:** Forecasting staffing needs for human-handled volume - **Complexity analysis:** Identifying which issue types take longest to resolve and may benefit from automation or process improvement - **Channel comparison:** Understanding relative efficiency across chat, email, phone, and AI-handled interactions **Related terms:** [first contact resolution](https://www.lorikeetcx.ai/glossary/first-contact-resolution), [cost per resolution](https://www.lorikeetcx.ai/glossary/cost-per-resolution), [resolution rate](https://www.lorikeetcx.ai/glossary/resolution-rate) ## Average Resolution Time (ART) URL: https://www.lorikeetcx.ai/glossary/average-resolution-time-art **Understand what average resolution time measures, how ART differs from first response time, and how to optimize resolution speed.** Average Resolution Time (ART) is the average elapsed time from when a customer submits a ticket to when their issue is fully resolved, measured across your entire ticket population. ART captures the customer's experience of time, not just agent effort. A ticket might have 15 minutes of actual handling time but take 3 days to resolve because it sat in queues, waited for approvals, or required information from another team. ART exposes those gaps. The formula is straightforward: Total time to resolution across all tickets / Number of tickets resolved. But "resolution" needs the same rigor as FCR—exclude tickets that reopen within your lookback window. Most teams measure ART in hours for live channels and days for async channels like email. ART is most useful when segmented. Your overall ART might be 18 hours, but that average hides a bimodal distribution: simple tickets closing in 2 hours and complex ones taking 4 days. Track ART by complexity tier, product area, and channel. Then attack the outliers. Often, 10% of your ticket types drive 50% of your total resolution time. **Related terms:** [Average Handling Time (AHT)](https://www.lorikeetcx.ai/glossary/average-handling-time-aht), [First Response Time (FRT)](https://www.lorikeetcx.ai/glossary/first-response-time-frt), [First Contact Resolution (FCR)](https://www.lorikeetcx.ai/glossary/first-contact-resolution-fcr) ## Average Speed of Answer (ASA) URL: https://www.lorikeetcx.ai/glossary/average-speed-of-answer-asa **Learn what average speed of answer (ASA) is, how to measure it, and why ASA directly impacts customer satisfaction and abandonment rates.** **Average speed of answer (ASA)** is the mean time customers wait in queue before connecting with a live agent, measured from the moment they enter the queue until an agent picks up. ASA matters because it directly correlates with customer frustration and abandon rates. Every second in queue erodes patience—research consistently shows abandonment spikes after 60-90 seconds for most contact centers. For fintech and insurance companies handling urgent issues like fraud alerts or claims, a high ASA can turn a recoverable situation into a lost customer. The trap with ASA is optimizing it in isolation. Teams that staff up purely to hit an ASA target often tank their agent occupancy and blow their cost-per-contact. The better approach: use ASA as a diagnostic alongside first reply time and call abandon rate to understand where queue friction actually lives. If your ASA is high but abandon rate is low, customers might be more patient than you think—or your IVR is doing its job filtering simple queries. **Related terms:** [Average wait time](https://www.lorikeetcx.ai/glossary/average-wait-time), [First reply time](https://www.lorikeetcx.ai/glossary/first-reply-time), [Call abandon rate](https://www.lorikeetcx.ai/glossary/call-abandon-rate), [Agent occupancy](https://www.lorikeetcx.ai/glossary/agent-occupancy) ## Average Virtual Wait Time URL: https://www.lorikeetcx.ai/glossary/average-virtual-wait-time **Discover what average virtual wait time measures in telehealth, how to reduce patient wait times, and benchmarks for virtual care.** **Average virtual wait time** is the elapsed time between a patient's scheduled appointment start and when the provider actually joins the virtual visit. This is the telehealth equivalent of sitting in a waiting room—except patients are staring at their own reflection in a paused video screen, wondering if something broke. Every minute of wait time erodes trust and satisfaction more aggressively than in-person waits because patients have nothing to do and constant uncertainty about whether the visit will happen at all. Industry benchmarks target under 5 minutes average wait time, with high performers achieving under 2 minutes. Track your 90th percentile, not just averages—a handful of 30-minute waits can hide behind an acceptable average while destroying those patients' experience and generating complaints. Reducing wait times requires addressing provider schedule compression, overbooking practices, and technical handoff processes. Some platforms use "virtual waiting rooms" with estimated wait times and educational content, which improves perceived wait experience even when actual times stay constant. **Related terms:** [Patient satisfaction score](https://www.lorikeetcx.ai/glossary/patient-satisfaction-score), [Provider utilization rate](https://www.lorikeetcx.ai/glossary/provider-utilization-rate), [Appointment completion rate](https://www.lorikeetcx.ai/glossary/appointment-completion-rate) ## Average Wait Time URL: https://www.lorikeetcx.ai/glossary/average-wait-time **Learn what average wait time is in customer service, how to calculate it across channels, and strategies to minimize customer wait times.** **Average wait time** is the total time customers spend waiting for service across all touchpoints—including IVR navigation, queue holds, and transfers—before their issue begins being addressed. This metric is broader than ASA, which only captures queue time before agent pickup. Average wait time accounts for the full customer experience: the 45 seconds navigating phone menus, the hold while being transferred to a specialist, the dead air between responses in chat. It's the honest answer to "how long did this actually take?" For omnichannel operations, average wait time reveals friction that channel-specific metrics miss. Your chat ASA might look great at 30 seconds, but if customers waited 5 minutes for a callback that then took 2 more minutes to connect, the experience was far worse than the numbers suggest. Healthtech and insurance companies with complex routing logic should track this closely—transfers and escalations are where wait time compounds. **Related terms:** [Average speed of answer](https://www.lorikeetcx.ai/glossary/average-speed-of-answer), [First reply time](https://www.lorikeetcx.ai/glossary/first-reply-time), [Escalation rate](https://www.lorikeetcx.ai/glossary/escalation-rate) ## Behavioural Signals URL: https://www.lorikeetcx.ai/glossary/behavioural-signals **Understand what behavioural signals are in customer service, how AI detects frustration and intent from user behavior, and why signals improve CX.** **Behavioural signals** are indicators within customer messages—word choice, sentiment shifts, urgency markers, implicit requests—that inform how an AI support system should respond beyond the literal content of the query. A customer asking "how do I cancel?" might be seeking cancellation instructions or might be expressing frustration that signals a retention opportunity. The words are identical; the behavioral signals differ. Prior purchase history, message tone, time since last contact, and comparison to similar customers all contribute to interpreting intent. This matters because customer support isn't information retrieval—it's relationship management. A system that responds only to explicit requests misses the 80% of communication that's implicit. Behavioral signal detection lets AI systems adjust tone, escalate proactively, or offer retention incentives before the customer explicitly asks to leave. The risk is overreach. Behavioral inference can feel invasive if misapplied. The best implementations use signals to route and adapt, not to manipulate. Detecting frustration should trigger empathy and faster resolution, not aggressive upselling. **Related terms:** [Ticket Complexity](https://www.lorikeetcx.ai/glossary/ticket-complexity), [Warm Handoff](https://www.lorikeetcx.ai/glossary/warm-handoff), [Conversational Configuration](https://www.lorikeetcx.ai/glossary/conversational-configuration) --- # AI & Automation Terms ## Business process outsourcing (BPO) URL: https://www.lorikeetcx.ai/glossary/business-process-outsourcing-bpo **Learn what business process outsourcing (BPO) is, how it works for customer service, and how AI is transforming the BPO model.** Business process outsourcing (BPO) is the practice of contracting customer service operations to third-party providers, typically to reduce costs through labor arbitrage — hiring agents in lower-cost markets (Philippines, India, Latin America) rather than domestically. BPO has been the dominant scaling strategy for customer service for over two decades. The economics of BPO are straightforward: an offshore agent costs $8-15/hour compared to $18-35/hour domestically. For a company processing 10,000 tickets per month, the savings are substantial. However, the BPO model carries well-documented tradeoffs: - **Quality control:** Outsourced agents typically have less product knowledge and cultural context, leading to lower CSAT and higher escalation rates - **Training overhead:** High turnover at BPO providers (often 30-60% annually) means continuous training investment - **Communication gaps:** Time zone differences, language barriers, and organizational distance create friction - **Brand risk:** Customers don't distinguish between in-house and outsourced agents — a poor BPO interaction is a poor brand experience AI customer service agents are fundamentally reshaping the BPO equation. Instead of choosing between expensive domestic agents and cheaper-but-lower-quality offshore agents, companies can deploy AI that handles a significant portion of volume at a fraction of the cost of either option — often with higher consistency and 24/7 availability. This doesn't mean BPO disappears. It means the role shifts: BPOs increasingly handle the complex, high-judgment interactions that AI escalates, while AI handles the high-volume, process-driven work that BPOs traditionally staffed. Companies that previously outsourced 80% of volume may now outsource 20-30%, with AI handling the rest. **Related terms:** [cost per resolution](https://www.lorikeetcx.ai/glossary/cost-per-resolution), [AI agent](https://www.lorikeetcx.ai/glossary/ai-agent), [escalation rate](https://www.lorikeetcx.ai/glossary/escalation-rate) ## Call Abandon Rate URL: https://www.lorikeetcx.ai/glossary/call-abandon-rate **Discover what call abandon rate measures, how to calculate it, and proven strategies to reduce call abandonment in your contact center.** **Call abandon rate** is the percentage of inbound calls where customers hang up before connecting with an agent, typically excluding disconnects within the first 5-10 seconds (short abandons). Abandon rate is your canary in the coal mine for staffing problems. When it spikes, customers are telling you—by leaving—that your wait times are unacceptable. The standard benchmark of 5-8% is meaningless without context; what matters is your baseline and how it moves. A sudden jump from 4% to 7% demands investigation even if 7% looks "normal" on paper. The short-abandon exclusion matters because misdials and accidental callbacks inflate the raw number. Most platforms let you set a threshold (5-10 seconds) below which disconnects don't count. But don't set it too high—a customer who hangs up at 25 seconds after hearing "estimated wait time: 20 minutes" made a real choice to leave. For insurance and healthtech companies with urgent use cases, segment abandon rate by call type. Fraud line abandons are a different severity than billing inquiry abandons. **Related terms:** [Average speed of answer](https://www.lorikeetcx.ai/glossary/average-speed-of-answer), [Average wait time](https://www.lorikeetcx.ai/glossary/average-wait-time), [Contact rate](https://www.lorikeetcx.ai/glossary/contact-rate) ## Care Plan Adherence Rate URL: https://www.lorikeetcx.ai/glossary/care-plan-adherence-rate **Learn what care plan adherence rate measures in healthcare, how to track patient compliance, and strategies to improve health outcomes.** **Care plan adherence rate** is the percentage of patients who follow their prescribed treatment protocols, including medication compliance, follow-up visits, and lifestyle modifications. Adherence is where telehealth demonstrates its clinical value proposition. The ability to check in with patients between visits, monitor progress remotely, and intervene at early signs of non-adherence creates opportunities for care management that in-person practices can't match without prohibitive costs. Measure adherence across multiple dimensions: medication adherence (prescription fills and refills), visit adherence (follow-up appointments kept), and protocol adherence (labs completed, lifestyle changes documented). Each reveals different intervention opportunities. High adherence rates correlate with better outcomes, lower downstream costs, and stronger payer relationships. Value-based contracts increasingly reward adherence improvements, making this metric financially material beyond its clinical importance. Telehealth organizations that can demonstrate superior adherence rates have a compelling story for payer contracting negotiations. **Related terms:** [Prescription fulfillment rate](https://www.lorikeetcx.ai/glossary/prescription-fulfillment-rate), [Virtual visit completion rate](https://www.lorikeetcx.ai/glossary/virtual-visit-completion-rate), [Patient satisfaction score](https://www.lorikeetcx.ai/glossary/patient-satisfaction-score) ## Channel switching URL: https://www.lorikeetcx.ai/glossary/channel-switching **Understand what channel switching is in customer support, why customers switch between channels, and how to deliver seamless omnichannel experiences.** Channel switching occurs when a customer moves between support channels during the course of resolving a single issue — starting on chat, following up via email, then calling in. High channel-switching rates indicate friction in the service experience and are a strong predictor of customer dissatisfaction. Customers switch channels for several reasons: - **Resolution failure:** The first channel didn't resolve their issue, so they try another - **Accessibility:** The preferred channel isn't available (chat outside business hours, phone queue too long) - **Escalation:** The issue is too complex for the current channel (e.g., needs a phone conversation, not chat) - **Frustration:** The customer is unhappy with the experience on the current channel and hopes another will be better Channel switching is expensive for the business and frustrating for the customer. Each switch often means re-explaining the issue, re-authenticating, and starting the resolution process over. The business incurs multiple costs for a single issue, and the customer's effort multiplies. AI can both reduce and increase channel switching, depending on implementation quality: - **Reducing:** AI that resolves issues on the first channel contact eliminates the need to switch - **Increasing:** Poor AI experiences that fail to resolve issues drive customers to try other channels For CX teams, channel-switching rate is a diagnostic metric. A sudden increase in switching from chat to phone, for example, might indicate that the AI chat experience has degraded. Tracking switching patterns reveals where the service experience is breaking down and where investment (in AI capability, in staffing, in self-service content) is most needed. **Related terms:** [omnichannel customer support](https://www.lorikeetcx.ai/glossary/omnichannel-customer-support), [customer effort score](https://www.lorikeetcx.ai/glossary/customer-effort-score), [first contact resolution](https://www.lorikeetcx.ai/glossary/first-contact-resolution) ## Chargeback Rate URL: https://www.lorikeetcx.ai/glossary/chargeback-rate **Learn what chargeback rate is, how to calculate it for financial services, and strategies to reduce chargebacks and protect revenue.** **Chargeback rate** is the percentage of transactions disputed by cardholders and reversed by their issuing bank, calculated as total chargebacks divided by total transactions over a given period. Card networks track this number closely. Visa and Mastercard will place you in monitoring programs if your chargeback rate exceeds 0.9-1%, and excessive chargebacks can get you dropped by payment processors entirely. For fintech CX leaders, this metric is existential. Most chargebacks fall into three categories: fraud (unauthorized transactions), merchant error (duplicate charges, wrong amounts), and friendly fraud (customers disputing legitimate purchases). Your support team's job is to prevent the preventable ones. Clear transaction descriptors reduce "I don't recognize this charge" disputes. Proactive communication about billing reduces subscription chargebacks. Fast refund processes stop customers from going to their bank instead of you. The real leverage is in your dispute response. When chargebacks come in, you have a narrow window to submit evidence. Teams that systematically gather transaction logs, delivery confirmations, and communication records win more representments. Track your win rate on disputes—it should be 40% or higher for non-fraud cases. **Related terms:** [Payment failure rate](https://www.lorikeetcx.ai/glossary/payment-failure-rate), [Transaction approval rate](https://www.lorikeetcx.ai/glossary/transaction-approval-rate), [Fraud prevention](https://www.lorikeetcx.ai/glossary/fraud-prevention) ## Chatbot URL: https://www.lorikeetcx.ai/glossary/chatbot **Discover what a chatbot is, how modern AI chatbots differ from rule-based bots, and why chatbots are essential for scalable customer support.** **A chatbot** is a software application that simulates text-based conversation with users, ranging from simple rule-based responders to sophisticated AI-powered agents capable of complex dialogue. The term "chatbot" covers enormous variance. At one end: keyword-matching systems that recognize "hours" and return store hours, breaking down on any variation. At the other end: large language model-powered agents that understand nuanced requests, maintain multi-turn context, and execute complex workflows. Calling both "chatbots" obscures critical capability differences. For evaluating chatbots, focus on specific capabilities: Can it handle questions not explicitly programmed? Can it maintain context across multiple turns? Can it execute actions, not just provide information? Can it recognize when it's failing and escalate gracefully? Can it improve from feedback? The answers to these questions determine whether a chatbot will create value or frustration. Chatbot reputation has been damaged by years of poor implementations. Many customers approach chatbots expecting to be frustrated and looking for "talk to human" options. Overcoming this perception requires delivering actually useful interactions—which is possible with modern AI, but requires more than just deploying a chatbot. **Related terms:** [Conversational AI](https://www.lorikeetcx.ai/glossary/conversational-ai), [Intelligent Virtual Agent](https://www.lorikeetcx.ai/glossary/intelligent-virtual-agent), [Fallback intent](https://www.lorikeetcx.ai/glossary/fallback-intent) ## Claims Denial Rate URL: https://www.lorikeetcx.ai/glossary/claims-denial-rate **Understand what claims denial rate measures in healthcare, common reasons for denials, and how to reduce denial rates to improve revenue.** **Claims denial rate** is the percentage of submitted insurance claims that are rejected or denied by payers on initial submission. Telehealth claims face unique denial risks: incorrect place-of-service codes, missing documentation of patient location, payer-specific telehealth coverage limitations, and provider credentialing gaps for out-of-state patients. A healthy telehealth denial rate is under 5%; rates above 10% indicate systematic billing process failures requiring immediate attention. Not all denials are equal. Distinguish between "soft" denials (missing information, easily corrected) and "hard" denials (coverage exclusions, medical necessity rejections). Soft denials indicate process gaps; hard denials indicate eligibility verification or clinical documentation problems. Track denials by payer and denial reason code. Patterns will emerge—one payer rejecting all audio-only visits, another denying specific CPT codes, a third requiring documentation your templates don't capture. These patterns turn an overwhelming pile of denials into a focused improvement roadmap. **Related terms:** [Clean claim rate](https://www.lorikeetcx.ai/glossary/clean-claim-rate), [First-pass resolution rate](https://www.lorikeetcx.ai/glossary/first-pass-resolution-rate), [Prior authorization success rate](https://www.lorikeetcx.ai/glossary/prior-authorization-success-rate) ## Claims Leakage URL: https://www.lorikeetcx.ai/glossary/claims-leakage **Learn what claims leakage is in insurance, how it impacts profitability, and strategies to detect and prevent unnecessary claims overpayment.** **Claims leakage** is the difference between what an insurer actually pays on claims and what should have been paid based on accurate assessment of coverage, damages, and liability. Claims leakage includes both overpayment (paying more than owed due to adjuster error, fraud, or inflated estimates) and underpayment (paying less than owed, which creates compliance risk and customer dissatisfaction). Most insurers focus on overpayment when measuring leakage, but underpayment can be equally costly through regulatory action and litigation. Industry estimates suggest leakage ranges from 5-10% of total claims spend for average performers. For a $1 billion book, that's $50-100 million in unnecessary payouts annually. Even small improvements in leakage have outsized P&L impact because every dollar of reduced leakage flows directly to underwriting profit. Leakage is difficult to measure precisely because you're comparing actual payments to a hypothetical correct amount. Most insurers use sampling-based audits, comparing closed claims against independent re-adjudication. AI-powered claims review can now do this continuously rather than through periodic audits, flagging potential overpayments before they occur. **Related terms:** [Loss Adjustment Expense ratio](https://www.lorikeetcx.ai/glossary/loss-adjustment-expense-ratio), [Fraud detection](https://www.lorikeetcx.ai/glossary/fraud-detection), [Claims resolution rate](https://www.lorikeetcx.ai/glossary/claims-resolution-rate) ## Claims Reopening Rate URL: https://www.lorikeetcx.ai/glossary/claims-reopening-rate **Discover what claims reopening rate measures, why reopened claims increase costs, and how to improve first-time claims resolution.** **Claims reopening rate** is the percentage of closed claims that are subsequently reopened due to new information, disputed outcomes, additional damages, or errors in the original settlement. Reopening rate is a lagging indicator of claims quality. High reopening rates signal problems upstream: insufficient investigation, premature closure pressure, poor communication with policyholders, or adjuster inexperience. Every reopened claim costs significantly more to process than getting it right the first time. The target for most lines of business is a reopening rate below 3%. Rates above 5% indicate systematic problems worth investigating. Segment by reason code: are claims reopening because of adjuster error, new damage discovery, or policyholder disputes? Each cause requires different interventions. Reopening rate creates tension with cycle time metrics. Adjusters pressured to close claims quickly may close them incompletely, driving up reopening rates. Effective claims operations balance both metrics, tracking them together rather than optimizing one at the expense of the other. **Related terms:** [Claims resolution rate](https://www.lorikeetcx.ai/glossary/claims-resolution-rate), [Claims settlement cycle time](https://www.lorikeetcx.ai/glossary/claims-settlement-cycle-time), [Claims leakage](https://www.lorikeetcx.ai/glossary/claims-leakage) ## Claims Resolution Rate URL: https://www.lorikeetcx.ai/glossary/claims-resolution-rate **Learn what claims resolution rate is, how to calculate it, and best practices for improving the speed and accuracy of claims processing.** **Claims resolution rate** is the percentage of claims closed within a defined time period, typically measured as the ratio of claims resolved to claims received. This metric answers a simple question: is your claims operation keeping pace with incoming volume, or is backlog growing? A resolution rate below 100% sustained over time indicates systemic capacity problems. A rate significantly above 100% suggests you're working through backlog—or potentially rushing claims to closure. The danger with claims resolution rate is optimizing for speed at the expense of accuracy. Insurers that pressure adjusters to close claims quickly often see increases in claims reopening rates and customer complaints. The metric should always be paired with quality indicators: reopening rate, customer satisfaction, and appeals volume. Resolution rate becomes more useful when segmented. Tracking it by claim type, region, adjuster, and complexity tier reveals where bottlenecks actually exist. A healthy operation might show 98% resolution on simple claims but only 70% on complex claims—and that's fine, as long as you've staffed accordingly. **Related terms:** [Claims settlement cycle time](https://www.lorikeetcx.ai/glossary/claims-settlement-cycle-time), [Claims reopening rate](https://www.lorikeetcx.ai/glossary/claims-reopening-rate), [Claims triage](https://www.lorikeetcx.ai/glossary/claims-triage) ## Claims Settlement Cycle Time URL: https://www.lorikeetcx.ai/glossary/claims-settlement-cycle-time **Understand what claims settlement cycle time measures, industry benchmarks, and how automation reduces time from filing to payment.** **Claims settlement cycle time** is the total elapsed time from when a claim is filed to when the final payment is issued to the policyholder. This is the metric policyholders care about most—and the one that drives the largest variance in NPS scores across insurers. Cycle time compounds: every additional day a claim remains open increases operational cost, policyholder frustration, and the likelihood of escalation or litigation. Cycle time varies dramatically by line of business. Auto glass claims might close in hours; complex liability claims can stretch for years. The useful comparison is against your own historical performance within each claim type, not against industry averages that blend simple and complex claims together. Reducing cycle time requires attacking handoff delays, not just processing speed. Most claims spend the majority of their lifecycle waiting—for documents, for adjuster availability, for approvals. Insurers achieving best-in-class cycle times have typically automated document collection, implemented parallel processing, and given front-line adjusters broader settlement authority. **Related terms:** [Claims resolution rate](https://www.lorikeetcx.ai/glossary/claims-resolution-rate), [Straight-through processing rate](https://www.lorikeetcx.ai/glossary/straight-through-processing-rate), [Claims leakage](https://www.lorikeetcx.ai/glossary/claims-leakage) ## Claims Triage URL: https://www.lorikeetcx.ai/glossary/claims-triage **Learn what claims triage is in insurance, how AI automates initial claims assessment, and why triage improves processing speed and accuracy.** **Claims triage** is the initial assessment and routing of incoming claims based on complexity, severity, fraud indicators, and required expertise to ensure appropriate handling from the start. Effective triage is the foundation of efficient claims operations. Routing a simple claim to a senior adjuster wastes expensive expertise. Routing a complex claim to a junior adjuster creates errors, delays, and poor customer outcomes. Triage decisions made in the first minutes of a claim determine much of its eventual cost and cycle time. Modern triage combines rules-based logic with machine learning. Rules handle the obvious: glass-only auto claims route to the glass vendor, water damage claims route to the property team. ML adds nuance: this claim looks simple but has fraud indicators; this claimant has a litigation history; this damage description suggests coverage complexity. The goal is matching claim characteristics to the right handling path. Triage should also identify straight-through processing candidates. Claims meeting criteria for automation should bypass human adjusters entirely. The triage decision isn't just "which adjuster?" but "does this need an adjuster at all?" **Related terms:** [Straight-through processing rate](https://www.lorikeetcx.ai/glossary/straight-through-processing-rate), [FNOL response time](https://www.lorikeetcx.ai/glossary/fnol-response-time), [Fraud detection](https://www.lorikeetcx.ai/glossary/fraud-detection) ## Clean Claim Rate URL: https://www.lorikeetcx.ai/glossary/clean-claim-rate **Discover what clean claim rate measures in healthcare billing, how to calculate it, and strategies to submit error-free claims consistently.** **Clean claim rate** is the percentage of insurance claims submitted that are accepted and processed without requiring additional information, corrections, or resubmission. A clean claim passes through payer adjudication on first touch. Every claim that isn't clean requires human intervention—someone researching the rejection, correcting information, resubmitting, and tracking. This rework is expensive, delays payment, and scales linearly with volume. Telehealth organizations should target 95%+ clean claim rates. Below 90% signals systematic problems with your billing templates, eligibility verification, or documentation workflows. Common telehealth-specific clean claim failures include incorrect modifier usage (GT, 95, etc.), missing originating site documentation, and state-specific telehealth coverage verification. Improving clean claim rate often delivers the highest ROI of any revenue cycle investment because it reduces both denial rework and Days in AR simultaneously while improving payer relationships. **Related terms:** [Claims denial rate](https://www.lorikeetcx.ai/glossary/claims-denial-rate), [First-pass resolution rate](https://www.lorikeetcx.ai/glossary/first-pass-resolution-rate), [Reimbursement cycle time](https://www.lorikeetcx.ai/glossary/reimbursement-cycle-time) ## Compliance automation URL: https://www.lorikeetcx.ai/glossary/compliance-automation **Learn what compliance automation is, how AI automates regulatory checks in customer service, and why it reduces risk and manual effort.** Compliance automation uses AI and rules-based systems to enforce regulatory requirements within customer service interactions — ensuring that mandatory disclosures are delivered, prohibited actions are prevented, required verifications are performed, and interaction records meet regulatory standards, all without depending on individual agent knowledge or memory. In regulated industries, compliance failures in customer service create real liability: - **Financial services:** Failing to disclose fees, providing unsuitable product advice, or processing unauthorized transactions - **Healthcare:** Sharing medical information without proper consent, providing clinical advice outside scope, or violating HIPAA data handling requirements - **Insurance:** Misrepresenting coverage, failing to deliver mandatory disclosures, or improperly handling claims Traditional compliance in customer service relies on agent training and manual QA — both of which are inherently limited. Agents forget procedures under pressure, new hires make mistakes, and QA sampling catches only a fraction of violations. AI-powered compliance automation can: - **Enforce disclosure requirements:** Automatically including required language in relevant interactions - **Block prohibited actions:** Preventing the AI (or flagging for human agents) actions that violate regulatory rules - **Verify identity:** Ensuring proper authentication before accessing sensitive information - **Maintain records:** Creating audit-ready logs of every compliance-relevant decision and action - **Monitor continuously:** Reviewing 100% of interactions for compliance violations, not just a sample For CX leaders in regulated industries, compliance automation isn't just an efficiency tool — it's a risk management strategy. The cost of a compliance violation (fines, litigation, reputational damage) dwarfs the cost of implementing automated controls. Compliance should be a primary evaluation criterion when selecting AI customer service platforms, not an afterthought. **Related terms:** [AI compliance](https://www.lorikeetcx.ai/glossary/ai-compliance), [AI audit trail](https://www.lorikeetcx.ai/glossary/ai-audit-trail), [AI guardrails](https://www.lorikeetcx.ai/glossary/ai-guardrails), [automated quality assurance](https://www.lorikeetcx.ai/glossary/automated-quality-assurance) ## Contact center as a service (CCaaS) URL: https://www.lorikeetcx.ai/glossary/contact-center-as-a-service-ccaas **Understand what CCaaS is, how cloud contact center platforms work, and why businesses are migrating from on-premise to CCaaS solutions.** Contact center as a service (CCaaS) is a cloud-based software model that provides the technology infrastructure for running a customer service operation — including call routing, agent desktop, workforce management, reporting, and omnichannel capabilities — without requiring on-premises hardware or telephony infrastructure. Major CCaaS platforms include Amazon Connect, Genesys Cloud, NICE CXone, Five9, and Talkdesk. These platforms replaced the previous generation of on-premises contact center systems (Avaya, Cisco) by offering lower upfront costs, faster deployment, and easier scaling. CCaaS platforms are now evolving to integrate AI capabilities, but most approach this by bolting AI features onto existing architectures — adding a chatbot layer on top of ticketing, or using AI for agent assist within the existing workflow. This creates limitations: the AI operates within the constraints of the platform rather than being purpose-built for autonomous resolution. For CX leaders evaluating their technology stack, the key question is whether their CCaaS platform is the right layer at which to introduce AI automation. Some teams find that a dedicated AI agent platform that integrates with their existing CCaaS (handling the AI-first conversations and escalating complex cases back to the CCaaS for human handling) delivers better results than relying on their CCaaS provider's native AI features. The market is actively consolidating, with CCaaS providers acquiring AI capabilities and AI-native platforms building contact center features. The architectural distinction — AI bolted onto legacy workflows vs. AI-native with human escalation — will likely define the next generation of the category. **Related terms:** [omnichannel customer support](https://www.lorikeetcx.ai/glossary/omnichannel-customer-support), [interactive voice response](https://www.lorikeetcx.ai/glossary/interactive-voice-response), [AI agent](https://www.lorikeetcx.ai/glossary/ai-agent) ## Contact Rate URL: https://www.lorikeetcx.ai/glossary/contact-rate **Learn what contact rate measures, how to calculate it across support channels, and strategies to optimize your customer contact volume.** **Contact rate** is the ratio of customer support contacts to a business activity metric—typically contacts per order, per active user, per transaction, or per policy. This is the metric that connects support operations to business outcomes. Raw ticket volume is noise without context; contact rate tells you whether your support burden is growing faster than your business. A fintech processing 100,000 transactions with 500 support tickets (0.5% contact rate) is in a fundamentally different position than one with the same volume generating 5,000 tickets. Contact rate trends reveal product and process health. Rising contact rate often predates churn spikes—customers hit friction, contact support, and if unresolved, leave. For healthtech and insurance, segment by journey stage: enrollment contact rate, claims contact rate, renewal contact rate. This exposes where your product experience breaks down. The goal isn't zero contact rate—some contact is valuable (upsell opportunities, feedback)—but unnecessary contact rate should trend down as you improve self-service and product clarity. **Related terms:** [Self-service ratio](https://www.lorikeetcx.ai/glossary/self-service-ratio), [Ticket volume](https://www.lorikeetcx.ai/glossary/ticket-volume), [First reply time](https://www.lorikeetcx.ai/glossary/first-reply-time) ## Containment Rate URL: https://www.lorikeetcx.ai/glossary/containment-rate **Discover what containment rate measures in AI customer service, how to calculate it, and why higher containment means more efficient automation.** Containment Rate measures the percentage of support interactions that stay within a self-service or AI channel without escalating to human agents. It answers: "How many customers did we keep out of the human queue?" Containment is a queue management metric, not a quality metric. A ticket is "contained" if it never reaches a human, regardless of whether the customer's issue was actually resolved. The customer might have found their answer, given up in frustration, or abandoned after three failed bot loops. All three count as contained. This distinction matters. High containment with low customer satisfaction means you're trapping customers, not helping them. Containment rate should always be paired with quality metrics: CSAT for contained interactions, repeat contact rate, or downstream churn analysis. That said, containment rate is valuable for capacity planning. If you're trying to reduce human agent load, containment rate directly measures that outcome. Just don't mistake it for customer success. **Related terms:** [Deflection Rate](https://www.lorikeetcx.ai/glossary/deflection-rate), [Automation Rate](https://www.lorikeetcx.ai/glossary/automation-rate), [Resolution Rate](https://www.lorikeetcx.ai/glossary/resolution-rate), [Independently Resolved (IR)](https://www.lorikeetcx.ai/glossary/independently-resolved-ir) ## Context Window URL: https://www.lorikeetcx.ai/glossary/context-window **Learn what a context window is in AI, how token limits affect LLM conversations, and why context window size matters for customer service.** **Context window** is the maximum amount of text (measured in tokens) that a large language model can process in a single interaction—including both the input provided and the output generated. Context window determines how much information the AI can "see" at once. A customer service AI needs to hold: the system prompt defining behavior, the conversation history, retrieved knowledge articles, customer data, and any other relevant context—all while leaving room for generating a response. Longer conversations, more retrieved documents, more customer data all consume context window. Practical implications: If the context window fills up, older information gets dropped. The AI might "forget" what the customer said at the start of a long conversation. Knowledge retrieval must be selective—you can't feed the AI your entire knowledge base, only the most relevant chunks. Complex workflows that require reasoning across many data points need sufficient context to hold all relevant information. Context window sizes have expanded dramatically with newer models, from thousands of tokens to hundreds of thousands. This shifts the constraint from "can the AI fit the context" to "can we efficiently find and structure the relevant context." Larger windows don't automatically mean better performance—the AI still needs help finding what matters in a sea of potentially relevant information. **Related terms:** [AI agent memory](https://www.lorikeetcx.ai/glossary/ai-agent-memory), [Multi-turn conversation](https://www.lorikeetcx.ai/glossary/multi-turn-conversation), [Retrieval augmented generation](https://www.lorikeetcx.ai/glossary/retrieval-augmented-generation) ## Conversational AI URL: https://www.lorikeetcx.ai/glossary/conversational-ai **Understand what conversational AI is, how it powers natural customer interactions across channels, and why it is transforming customer experience.** Conversational AI is the broad category of AI technologies that enable machines to understand, process, and generate human language in dialogue. It encompasses the natural language processing, understanding, and generation capabilities that power AI agents, virtual assistants, and automated customer interactions. The technology stack behind conversational AI has evolved dramatically. Early systems relied on pattern matching and decision trees — if the customer says X, respond with Y. Modern conversational AI uses large language models that understand intent, context, and nuance, enabling genuine multi-turn conversations that feel natural. Key capabilities of modern conversational AI in customer service include: - **Intent understanding:** Recognizing what the customer wants, even when expressed ambiguously or colloquially - **Context tracking:** Maintaining awareness of the full conversation history, including topic switches and follow-up questions - **Tone adaptation:** Adjusting communication style based on the customer's emotional state and the nature of the interaction - **Multi-language support:** Handling conversations across languages without separate models for each - **Channel flexibility:** Operating consistently across chat, email, voice, and social media The distinction that matters for CX teams is between conversational AI as a front-end experience (making interactions feel more natural) and conversational AI as a resolution engine (actually completing the work the customer needs done). Many systems excel at the former while falling short on the latter. A smooth, natural conversation that ends with "I've transferred you to a specialist" is still a deflection, regardless of how conversational it felt. **Related terms:** [AI agent](https://www.lorikeetcx.ai/glossary/ai-agent), [natural language processing](https://www.lorikeetcx.ai/glossary/natural-language-processing), [intent detection](https://www.lorikeetcx.ai/glossary/intent-detection) ## Conversational AI Design URL: https://www.lorikeetcx.ai/glossary/conversational-ai-design **Learn what conversational AI design is, key principles for designing effective AI conversations, and how good design improves resolution rates.** **Conversational AI design** is the discipline of architecting AI-powered conversation systems that effectively understand user needs, manage dialogue flow, and deliver satisfying customer experiences. Good conversational AI design is invisible—the conversation feels natural, the system understands what you mean, issues get resolved without friction. Bad design is immediately obvious: the AI misunderstands, asks redundant questions, gets stuck in loops, or provides generic non-answers. Key design considerations include: conversation architecture (how dialogues flow between topics and states), error handling (what happens when the AI doesn't understand), escalation design (when and how to involve humans), personality and tone (how the AI should "sound"), and edge case handling (the weird queries that don't fit normal patterns). Design must also account for multi-turn dynamics—conversations aren't isolated exchanges but evolving dialogues where context accumulates. The shift to generative AI changes design priorities. Less time goes into scripting every possible response; more goes into guardrails, persona definition, and handling the unexpected. The AI has more capability but needs more guidance on appropriate boundaries. Design becomes less about what the AI says and more about what it shouldn't say. **Related terms:** [Conversational AI](https://www.lorikeetcx.ai/glossary/conversational-ai), [Multi-turn conversation](https://www.lorikeetcx.ai/glossary/multi-turn-conversation), [Fallback intent](https://www.lorikeetcx.ai/glossary/fallback-intent) ## Conversational Analytics URL: https://www.lorikeetcx.ai/glossary/conversational-analytics **Discover what conversational analytics is, how AI analyzes customer conversations at scale, and the insights it unlocks for CX improvement.** **Conversational analytics** is the analysis of text and voice conversations to extract insights about customer needs, agent performance, and operational opportunities. Conversational analytics transforms chat logs and call recordings from archives into intelligence. What are customers actually asking about? Where do conversations get stuck? Which issues correlate with negative sentiment? What language predicts escalation? Analytics answers these questions at scale, processing thousands of conversations to surface patterns. Core capabilities include: topic clustering (grouping conversations by subject), sentiment tracking (monitoring customer emotion across interactions), intent analysis (understanding what customers want), agent evaluation (comparing handling approaches), and trend detection (identifying emerging issues). Advanced analytics adds predictive modeling—forecasting volume, predicting escalation risk, identifying churn signals. For CX leaders, conversational analytics is the foundation of continuous improvement. Without it, you're relying on anecdotes and sampled QA reviews. With it, you can identify that 23% of conversations about topic X result in escalation, or that customers who use phrase Y are significantly more likely to churn. This drives targeted intervention rather than generic improvement efforts. **Related terms:** [AI observability](https://www.lorikeetcx.ai/glossary/ai-observability), [Sentiment analysis](https://www.lorikeetcx.ai/glossary/sentiment-analysis), [Intent recognition](https://www.lorikeetcx.ai/glossary/intent-recognition) ## Conversational configuration URL: https://www.lorikeetcx.ai/glossary/conversational-configuration **Learn what conversational configuration is, how natural language setup simplifies AI agent management, and why it reduces technical overhead.** Conversational configuration is an approach to setting up and managing AI systems where teams configure behavior through natural language conversation rather than code, rule builders, or complex admin interfaces. Instead of writing if-then rules or navigating settings panels, a team member describes what they want the AI to do in plain language, and the system interprets and implements the instruction. In customer service, conversational configuration changes who can manage the AI system. Traditional automation tools require technical skills — building decision trees, writing rules, configuring integrations. Conversational configuration puts control in the hands of CX leaders and frontline managers who understand the business deeply but may not be technical. Examples of conversational configuration include: - "When a customer asks about our refund policy after 30 days, explain that we offer store credit instead of cash refunds" - "If a customer mentions they're a Gold member, skip the standard verification and use the expedited process" - "We just changed our shipping policy for international orders — here's the new policy document, please update how you handle these inquiries" The advantages are speed and accessibility: changes that previously required a development cycle (write requirement, build rule, test, deploy) can be made in a conversation. CX teams can respond to policy changes, new product launches, or emerging issues in minutes rather than days. The challenge is precision. Natural language is inherently ambiguous, so conversational configuration systems need mechanisms to confirm interpretation, test changes before they go live, and track what was configured and by whom. The best implementations combine the accessibility of conversational input with the rigor of formal testing and audit trails. **Related terms:** [AI agent](https://www.lorikeetcx.ai/glossary/ai-agent), [AI guardrails](https://www.lorikeetcx.ai/glossary/ai-guardrails), [knowledge base](https://www.lorikeetcx.ai/glossary/knowledge-base) ## Conversational IVR URL: https://www.lorikeetcx.ai/glossary/conversational-ivr **Understand what conversational IVR is, how AI-powered voice menus replace touch-tone systems, and why conversational IVR improves caller experience.** **Conversational IVR** is an interactive voice response system that uses AI and natural language understanding to allow callers to speak naturally rather than navigate touchtone menus. Traditional IVR ("Press 1 for billing, press 2 for technical support...") is universally hated. Conversational IVR replaces button-pressing with natural speech: "What can I help you with today?" The caller explains their issue in their own words, and the system routes or responds appropriately. Conversational IVR sits between traditional IVR and full AI voice agents. It handles the front-door experience—understanding why the customer called and either resolving simple requests, routing to the right queue, or collecting information before handoff to an agent. It doesn't necessarily handle complete resolution, but it eliminates menu navigation and improves triage accuracy. Implementation complexity varies. Basic conversational IVR handles common intents (billing question, account inquiry, technical issue) and routes accordingly. Advanced implementations integrate with backend systems to verify identity, pull account information, and resolve simple requests entirely within the IVR. For support operations with high phone volume, conversational IVR is often the first step in voice automation—lower risk than full voice agents, clear ROI from improved routing and reduced handle time. **Related terms:** [AI voice agent](https://www.lorikeetcx.ai/glossary/ai-voice-agent), [Intent detection](https://www.lorikeetcx.ai/glossary/intent-detection), [NLU](https://www.lorikeetcx.ai/glossary/nlu) ## Cost per Resolution URL: https://www.lorikeetcx.ai/glossary/cost-per-resolution **Learn what cost per resolution measures, how to calculate it for AI and human support, and strategies to reduce your cost per resolved ticket.** Cost per Resolution (CPR) measures the total cost to fully resolve a customer support issue, including all touches across channels and agents. It answers: "What does it actually cost us to solve a customer problem?" CPR differs from cost per ticket because one customer issue might generate multiple tickets, transfers, and follow-ups before resolution. A ticket might cost $5 to handle, but if the issue requires three tickets to resolve, the CPR is $15 plus the overhead of re-explaining context. Calculating CPR requires tracking resolution at the issue level, not the ticket level. This means linking related tickets, accounting for reopens, and attributing costs across the full resolution journey. Components typically include: agent time (fully loaded cost), technology costs (per-interaction AI costs, platform fees), and quality costs (rework, escalations, supervisor time). For AI support, CPR is the ROI metric. If AI resolves issues at $0.50 CPR versus $8 for human agents, the business case is clear—as long as the AI resolutions actually stick. A low CPR with high reopen rates means you're paying twice. **Related terms:** [Cost per Ticket](https://www.lorikeetcx.ai/glossary/cost-per-ticket), [Effective Automation Rate](https://www.lorikeetcx.ai/glossary/effective-automation-rate), [Resolution Rate](https://www.lorikeetcx.ai/glossary/resolution-rate) ## Cost per resolution (CPR) URL: https://www.lorikeetcx.ai/glossary/cost-per-resolution-cpr **Discover how cost per resolution (CPR) is calculated, why it matters for support ROI, and how AI-driven automation lowers CPR significantly.** Cost per resolution (CPR) measures the total cost a business incurs to fully resolve a single customer issue. Unlike cost-per-contact or cost-per-ticket, CPR focuses on the outcome — a resolved issue — rather than the activity. **CPR = Total support costs / Number of resolved issues** Total support costs include agent salaries and benefits, technology costs (CCaaS platform, ticketing system, AI tools), training, quality assurance, management overhead, and any outsourcing fees. The denominator — resolved issues — requires a clear definition of what constitutes "resolution" (customer confirmed satisfied, no reopen within 7 days, etc.). CPR varies significantly by channel, complexity, and resolution method: | Channel | Typical CPR | |---------|------------| | AI-resolved (chat/email) | $0.50 – $2.00 | | Human agent (chat) | $5 – $12 | | Human agent (phone) | $8 – $25 | | BPO agent (offshore) | $3 – $8 | | Escalated/multi-touch | $15 – $50+ | The strategic value of CPR is that it connects service quality to unit economics. A low CPR achieved through deflection (bouncing customers to self-service that doesn't actually resolve their issue) is illusory — those customers either come back (increasing total cost) or churn (increasing customer acquisition cost). A genuinely low CPR, achieved through efficient resolution, is one of the most powerful financial levers in a customer service operation. CPR is also the natural metric for evaluating AI agent ROI. If AI can resolve issues at $1 that previously cost $12 through human agents, the business case is immediate and measurable — provided the AI is actually resolving issues, not just deflecting them. **Related terms:** [resolution rate](https://www.lorikeetcx.ai/glossary/resolution-rate), [deflection rate](https://www.lorikeetcx.ai/glossary/deflection-rate), [first contact resolution](https://www.lorikeetcx.ai/glossary/first-contact-resolution), [per-resolution pricing](https://www.lorikeetcx.ai/glossary/per-resolution-pricing) ## Cost per Ticket URL: https://www.lorikeetcx.ai/glossary/cost-per-ticket **Understand what cost per ticket measures, how to calculate it across channels, and proven strategies to reduce support costs without losing quality.** Cost per Ticket measures the average cost to handle a single support ticket, regardless of whether it resolves the customer's issue. It's the most common cost metric in support operations and the easiest to calculate. The basic formula divides total support costs by ticket volume. Total costs include agent compensation (fully loaded with benefits and overhead), technology costs, facilities, and management. More granular calculations allocate costs by channel: phone tickets cost more than chat, which costs more than email. Cost per ticket is useful for budgeting and capacity planning but misleading for efficiency analysis. A team could reduce cost per ticket by closing tickets faster—even if customers have to open new tickets because their issues weren't resolved. Cost per ticket measures transaction efficiency, not customer efficiency. For AI support, cost per ticket comparisons show dramatic differences: AI-handled tickets might cost $0.20-0.50 versus $5-15 for human-handled tickets. But this only matters if the AI tickets are actually resolved. Comparing cost per ticket without comparing resolution quality is comparing apples to rotten apples. **Related terms:** [Cost per Resolution](https://www.lorikeetcx.ai/glossary/cost-per-resolution), [Resolution Rate](https://www.lorikeetcx.ai/glossary/resolution-rate), [Automation Rate](https://www.lorikeetcx.ai/glossary/automation-rate), [Effective Automation Rate](https://www.lorikeetcx.ai/glossary/effective-automation-rate) ## Cost to Collect URL: https://www.lorikeetcx.ai/glossary/cost-to-collect **Learn what cost to collect measures in healthcare revenue cycle, how to calculate it, and strategies to reduce collection costs.** **Cost to collect** is the total expense of revenue cycle operations divided by total collections, expressed as a percentage or dollar amount per dollar collected. This metric captures the efficiency of your revenue cycle. A cost to collect of 4% means you're spending four cents to collect every dollar. Industry benchmarks for well-run operations range from 3-5%; above 7% indicates significant optimization opportunities. Telehealth revenue cycle costs can be higher than traditional practices due to more complex claim requirements, higher patient responsibility portions requiring more collection effort, and lower average reimbursement per visit. However, telehealth also offers automation opportunities—AI-powered coding assistance, automated eligibility verification, digital patient payment collection—that can drive costs below in-person benchmarks when implemented well. Reduce cost to collect by improving clean claim rate (less rework), collecting patient payments at time of service (less billing effort), and automating repetitive revenue cycle tasks. **Related terms:** [Net collection rate](https://www.lorikeetcx.ai/glossary/net-collection-rate), [Patient payment collection rate](https://www.lorikeetcx.ai/glossary/patient-payment-collection-rate), [Clean claim rate](https://www.lorikeetcx.ai/glossary/clean-claim-rate) ## Credit Consumption URL: https://www.lorikeetcx.ai/glossary/credit-consumption **Discover what credit consumption means in AI platforms, how usage-based pricing works, and how to optimize AI credit spend for support operations.** **Credit consumption** is the metric that tracks how many AI processing units a customer support automation platform uses when handling tickets, typically measured per interaction or per task completed. Unlike seat-based pricing, credit consumption aligns costs with actual usage. When an AI agent resolves a ticket, it consumes credits based on the complexity of the reasoning required—a simple FAQ lookup uses fewer credits than a multi-step order modification that requires API calls and policy evaluation. The shift to credit-based models reflects a fundamental truth about AI support: value comes from resolution, not availability. A traditional help desk charges per agent seat regardless of whether that agent handles 10 tickets or 100. Credit consumption flips this, making costs proportional to work performed. For high-volume support teams, this creates predictable unit economics—you know exactly what each resolved ticket costs. Watch for hidden credit multipliers. Some platforms charge extra credits for tool usage, knowledge retrieval, or handoffs. The best models are transparent: one credit equals one complete interaction, regardless of what happens under the hood. **Related terms:** [Agent Handle Time](https://www.lorikeetcx.ai/glossary/agent-handle-time), [Ticket Complexity](https://www.lorikeetcx.ai/glossary/ticket-complexity), [Warm Handoff](https://www.lorikeetcx.ai/glossary/warm-handoff) ## Customer churn rate URL: https://www.lorikeetcx.ai/glossary/customer-churn-rate **Learn what customer churn rate is, how to calculate it, and proven retention strategies to reduce churn and increase customer lifetime value.** Customer churn rate measures the percentage of customers who stop doing business with a company over a given period. It is one of the most critical health metrics for subscription and recurring-revenue businesses. **Churn rate = (Customers lost during period / Customers at start of period) x 100** In B2C businesses, churn is often driven by poor customer experience — and customer service interactions are one of the most direct levers. Research consistently shows that customer service is both the leading cause of churn (when it's bad) and one of the strongest retention tools (when it's good). A single unresolved issue can trigger a cancellation; a single exceptional interaction can save one. The relationship between customer service quality and churn is not linear. There's a threshold effect: adequate service prevents churn but doesn't build loyalty. Exceptional service — resolving issues quickly, proactively addressing problems, making customers feel heard — actively reduces churn below baseline rates. - For CX leaders, the connection to AI is significant. AI agents that genuinely resolve issues (rather than deflecting them) can reduce churn by: - Providing 24/7 availability, eliminating the frustration of waiting for business hours - Resolving issues in the first interaction more consistently than human agents - Identifying at-risk customers through interaction patterns and triggering retention interventions - Maintaining consistent quality across all interactions regardless of volume spikes The counterpoint: poorly implemented AI can increase churn. Customers forced through unhelpful chatbot loops, unable to reach a human, or given incorrect information will churn faster than if they'd waited in a phone queue. The implementation quality matters as much as the technology choice. **Related terms:** [customer retention rate](https://www.lorikeetcx.ai/glossary/customer-retention-rate), [customer lifetime value](https://www.lorikeetcx.ai/glossary/customer-lifetime-value), [customer satisfaction score](https://www.lorikeetcx.ai/glossary/customer-satisfaction-score) ## Customer effort score (CES) URL: https://www.lorikeetcx.ai/glossary/customer-effort-score-ces **Understand what customer effort score (CES) measures, how to survey for it, and why reducing effort drives loyalty and repeat business.** Customer effort score (CES) measures how easy it is for a customer to get their issue resolved. Typically captured through a post-interaction survey asking "How easy was it to handle your issue?" on a scale of 1-7 (or 1-5), CES reflects the friction a customer experiences — not just whether they're satisfied, but how much work they had to do to get there. CES was introduced by the Corporate Executive Board (now Gartner) based on research showing that reducing customer effort is a stronger driver of loyalty than exceeding expectations. The finding challenged the prevailing "delight" strategy — it turns out that making things easy matters more than making things impressive. CES is particularly valuable for evaluating AI customer service implementations because it captures something CSAT misses. A customer might rate an interaction as "satisfactory" (decent CSAT) while noting it required three attempts, a channel switch, and 45 minutes of total effort. CES catches that gap. - High-effort experiences that CES detects include: - Having to repeat information across channels or agents - Being transferred multiple times before reaching someone who can help - Needing to follow up on an unresolved issue - Navigating complex self-service systems that don't lead to resolution - Being forced through AI chatbot loops before reaching a human For AI-automated customer service, CES is arguably the most important quality metric. The promise of AI is frictionless resolution — customer describes their issue, AI handles it, done. If the AI creates friction (asking redundant questions, failing to understand intent, requiring customers to start over), it defeats the purpose regardless of whether it technically "resolves" the ticket. **Related terms:** [customer satisfaction score](https://www.lorikeetcx.ai/glossary/customer-satisfaction-score), [Net Promoter Score](https://www.lorikeetcx.ai/glossary/net-promoter-score), [first contact resolution](https://www.lorikeetcx.ai/glossary/first-contact-resolution) ## Customer Health Score URL: https://www.lorikeetcx.ai/glossary/customer-health-score **Learn what a customer health score is, how to build a scoring model, and why health scores predict churn and expansion opportunities.** Customer Health Score is a composite metric combining multiple signals—product usage, support interactions, billing status, engagement patterns—into a single indicator predicting whether a customer is likely to renew, expand, or churn. Health scores attempt to answer the question CSAT and NPS can't: what is this customer actually going to do? A customer might report high satisfaction but show declining usage. Another might complain constantly but expand every year. Behavior matters more than sentiment. There's no standard formula—every company builds their own based on what predicts outcomes in their business. Common inputs include: login frequency, feature adoption depth, support ticket volume and sentiment, billing delinquency, stakeholder engagement, and NPS/CSAT trends. Weight the inputs based on which correlate most strongly with renewal in your historical data. The risk with health scores is false precision. A score of 73 vs 75 is meaningless if your model isn't validated. Start simple: red/yellow/green based on 3-5 signals you know matter. Validate against actual churn. Refine over time. The goal is to trigger intervention before at-risk customers leave, not to create a perfect predictive model. **Related terms:** [Customer Churn Rate](https://www.lorikeetcx.ai/glossary/customer-churn-rate), [Lifetime Value (LTV)](https://www.lorikeetcx.ai/glossary/lifetime-value-ltv), [Net Promoter Score (NPS)](https://www.lorikeetcx.ai/glossary/net-promoter-score-nps) ## Customer journey mapping URL: https://www.lorikeetcx.ai/glossary/customer-journey-mapping **Discover what customer journey mapping is, how to create journey maps, and why mapping touchpoints reveals friction and improvement opportunities.** Customer journey mapping is the practice of documenting and visualizing every interaction a customer has with a business — from initial awareness through purchase, onboarding, ongoing usage, support, and renewal or churn. It identifies touchpoints, pain points, emotional states, and opportunities across the full lifecycle. In the context of customer service and AI, journey mapping serves two critical functions: First, it reveals where automation will have the most impact. Not all customer interactions are equal — some are high-volume and process-driven (password resets, order tracking), while others are complex, emotional, or high-stakes (billing disputes, medical questions, insurance claims). Journey mapping helps teams prioritize which interactions to automate first and which require human handling. Second, it exposes gaps between how a company organizes internally (marketing, sales, support, success as separate departments) and how customers experience the relationship (one continuous journey). These gaps create friction: a customer who just purchased shouldn't have to re-explain their situation to the support team. A customer nearing renewal shouldn't be treated like a first-time caller. Modern journey mapping is increasingly data-driven rather than workshop-driven. Instead of guessing the customer journey in a whiteboard session, teams analyze actual interaction data — ticket volumes by stage, channel switching patterns, time between touchpoints, and resolution outcomes — to build an evidence-based map. The most forward-thinking CX organizations use journey maps to design AI interactions that span the full lifecycle — not just reactive support, but proactive onboarding assistance, usage-based recommendations, and retention interventions — creating a coherent experience regardless of which department "owns" each stage. **Related terms:** [AI concierge](https://www.lorikeetcx.ai/glossary/ai-concierge), [customer lifetime value](https://www.lorikeetcx.ai/glossary/customer-lifetime-value), [omnichannel customer support](https://www.lorikeetcx.ai/glossary/omnichannel-customer-support) ## Customer lifetime value (CLV) URL: https://www.lorikeetcx.ai/glossary/customer-lifetime-value-clv **Learn what customer lifetime value (CLV) is, how to calculate it, and why CLV should guide your customer service investment decisions.** Customer lifetime value (CLV or LTV) is the total revenue a business expects to earn from a single customer over the full duration of their relationship. It's a foundational metric for customer-centric businesses because it quantifies the economic impact of retention, satisfaction, and experience quality. **CLV = Average revenue per customer × Average customer lifespan** More sophisticated models account for gross margin, discount rates, expansion revenue, and the probability of retention at each period. For subscription businesses, CLV is directly linked to churn rate: reducing monthly churn from 5% to 3% can nearly double CLV. The connection to customer service is direct and measurable. Every customer service interaction is either building or eroding CLV: - **Positive resolution** extends the customer relationship, increases likelihood of expansion, and generates referrals - **Poor experience** accelerates churn, reduces expansion, and creates negative word-of-mouth - **Proactive service** (catching and resolving issues before the customer contacts you) has the highest CLV impact because it signals the company is paying attention For CX leaders, CLV reframes the cost conversation around customer service. When a CEO asks "why are we spending $X on support?", CLV provides the answer: because each resolved interaction protects $Y in future revenue. This is especially powerful when evaluating AI investments — if AI resolution improves CSAT and reduces churn, the CLV impact often exceeds the direct cost savings by an order of magnitude. **Related terms:** [customer churn rate](https://www.lorikeetcx.ai/glossary/customer-churn-rate), [customer retention rate](https://www.lorikeetcx.ai/glossary/customer-retention-rate), [cost per resolution](https://www.lorikeetcx.ai/glossary/cost-per-resolution) ## Customer onboarding URL: https://www.lorikeetcx.ai/glossary/customer-onboarding **Understand what customer onboarding is, best practices for onboarding flows, and how great onboarding reduces churn and accelerates time to value.** Customer onboarding is the process of guiding new customers from purchase to successful product adoption. It encompasses account setup, initial configuration, training, first-value delivery, and early engagement — the period that most strongly predicts long-term retention and expansion. Onboarding is a critical customer lifecycle stage because it's where expectations set during sales are either met or broken. Research from Wyzowl shows that 86% of customers say they'd be more loyal to a company that invests in onboarding content. Conversely, poor onboarding is among the top reasons customers churn within the first 90 days. In the context of AI customer service, onboarding represents both a challenge and an opportunity: **Challenge:** Onboarding interactions are often complex, sequential, and vary significantly by customer segment. A customer setting up a financial product needs different guidance than one configuring a software tool. Generic onboarding flows frustrate customers who need tailored guidance. **Opportunity:** AI agents can deliver personalized, adaptive onboarding at scale — checking in with customers at the right moments, answering questions in context, detecting when a customer is stuck, and adjusting the flow based on the customer's progress. This is proactive service at its most impactful. Effective AI-assisted onboarding moves beyond reactive support (answering questions when customers get stuck) to proactive guidance (reaching out when usage patterns suggest the customer hasn't completed critical setup steps, or when a customer hasn't logged in after signing up). For businesses with complex products or regulated onboarding processes (financial account opening, healthcare enrollment), AI onboarding agents can ensure every required step is completed, every disclosure is delivered, and every verification is performed — consistently, at scale. **Related terms:** [AI concierge](https://www.lorikeetcx.ai/glossary/ai-concierge), [proactive customer service](https://www.lorikeetcx.ai/glossary/proactive-customer-service), [customer lifetime value](https://www.lorikeetcx.ai/glossary/customer-lifetime-value) ## Customer retention rate URL: https://www.lorikeetcx.ai/glossary/customer-retention-rate **Learn what customer retention rate is, how to calculate it, and proven strategies to improve retention through better customer experience.** Customer retention rate measures the percentage of customers a business retains over a specific period. It's the inverse of churn and one of the most important metrics for subscription and recurring-revenue businesses. **Retention rate = ((Customers at end of period - New customers acquired) / Customers at start of period) x 100** A 95% monthly retention rate might sound strong, but it compounds to only 54% annual retention — meaning nearly half of customers leave each year. This math is why small improvements in retention have outsized business impact. Customer service is one of the most direct levers for retention. Research from Bain & Company and others consistently shows that customers who contact support and have a good experience are actually more loyal than customers who never had an issue — the "service recovery paradox." The key word is "good experience." A bad support interaction is one of the strongest predictors of churn. For CX teams deploying AI, retention rate is the north star metric. AI that resolves issues quickly and completely improves retention. AI that frustrates customers through unhelpful loops, inaccurate information, or inability to reach a human hurts retention. The technology is neutral — the implementation determines the outcome. Retention is also where the value of proactive customer service becomes visible. Detecting early warning signals (decreased usage, unresolved complaints, billing friction) and intervening before the customer decides to leave is significantly more effective than reacting after a cancellation request. AI systems with customer lifecycle visibility can identify and act on these signals at scale. **Related terms:** [customer churn rate](https://www.lorikeetcx.ai/glossary/customer-churn-rate), [customer lifetime value](https://www.lorikeetcx.ai/glossary/customer-lifetime-value), [customer satisfaction score](https://www.lorikeetcx.ai/glossary/customer-satisfaction-score) ## Customer satisfaction score (CSAT) URL: https://www.lorikeetcx.ai/glossary/customer-satisfaction-score-csat **Discover what CSAT is, how to measure customer satisfaction, and why CSAT remains one of the most important CX metrics for support teams.** Customer satisfaction score (CSAT) measures how satisfied a customer is with a specific interaction, product, or overall experience. It's typically captured through a post-interaction survey: "How satisfied were you with your experience?" on a 1-5 scale, with results reported as the percentage of respondents who selected 4 (satisfied) or 5 (very satisfied). **CSAT = (Number of satisfied responses / Total responses) x 100** CSAT is the most widely used customer service quality metric, and for good reason — it's simple, intuitive, and directly captures the customer's perspective. However, it has well-known limitations: - **Response bias:** Typically only 10-30% of customers respond, and respondents tend to skew toward extremes (very satisfied or very dissatisfied) - **Recency bias:** The score reflects the last moment of the interaction, not the overall experience - **Inflated baselines:** A CSAT of 85% is often reported as "good," but since only engaged customers respond, the true satisfaction rate across all customers may be significantly lower CSAT is most valuable when tracked as a trend over time and segmented by channel, topic, agent, or customer segment. A drop in CSAT for a specific issue category signals a systematic problem. A gap in CSAT between human-handled and AI-handled interactions tells you whether AI is meeting customer expectations. For AI customer service deployments, CSAT is the accountability metric. If AI resolution rates are climbing but CSAT is falling, the AI is closing tickets without actually satisfying customers. The metrics need to move together. **Related terms:** [Net Promoter Score](https://www.lorikeetcx.ai/glossary/net-promoter-score), [customer effort score](https://www.lorikeetcx.ai/glossary/customer-effort-score), [resolution rate](https://www.lorikeetcx.ai/glossary/resolution-rate) ## Days in Accounts Receivable (Days in AR) URL: https://www.lorikeetcx.ai/glossary/days-in-accounts-receivable-days-in-ar **Learn what days in accounts receivable measures in healthcare, how to calculate it, and strategies to accelerate payment collection.** **Days in Accounts Receivable** is the average number of days between the date of service and when payment is collected, measuring how quickly your organization converts completed visits into cash. For telehealth organizations, Days in AR often runs higher than traditional practices due to payer unfamiliarity with telehealth codes, place-of-service complexities, and higher rates of patient self-pay requiring collection efforts. A healthy telehealth operation targets 30-40 days; above 50 days signals billing process problems or payer mix issues requiring attention. This metric directly impacts cash flow and your ability to fund operations. High Days in AR means you're essentially providing free financing to payers and patients. Calculate it monthly: (Total AR ÷ Average Daily Charges) gives you the number. Watch the trend, not just the absolute number. Increasing Days in AR month-over-month often signals emerging problems with a specific payer, new claim denial patterns, or staffing issues in your billing operation before those problems become obvious elsewhere. **Related terms:** [AR aging (120+ days)](https://www.lorikeetcx.ai/glossary/ar-aging-120-days), [Net collection rate](https://www.lorikeetcx.ai/glossary/net-collection-rate), [Reimbursement cycle time](https://www.lorikeetcx.ai/glossary/reimbursement-cycle-time) ## Deflection rate URL: https://www.lorikeetcx.ai/glossary/deflection-rate **Understand what deflection rate measures in customer service, how to calculate it, and why higher deflection means more effective self-service.** Deflection rate measures the percentage of customer inquiries that are redirected away from human agents — typically to self-service resources, FAQ pages, or automated responses. It's one of the most commonly cited metrics in customer service automation, and one of the most misleading. **Deflection rate = (Inquiries handled without a human agent / Total inquiries) x 100** The problem with deflection rate as a success metric is that it measures activity, not outcomes. A customer who is shown a help article and gives up is counted as "deflected" — even though their issue is unresolved and their experience was negative. A customer who is routed through a chatbot loop and eventually abandons the conversation is "deflected." A customer who calls in after being deflected creates a new interaction that may also be deflected. This is why leading CX organizations are shifting from deflection to resolution as their primary automation metric. The question isn't "did we keep the customer away from a human?" but "did we actually solve their problem?" - The distinction has real business consequences. High deflection rates often mask: - **Hidden demand:** Customers who need help but can't get it, leading to churn - **Channel shifting:** Customers deflected from chat who call in (or vice versa), inflating total contact volume - **Repeat contacts:** Customers who come back because their issue wasn't actually resolved AI vendors who lead with deflection rate as their primary metric are often optimizing for the wrong thing. The more meaningful metrics are resolution rate (was the issue solved?) and customer effort score (how hard did the customer have to work?). **Related terms:** [resolution rate](https://www.lorikeetcx.ai/glossary/resolution-rate), [self-service rate](https://www.lorikeetcx.ai/glossary/self-service-rate), [customer effort score](https://www.lorikeetcx.ai/glossary/customer-effort-score), [cost per resolution](https://www.lorikeetcx.ai/glossary/cost-per-resolution) ## Effective Automation Rate URL: https://www.lorikeetcx.ai/glossary/effective-automation-rate **Learn what effective automation rate measures, how it differs from basic automation rate, and why quality-adjusted metrics matter for AI support.** Effective Automation Rate (EAR) measures the percentage of support tickets fully resolved by AI without human intervention and without a quality failure. It answers the question: "How much of my ticket volume is AI actually handling well?" Most vendors report automation rate or deflection rate, which count any ticket touched by AI—including tickets where the AI failed, hallucinated, or frustrated the customer into abandoning the conversation. EAR is more rigorous. It only counts tickets where three conditions are met: the AI resolved the issue, the customer didn't return within a defined window (typically 72 hours), and the resolution passed quality checks. Calculating EAR requires defining what counts as "independently resolved." At minimum, this means: (1) the ticket was closed without human involvement, (2) no reopens within the lookback window, and (3) no negative signals like CSAT bombs or escalation requests. Some teams add QA sampling to catch hallucinations that customers didn't catch. EAR is the metric that matters for ROI conversations. If you're paying for AI support and only tracking automation rate, you're measuring activity, not outcomes. A 60% automation rate with a 30% quality failure rate means you're only getting 42% effective automation—and burning customer trust on the rest. **Related terms:** [Independently Resolved (IR)](https://www.lorikeetcx.ai/glossary/independently-resolved-ir), [Resolution Rate](https://www.lorikeetcx.ai/glossary/resolution-rate), [Automation Rate](https://www.lorikeetcx.ai/glossary/automation-rate), [Total Quality Score (TQS)](https://www.lorikeetcx.ai/glossary/total-quality-score-tqs) ## Error Rate / Hallucination Rate URL: https://www.lorikeetcx.ai/glossary/error-rate-hallucination-rate **Discover what error rate and hallucination rate measure in AI, how to track them, and strategies to minimize AI inaccuracies in support.** Error Rate measures the percentage of AI responses containing factually incorrect information, policy violations, or fabricated details. Hallucination Rate specifically refers to errors where the AI confidently stated false information as if it were true. These metrics require active QA. Unlike CSAT or resolution rate, errors don't surface automatically—customers often accept confident-sounding wrong answers. Measuring error rate means sampling AI conversations and checking responses against source data, policies, and known facts. Common error categories include: wrong product information, incorrect policy statements, fabricated order statuses, promised actions the AI can't perform, and invented features or capabilities. Hallucinations are errors where the AI generated plausible-sounding but entirely false information, like a tracking number that doesn't exist. Error rate is the safety metric for AI support. A 5% error rate means 1 in 20 customers receives wrong information. For regulated industries—fintech, healthcare, insurance—this creates compliance risk. For any business, errors erode trust faster than good interactions build it. **Related terms:** [Total Quality Score (TQS)](https://www.lorikeetcx.ai/glossary/total-quality-score-tqs), [Effective Automation Rate](https://www.lorikeetcx.ai/glossary/effective-automation-rate), [Success Rate](https://www.lorikeetcx.ai/glossary/success-rate) ## Escalation rate URL: https://www.lorikeetcx.ai/glossary/escalation-rate **Learn what escalation rate measures, how to calculate it, and why monitoring escalations reveals gaps in your AI and self-service coverage.** Escalation rate measures the percentage of customer interactions that are transferred from one level of support to another — typically from AI or frontline agents to specialized or senior agents. It's a key indicator of how well the first point of contact can handle incoming volume. **Escalation rate = (Escalated interactions / Total interactions) x 100** A healthy escalation rate depends on context. For an AI-first support model, some escalation is expected and desirable — the AI should escalate when it encounters situations outside its competence, when the customer requests a human, or when the conversation involves sensitive judgment calls. An escalation rate of zero would indicate either trivial ticket volume or an AI that's overstepping its boundaries. What matters more than the raw rate is the quality of escalation: - **Clean handoff:** Does the escalated agent receive the full context of the AI conversation, or does the customer have to start over? - **Appropriate triggers:** Is the AI escalating for the right reasons (genuine complexity, customer distress, policy exceptions) or for the wrong ones (knowledge gaps, integration failures)? - **Resolution at escalation:** Are escalated interactions resolved on the first human touch, or do they bounce further? For CX leaders managing AI deployments, tracking escalation rate over time reveals the AI's learning trajectory. A decreasing escalation rate (with stable CSAT) indicates the AI is getting better at handling more complex cases. A stable rate despite expanding AI scope suggests the AI is being deployed responsibly. A rising rate despite stable scope signals a problem. **Related terms:** [first contact resolution](https://www.lorikeetcx.ai/glossary/first-contact-resolution), [AI guardrails](https://www.lorikeetcx.ai/glossary/ai-guardrails), [human-in-the-loop](https://www.lorikeetcx.ai/glossary/human-in-the-loop) ## Fallback Intent URL: https://www.lorikeetcx.ai/glossary/fallback-intent **Understand what a fallback intent is in conversational AI, when it triggers, and how to design effective fallback experiences for customers.** **Fallback intent** is the default handling path triggered when an AI system cannot confidently classify a user's request into any defined intent category. Every AI system needs a plan for "I don't know what you're asking." Fallback intent is that plan—the response and routing when confidence is low or the request doesn't match known patterns. Good fallback design is critical to customer experience; bad fallback design creates frustration spirals where customers repeat themselves to an uncomprehending bot. Fallback strategies include: clarification (asking the user to rephrase or provide more detail), suggestion (offering likely intents based on partial understanding), escalation (routing to a human agent), and graceful acknowledgment (admitting confusion without making it the user's problem). The best approach depends on context—high-confidence near-misses warrant clarification, while complete confusion warrants fast escalation. Fallback rate is a key health metric for conversational AI. A high fallback rate signals either narrow intent coverage (the AI doesn't handle what customers actually ask) or weak intent recognition (the AI can't understand how customers phrase known requests). Both require different remediation. Monitor fallback patterns to identify gaps: what are customers asking that the AI can't handle? **Related terms:** [Intent detection](https://www.lorikeetcx.ai/glossary/intent-detection), [Intent recognition](https://www.lorikeetcx.ai/glossary/intent-recognition), [Conversational AI design](https://www.lorikeetcx.ai/glossary/conversational-ai-design) ## False Positive Rate (Fraud) URL: https://www.lorikeetcx.ai/glossary/false-positive-rate-fraud **Learn what false positive rate means in fraud detection, how it impacts customer experience, and strategies to reduce false fraud alerts.** False Positive Rate (Fraud) is the percentage of legitimate transactions that your fraud system incorrectly flags or blocks as fraudulent. Calculate it as: (Legitimate transactions flagged as fraud / Total legitimate transactions) × 100. Industry benchmarks vary wildly—aggressive systems see 5-10% false positive rates, while customer-friendly systems target under 1%. False positives are where fraud prevention becomes a CX problem. Every declined legitimate transaction is a customer calling your support team, a cart abandonment, or a churned user. At scale, a 3% false positive rate on 1 million monthly transactions means 30,000 frustrated customers. Most of those customers won't call—they'll just leave. The real cost isn't the support ticket. It's the lifetime value of customers who stop trusting your platform after a false decline at checkout or a frozen account during a legitimate purchase. Sophisticated fraud teams track false positive rates by customer segment, transaction size, and time of day to identify where their rules are too aggressive without exposing actual vulnerabilities. **Related terms:** [Fraud Detection Rate](https://www.lorikeetcx.ai/glossary/fraud-detection-rate), [Customer Effort Score](https://www.lorikeetcx.ai/glossary/customer-effort-score), [Transaction Approval Rate](https://www.lorikeetcx.ai/glossary/transaction-approval-rate) ## First contact resolution (FCR) URL: https://www.lorikeetcx.ai/glossary/first-contact-resolution-fcr **Discover what first contact resolution (FCR) is, how to measure it, and why resolving issues on first contact drives satisfaction and efficiency.** First contact resolution (FCR) measures the percentage of customer issues that are fully resolved during the customer's first interaction — no follow-ups, callbacks, or transfers needed. **FCR = (Issues resolved on first contact / Total issues) x 100** FCR is widely considered the single most important quality metric in customer service. Research consistently links high FCR to higher customer satisfaction, lower operating costs, and reduced churn. Each percentage point improvement in FCR reduces repeat contacts, frees agent capacity, and improves the customer experience. FCR directly addresses what customers care about most: getting their problem solved without unnecessary effort. A customer who resolves their issue in one five-minute interaction is significantly more satisfied than one who resolves it across three two-minute interactions — even though the total time spent is shorter in the second scenario. - Measuring FCR accurately requires defining "resolved" carefully: - **Customer-confirmed:** The customer explicitly states the issue is resolved (most accurate, hardest to capture) - **No recontact:** The customer doesn't contact support about the same issue within a defined window (7, 14, or 30 days) - **Agent-marked:** The agent flags the interaction as resolved (easiest to capture, least reliable) For AI-powered customer service, FCR is the metric that separates genuine resolution from deflection. An AI system that achieves 80% FCR is delivering real value — it's solving customer problems end-to-end. An AI system that achieves 80% "deflection" may be closing tickets without resolving issues, pushing customers to multiple contacts to get help. **Related terms:** [resolution rate](https://www.lorikeetcx.ai/glossary/resolution-rate), [customer effort score](https://www.lorikeetcx.ai/glossary/customer-effort-score), [average handling time](https://www.lorikeetcx.ai/glossary/average-handling-time) ## First Notice of Loss (FNOL) Response Time URL: https://www.lorikeetcx.ai/glossary/first-notice-of-loss-fnol-response-time **Learn what FNOL response time measures in insurance, why fast first contact matters, and how AI accelerates claims intake.** **First Notice of Loss (FNOL) response time** is the elapsed time between when a policyholder first reports a claim and when the insurer acknowledges receipt and initiates the claims process. FNOL response time sets the tone for the entire claims experience. Research consistently shows that faster initial response correlates with higher policyholder satisfaction and lower litigation rates. The industry benchmark is under 15 minutes for acknowledgment, though many carriers still operate on a 24-48 hour response window. What matters more than the acknowledgment itself is what happens during it. A quick auto-reply that says "we received your claim" does nothing to reduce policyholder anxiety. Effective FNOL response includes confirming coverage eligibility, setting expectations for next steps, and—where possible—beginning triage automatically. Insurers using AI-powered claims intake can now gather supplementary information, schedule adjusters, and even issue partial payments within that first interaction. The metric loses meaning if you measure acknowledgment without measuring quality. Track time-to-meaningful-response, not time-to-auto-reply. **Related terms:** [Claims triage](https://www.lorikeetcx.ai/glossary/claims-triage), [Claims settlement cycle time](https://www.lorikeetcx.ai/glossary/claims-settlement-cycle-time), [Straight-through processing rate](https://www.lorikeetcx.ai/glossary/straight-through-processing-rate) ## First Reply Time URL: https://www.lorikeetcx.ai/glossary/first-reply-time **Understand what first reply time measures, how to calculate it, and why faster first responses improve customer satisfaction and loyalty.** **First reply time (FRT)** is the elapsed time between when a customer submits a request and when they receive the first meaningful response from your team, excluding auto-acknowledgments. FRT is the defining metric for async channels—email, chat, social, in-app messaging. Unlike phone where ASA measures queue wait, FRT captures the actual customer experience of "I asked a question, how long until someone engaged with it?" The key word is meaningful: automated "we received your message" replies don't count. Customers want evidence a human is working on their problem. For B2B fintech and enterprise insurance support, FRT expectations vary wildly by channel and severity. A 4-hour email FRT might be fine for billing questions but catastrophic for production outages. Smart teams segment FRT targets by priority tier and channel rather than chasing a single number. Watch the distribution, not just the average—a mean FRT of 2 hours hides the 10% of tickets sitting untouched for 24 hours. **Related terms:** [Average speed of answer](https://www.lorikeetcx.ai/glossary/average-speed-of-answer), [Average wait time](https://www.lorikeetcx.ai/glossary/average-wait-time), [Ticket backlog](https://www.lorikeetcx.ai/glossary/ticket-backlog) ## First response time (FRT) URL: https://www.lorikeetcx.ai/glossary/first-response-time-frt **Learn what first response time (FRT) is, how to measure it across channels, and proven strategies to reduce FRT in customer support.** First response time (FRT) measures the elapsed time between a customer submitting a support request and receiving the first meaningful response. It's a primary SLA metric for customer service operations, particularly for email and chat channels. FRT matters because customer expectations for response speed have shifted dramatically. Research shows that 90% of customers rate an "immediate" response as important when they have a support question, with "immediate" defined as within 10 minutes. For chat, the expectation is seconds, not minutes. Channels vary significantly in expected FRT: | Channel | Customer expectation | |---------|---------------------| | Live chat | Under 1 minute | | Social media | Under 1 hour | | Email | Under 4 hours | | Phone | Under 2 minutes (queue time) | AI fundamentally changes the FRT equation. AI agents can respond to chat and email inquiries in seconds, 24/7, with no queue time. This eliminates FRT as a constraint — the question shifts from "how fast can we respond?" to "how well can we resolve?" However, FRT still matters in hybrid models where AI handles initial triage and some conversations escalate to human agents. The handoff time — from AI escalation to human pickup — becomes the new FRT for those interactions. Teams that optimize AI resolution for easy cases but neglect the escalation path can inadvertently create worse FRT for customers with complex issues. The most useful FRT reporting segments by channel, complexity, and whether the interaction was AI-handled or human-handled, revealing where speed is excellent and where customers are waiting. **Related terms:** [average handling time](https://www.lorikeetcx.ai/glossary/average-handling-time), [escalation rate](https://www.lorikeetcx.ai/glossary/escalation-rate), [AI agent](https://www.lorikeetcx.ai/glossary/ai-agent) ## First-Pass Resolution Rate (Claims) URL: https://www.lorikeetcx.ai/glossary/first-pass-resolution-rate-claims **Discover what first-pass resolution rate measures in claims processing, and how to improve the percentage of claims resolved without rework.** **First-pass resolution rate** is the percentage of insurance claims that are paid correctly on initial submission without requiring appeal, resubmission, or additional documentation. This metric goes beyond clean claim rate by measuring not just whether a claim was accepted, but whether it was paid correctly the first time. A claim might be "clean" but paid at the wrong rate or denied for medical necessity—that's a first-pass failure. For telehealth, first-pass resolution rates typically run 5-10 percentage points lower than in-person claims due to payer processing complexity around telehealth services. Target 85%+ first-pass resolution; below 75% means your revenue cycle team is spending most of their time on rework rather than proactive improvements. Track this by payer to identify which relationships need attention. Some payers process telehealth claims smoothly; others systematically underpay or reject claims requiring appeals. That intelligence should inform both your contracting strategy and your billing workflow design. **Related terms:** [Clean claim rate](https://www.lorikeetcx.ai/glossary/clean-claim-rate), [Claims denial rate](https://www.lorikeetcx.ai/glossary/claims-denial-rate), [Net collection rate](https://www.lorikeetcx.ai/glossary/net-collection-rate) ## Fraud Detection (Insurance) URL: https://www.lorikeetcx.ai/glossary/fraud-detection-insurance **Learn what fraud detection means in insurance, how AI identifies suspicious claims patterns, and why automated detection reduces losses.** **Fraud detection in insurance** refers to the systems and processes used to identify fraudulent claims, applications, or provider billing before payment is made. Insurance fraud costs the industry an estimated $80-100 billion annually in the US alone. It spans opportunistic padding (inflating legitimate claims by 20%) to organized rings (staged accidents, fake identities, provider collusion). Detection capability directly impacts loss ratios and, by extension, pricing competitiveness. Traditional fraud detection relied on red-flag rules and Special Investigations Unit (SIU) referrals. These catch obvious fraud but miss sophisticated schemes and generate high false positive rates. Modern approaches layer machine learning on top: anomaly detection across claim patterns, network analysis linking related parties, and predictive models trained on historical fraud outcomes. The operational challenge is balancing detection with customer experience. Aggressive fraud flagging delays legitimate claims and frustrates honest policyholders. The goal is high-confidence identification: flag the clear fraud for investigation, fast-track the clearly legitimate claims, and focus human judgment on the uncertain middle. **Related terms:** [Claims leakage](https://www.lorikeetcx.ai/glossary/claims-leakage), [Claims triage](https://www.lorikeetcx.ai/glossary/claims-triage), [Straight-through processing rate](https://www.lorikeetcx.ai/glossary/straight-through-processing-rate) ## Fraud Detection Rate URL: https://www.lorikeetcx.ai/glossary/fraud-detection-rate **Understand what fraud detection rate measures, how AI improves detection accuracy, and strategies to catch more fraud while reducing false positives.** Fraud Detection Rate is the percentage of actual fraudulent transactions that your fraud prevention system successfully identifies and blocks before they cause financial loss. Calculate it as: (Fraudulent transactions detected / Total fraudulent transactions) × 100. A 95% fraud detection rate means 5% of fraud slips through undetected. For most fintechs, acceptable rates range from 90-98% depending on transaction type and risk tolerance. The tension every CX leader faces: higher detection rates typically mean more aggressive rules, which increase false positives and create friction for legitimate customers. A 99% detection rate sounds impressive until you realize you're declining 1 in 10 good transactions. The goal isn't maximum detection—it's optimal detection balanced against customer experience. Your fraud detection rate should be measured across transaction types separately. Card-not-present transactions, account takeovers, and new account fraud each have different baseline rates and require different thresholds. Blending them into a single metric obscures where your actual vulnerabilities are. **Related terms:** [False Positive Rate (Fraud)](https://www.lorikeetcx.ai/glossary/false-positive-rate-fraud), [Chargeback Rate](https://www.lorikeetcx.ai/glossary/chargeback-rate), [Account Takeover Rate](https://www.lorikeetcx.ai/glossary/account-takeover-rate) ## Generative AI for Customer Service URL: https://www.lorikeetcx.ai/glossary/generative-ai-for-customer-service **Learn what generative AI means for customer service, how LLMs power natural support conversations, and the business impact of generative AI.** **Generative AI for customer service** refers to the application of large language models and related technologies to generate natural, contextually appropriate responses to customer inquiries—rather than retrieving pre-written templates. This is the technology shift that's reshaping customer support. Traditional chatbots selected from pre-authored responses; generative AI writes new responses for each interaction. This enables handling of novel situations, natural conversation flow, and responses personalized to context—capabilities that template-based systems fundamentally cannot achieve. The opportunity is substantial: generative AI can handle the long tail of questions that traditional systems couldn't address without a matching template. The risk is equally substantial: generative AI can hallucinate incorrect information, generate inappropriate content, or make promises outside policy. Production deployment requires robust guardrails, grounding in verified knowledge sources, and quality monitoring. For CX leaders, generative AI changes the skills required. Less time scripting responses, more time defining policies, curating knowledge, and configuring guardrails. The bottleneck shifts from content creation to content governance. **Related terms:** [AI hallucinations](https://www.lorikeetcx.ai/glossary/ai-hallucinations), [Retrieval augmented generation](https://www.lorikeetcx.ai/glossary/retrieval-augmented-generation), [Prompt engineering](https://www.lorikeetcx.ai/glossary/prompt-engineering) ## Hallucination Detection URL: https://www.lorikeetcx.ai/glossary/hallucination-detection **Discover what hallucination detection is, how AI systems identify and flag inaccurate outputs, and why detection is critical for trustworthy AI.** **Hallucination detection** is the capability to identify when an AI system generates information that is factually incorrect, fabricated, or unsupported by its knowledge sources—before that information reaches the customer. Because hallucinations are inherent to how language models work, detection becomes a critical safety layer. Effective detection approaches include: cross-referencing generated claims against source documents, confidence scoring that flags uncertain assertions, semantic consistency checks that catch contradictions, and specialized classifiers trained to identify hallucination patterns. For customer service, hallucination detection must be real-time. Catching a hallucination after the customer receives it is too late—the trust damage is done. This requires inference-time detection integrated into the response generation pipeline, not post-hoc review. Detection accuracy varies by domain. Factual claims ("your order ships tomorrow") are more verifiable than judgment calls ("this is the best option for you"). The most rigorous approaches flag anything that isn't directly grounded in retrieved sources, accepting some false positives to minimize false negatives on high-stakes information. **Related terms:** [AI hallucinations](https://www.lorikeetcx.ai/glossary/ai-hallucinations), [AI grounding](https://www.lorikeetcx.ai/glossary/ai-grounding), [Retrieval augmented generation](https://www.lorikeetcx.ai/glossary/retrieval-augmented-generation) ## Human-in-the-loop (HITL) URL: https://www.lorikeetcx.ai/glossary/human-in-the-loop-hitl **Learn what human-in-the-loop (HITL) means, how humans supervise and improve AI systems, and why HITL is essential for reliable customer service.** Human-in-the-loop (HITL) is a system design approach where human oversight is built into an AI workflow — humans review, approve, or override AI decisions at defined points rather than the AI operating fully autonomously. In customer service, HITL ensures that high-stakes or ambiguous situations receive human judgment before actions are taken. HITL operates on a spectrum: - **Full HITL:** A human reviews and approves every AI-generated response before it's sent (agent assist model) - **Selective HITL:** The AI handles routine cases autonomously but flags uncertain or high-risk cases for human review - **Oversight HITL:** The AI operates autonomously but humans periodically review a sample of interactions for quality assurance - **Escalation HITL:** The AI handles the conversation until it encounters a situation requiring human judgment, then transfers seamlessly For most customer service deployments, selective HITL is the practical sweet spot. The AI resolves the straightforward cases (password resets, order tracking, account updates) without human involvement, while complex cases (billing disputes, compliance-sensitive requests, upset customers) are routed to human agents with full context. The key design decision is where to draw the line. Too much human involvement negates the efficiency benefits of AI; too little creates risk. The best implementations define clear criteria for when HITL triggers: monetary thresholds, compliance categories, customer sentiment scores, or AI confidence levels. HITL is also the mechanism that enables AI systems to improve over time. When humans review and correct AI decisions, that feedback can be used to refine the AI's behavior — creating a virtuous cycle where the AI gradually handles more cases correctly, reducing the need for human intervention. **Related terms:** [AI guardrails](https://www.lorikeetcx.ai/glossary/ai-guardrails), [escalation rate](https://www.lorikeetcx.ai/glossary/escalation-rate), [automated quality assurance](https://www.lorikeetcx.ai/glossary/automated-quality-assurance) ## Independently Resolved (IR) URL: https://www.lorikeetcx.ai/glossary/independently-resolved-ir **Understand what independently resolved (IR) means in AI support, how to measure it, and why IR rate is a key metric for AI effectiveness.** Independently Resolved (IR) tickets are support conversations fully handled by AI from first contact to resolution without any human agent involvement. This is the numerator in most automation rate calculations. A ticket qualifies as IR when the AI handled every message in the conversation, the ticket reached a closed state, and the customer accepted the outcome. The definition sounds simple but implementation gets messy fast. Does a ticket count as IR if it was auto-closed after 24 hours of inactivity? What if the customer silently abandoned in frustration? What if the AI confidently gave wrong information that the customer didn't catch? Mature teams add guardrails to their IR definition: no reopens within 72 hours, no negative CSAT scores, no detected escalation language. Some run QA sampling on a percentage of IR tickets to catch false positives—tickets marked resolved where the AI actually failed. The gap between reported IR and true IR is where vendor metrics diverge from reality. If your AI vendor claims 70% IR but your CSAT is dropping and repeat contacts are rising, the IR definition is too loose. **Related terms:** [Effective Automation Rate](https://www.lorikeetcx.ai/glossary/effective-automation-rate), [Resolution Rate](https://www.lorikeetcx.ai/glossary/resolution-rate), [Success Rate](https://www.lorikeetcx.ai/glossary/success-rate), [Match Rate](https://www.lorikeetcx.ai/glossary/match-rate) ## Intelligent Virtual Agent (IVA) URL: https://www.lorikeetcx.ai/glossary/intelligent-virtual-agent-iva **Learn what an intelligent virtual agent (IVA) is, how IVAs handle complex customer queries, and how they differ from simple chatbots.** **Intelligent Virtual Agent (IVA)** is an AI-powered software agent that can autonomously handle customer interactions across channels, combining understanding, reasoning, and action to resolve issues without human intervention. IVA is the industry term for sophisticated conversational AI that goes beyond simple chatbots. Where a basic chatbot matches keywords and returns templates, an IVA understands context, reasons about appropriate responses, and executes actions in backend systems. The "intelligent" qualifier signals capabilities like learning from interactions, handling ambiguity, and improving over time. In practice, IVA often describes vendor positioning rather than technical architecture. The term emerged to differentiate advanced solutions from the first generation of disappointing chatbots. When evaluating IVA claims, probe specifics: Can it execute transactions or just provide information? How does it handle multi-turn conversations? What's the actual resolution rate versus containment rate? For CX leaders, IVA represents the goal state: AI that genuinely resolves customer issues, not just deflects them. Getting there requires investment in knowledge, integration, conversation design, and ongoing optimization. The technology exists; the implementation discipline separates success from expensive disappointment. **Related terms:** [Chatbot](https://www.lorikeetcx.ai/glossary/chatbot), [Conversational AI](https://www.lorikeetcx.ai/glossary/conversational-ai), [AI workflow automation](https://www.lorikeetcx.ai/glossary/ai-workflow-automation) ## Intent detection URL: https://www.lorikeetcx.ai/glossary/intent-detection **Discover what intent detection is in conversational AI, how AI identifies customer intent from messages, and why accurate detection drives resolution.** Intent detection (also called intent recognition or intent classification) is the AI capability that identifies what a customer is trying to accomplish from their message. When a customer writes "I haven't received my package and it's been two weeks," the intent detection system classifies this as a shipping/delivery inquiry, distinguishing it from a return request, a billing question, or a product complaint. Intent detection is the first step in any AI-powered customer service interaction. The accuracy of intent detection directly determines whether the AI can route the conversation correctly, retrieve relevant information, and take appropriate actions. Modern intent detection has moved beyond keyword matching and simple classification. Large language models can: - **Detect multiple intents** in a single message ("I need to return this item AND update my shipping address") - **Handle implicit intent** where the customer doesn't state their goal directly ("This is the third time I've been charged twice" implies a refund request and a systematic billing issue) - **Distinguish intent from emotion** (a frustrated customer asking about a return vs. a neutral customer asking the same thing may need different handling) - **Recognize intent shifts** mid-conversation as the customer's needs evolve For CX teams, intent detection quality is a leading indicator of overall AI performance. If the AI frequently misidentifies what customers want, everything downstream fails — wrong knowledge articles are retrieved, wrong actions are taken, and customers are sent down incorrect conversation paths. Evaluating intent detection requires looking beyond top-line accuracy rates. A system that correctly identifies 95% of intents might still fail badly on the 5% it misses — particularly if those misses cluster in high-value or sensitive categories. **Related terms:** [natural language processing](https://www.lorikeetcx.ai/glossary/natural-language-processing), [sentiment analysis](https://www.lorikeetcx.ai/glossary/sentiment-analysis), [conversational AI](https://www.lorikeetcx.ai/glossary/conversational-ai) ## Intent Recognition URL: https://www.lorikeetcx.ai/glossary/intent-recognition **Learn what intent recognition is, how NLP models classify user intents, and best practices for training accurate intent recognition systems.** **Intent recognition** is the process by which an AI system interprets user input to determine the underlying purpose or goal, often used interchangeably with intent detection. Intent recognition emphasizes the interpretation process—understanding not just keywords but meaning. A customer saying "this isn't what I ordered" requires recognizing the complaint intent, but also the implicit context (order issue), emotion (frustration), and likely need (resolution or refund). Full intent recognition captures these layers. Modern intent recognition leverages large language models that understand semantic meaning rather than just pattern matching. This enables handling of novel phrasings, implicit requests, and context-dependent interpretation. The tradeoff: LLM-based recognition is more flexible but less predictable than rules-based approaches. Testing and monitoring become critical. For CX implementation, focus on the intents that matter most: high-volume requests, high-stakes interactions, and common failure points. You don't need to recognize every possible intent—start with the 20 intents that cover 80% of volume. Expand coverage based on what falls through to fallback or human agents. Intent libraries should reflect what customers actually ask, not what you think they should ask. **Related terms:** [Intent detection](https://www.lorikeetcx.ai/glossary/intent-detection), [NLU](https://www.lorikeetcx.ai/glossary/nlu), [Natural language processing](https://www.lorikeetcx.ai/glossary/natural-language-processing) ## Interactive voice response (IVR) URL: https://www.lorikeetcx.ai/glossary/interactive-voice-response-ivr **Understand what IVR is, how interactive voice response systems route phone calls, and how modern AI is replacing traditional IVR menus.** Interactive voice response (IVR) is a telephony system that interacts with callers through voice prompts and keypad inputs to route calls, provide information, or complete simple transactions without a human agent. "Press 1 for billing, press 2 for technical support" is the classic IVR experience. Traditional IVR systems use pre-recorded prompts and touch-tone (DTMF) input to navigate callers through a decision tree. More advanced systems use speech recognition to allow callers to state their intent verbally. Despite being a decades-old technology, IVR remains the first point of contact for a significant percentage of customer service interactions. IVR is also one of the most widely disliked customer experiences. Industry data consistently shows that navigating IVR menus is among the top frustrations customers cite. Long menu trees, poor speech recognition, and the inability to reach a human agent erode customer satisfaction before the actual service interaction even begins. Conversational AI is increasingly replacing traditional IVR. Instead of rigid menu trees, AI-powered voice systems can: - Understand natural language ("I need to check on a claim I filed last week") rather than requiring menu selection - Handle the interaction directly rather than just routing to the right department - Authenticate the caller through voice biometrics rather than asking for account numbers - Maintain context if the call is transferred to a human agent The transition from IVR to AI voice agents represents one of the highest-impact automation opportunities in customer service — it affects the highest-friction touchpoint (phone) at the moment of maximum customer frustration (calling in with a problem). **Related terms:** [conversational AI](https://www.lorikeetcx.ai/glossary/conversational-ai), [AI agent](https://www.lorikeetcx.ai/glossary/ai-agent), [first response time](https://www.lorikeetcx.ai/glossary/first-response-time) ## Knowledge base URL: https://www.lorikeetcx.ai/glossary/knowledge-base **Learn what a knowledge base is, how to build one for customer self-service, and why a well-maintained knowledge base reduces support volume.** A knowledge base is a centralized repository of information that customer service agents (human or AI) use to answer customer questions and resolve issues. It typically includes product documentation, FAQs, troubleshooting guides, policy documents, and standard operating procedures. For AI-powered customer service, the knowledge base is foundational. AI agents use knowledge base content as the source of truth for generating responses — through retrieval-augmented generation (RAG), the AI retrieves relevant articles and uses them to construct accurate, grounded answers. The quality of the knowledge base directly determines the quality of AI responses. Common knowledge base challenges include: - **Staleness:** Documentation that was accurate when written but hasn't been updated to reflect product changes, policy updates, or new edge cases - **Gaps:** Topics that customers frequently ask about but aren't documented — often discovered only when AI can't find relevant content - **Conflicting information:** Multiple articles covering the same topic with inconsistent guidance, forcing AI (or agents) to guess which is correct - **Organization:** Content structured for internal convenience rather than retrieval effectiveness, making it hard for both humans and AI to find the right answer A well-maintained knowledge base is one of the highest-leverage investments a CX team can make. It improves human agent consistency, AI accuracy, and self-service effectiveness simultaneously. The return on knowledge base quality compounds: every article improved helps every future interaction on that topic. Teams deploying AI agents should treat knowledge base quality as a continuous operation, not a one-time project — regularly identifying gaps, updating outdated content, and resolving conflicts based on AI performance data and customer feedback. **Related terms:** [retrieval-augmented generation](https://www.lorikeetcx.ai/glossary/retrieval-augmented-generation), [self-service rate](https://www.lorikeetcx.ai/glossary/self-service-rate), [AI hallucinations](https://www.lorikeetcx.ai/glossary/ai-hallucinations) ## Knowledge Graph URL: https://www.lorikeetcx.ai/glossary/knowledge-graph **Discover what a knowledge graph is, how structured data relationships improve AI understanding, and why knowledge graphs power better AI answers.** **Knowledge graph** is a structured representation of information as interconnected entities and relationships, enabling AI systems to reason about connections between concepts. Knowledge graphs go beyond document storage to model how things relate. Instead of an article about "premium subscriptions" and a separate article about "refund policy," a knowledge graph captures that premium_subscription has billing_frequency of monthly, includes feature_access to [list], has associated refund_policy of policy_x, and so on. The AI can traverse these relationships to answer questions requiring synthesis across multiple concepts. For customer service, knowledge graphs enable more intelligent responses. "What's included in my plan and can I get a refund if I cancel?" requires connecting account information to plan features to cancellation policy to refund rules. A knowledge graph models these connections explicitly rather than relying on the AI to infer them from document proximity. Building knowledge graphs requires investment in information architecture and ongoing maintenance. The payoff: more accurate, consistent responses to complex questions, and easier updates (change the relationship in one place, not across multiple documents). For organizations with complex products, policies, or regulatory requirements, knowledge graphs provide the structure that document-based retrieval lacks. **Related terms:** [Knowledge base](https://www.lorikeetcx.ai/glossary/knowledge-base), [Retrieval augmented generation](https://www.lorikeetcx.ai/glossary/retrieval-augmented-generation), [AI grounding](https://www.lorikeetcx.ai/glossary/ai-grounding) ## KYC/KYB Completion Rate URL: https://www.lorikeetcx.ai/glossary/kyc-kyb-completion-rate **Learn what KYC/KYB completion rate measures in financial services, how to improve verification completion, and reduce customer drop-off.** **KYC/KYB completion rate** measures the percentage of customers who successfully finish identity verification (Know Your Customer for individuals, Know Your Business for companies) out of all who start the process. This metric sits at the intersection of compliance, product, and customer experience. A low completion rate means you're bleeding potential customers at onboarding—either because your verification flow is too friction-heavy, your document capture fails on certain devices, or your identity provider rejects legitimate users. Every point you lose here is revenue that never arrives. The best fintech CX teams treat KYC/KYB completion as a support metric, not just a product metric. They monitor where users drop off, what error messages they see, and how many reach out to support mid-flow. They also track time-to-completion: a 95% completion rate means nothing if it takes users three days and four support tickets to get there. Benchmark targets vary by customer type. Consumer fintechs typically aim for 85-95% completion rates. B2B platforms doing KYB often see 70-85% due to the complexity of business documentation. If you're significantly below these, your verification partner or flow design needs work. **Related terms:** [First contact resolution](https://www.lorikeetcx.ai/glossary/first-contact-resolution), [Customer effort score](https://www.lorikeetcx.ai/glossary/customer-effort-score), [Onboarding time](https://www.lorikeetcx.ai/glossary/onboarding-time) ## Large language model (LLM) URL: https://www.lorikeetcx.ai/glossary/large-language-model-llm **Understand what a large language model (LLM) is, how LLMs like GPT and Claude work, and why LLMs are powering the next generation of CX.** A large language model (LLM) is a type of AI model trained on vast amounts of text data that can understand, generate, and reason about natural language. Models like GPT-4, Claude, Gemini, and Llama are examples. LLMs are the core technology enabling the current generation of AI customer service agents. LLMs differ from earlier NLP approaches in a fundamental way: they don't just classify or extract information from text — they can generate contextually appropriate responses, reason through multi-step problems, and handle the ambiguity inherent in natural language. This is what makes modern AI agents capable of genuine conversations rather than scripted interactions. In customer service applications, LLMs are used for: - **Conversation handling:** Understanding customer messages and generating appropriate responses - **Reasoning:** Breaking down complex requests into steps and determining the right course of action - **Knowledge retrieval:** Understanding which information is relevant to a given query and synthesizing it into a coherent answer - **Summarization:** Condensing long conversation histories or ticket notes into concise summaries for handoffs Key considerations for CX leaders when evaluating LLM-based AI solutions: - **Accuracy vs. fluency:** LLMs can generate convincing-sounding responses that are factually wrong (hallucinations). The system architecture matters more than the model choice. - **Cost:** LLM inference has a per-token cost that scales with conversation length and volume. Pricing models should account for this. - **Latency:** Response generation time varies by model size and complexity. For real-time chat, latency matters. - **Data privacy:** Understanding how customer data flows through LLM inference, whether conversations are used for model training, and where data is processed is critical for regulated industries. **Related terms:** [AI agent](https://www.lorikeetcx.ai/glossary/ai-agent), [AI hallucinations](https://www.lorikeetcx.ai/glossary/ai-hallucinations), [natural language processing](https://www.lorikeetcx.ai/glossary/natural-language-processing) ## Lifetime Value (LTV) URL: https://www.lorikeetcx.ai/glossary/lifetime-value-ltv **Learn what lifetime value (LTV) measures, how to calculate customer LTV, and why it should guide your service and retention investments.** Lifetime Value (LTV) is the total revenue a customer generates over their entire relationship with your company, calculated as average revenue per customer multiplied by average customer lifespan. LTV is why CX investment matters. Support costs are visible and immediate—headcount, tools, overhead. Support value is invisible and delayed—reduced churn, increased expansion, referrals. LTV quantifies that value. Every point of churn you prevent extends lifespan and adds directly to LTV. Simple formula: Average Revenue Per Account (ARPA) x Average Customer Lifespan. More sophisticated models factor in gross margin, expansion revenue, and time value of money. For decision-making, the LTV:CAC ratio matters most—if it costs $500 to acquire a customer with $5,000 LTV, you can afford to invest in retention. CX teams should think in terms of LTV impact, not cost savings. A $50/hour agent spending 30 minutes to save a $5,000 LTV customer isn't expensive—it's a 400x return. Build this framing into your business cases. When justifying headcount, tools, or process improvements, connect them to churn reduction and LTV protection. That's the language the CFO speaks. **Related terms:** [Customer Churn Rate](https://www.lorikeetcx.ai/glossary/customer-churn-rate), [Customer Health Score](https://www.lorikeetcx.ai/glossary/customer-health-score), [Net Promoter Score (NPS)](https://www.lorikeetcx.ai/glossary/net-promoter-score-nps) ## Loss Adjustment Expense (LAE) Ratio URL: https://www.lorikeetcx.ai/glossary/loss-adjustment-expense-lae-ratio **Discover what loss adjustment expense ratio measures in insurance, how to calculate LAE, and strategies to reduce claims processing costs.** **Loss Adjustment Expense (LAE) ratio** is the percentage of total claims payouts spent on the cost of investigating, processing, and settling those claims—including adjuster salaries, legal fees, and vendor costs. LAE ratio answers the question: for every dollar you pay out in claims, how much are you spending to administer the process? An LAE ratio of 12% means you spend $12 investigating and processing every $100 in claims paid. The industry average hovers between 10-15%, though this varies significantly by line of business. The metric splits into two components. Allocated LAE (ALAE) includes costs tied to specific claims—independent adjuster fees, legal defense, expert witnesses. Unallocated LAE (ULAE) covers overhead that can't be attributed to individual claims—staff claims salaries, technology, facilities. Both matter, but ULAE is where automation investments show the clearest returns. Reducing LAE ratio without increasing claims leakage is the core challenge of claims operations. Cutting adjuster headcount or investigation thoroughness might lower LAE but will likely increase indemnity costs through overpayment. The insurers winning on LAE are those automating low-complexity claims while directing human expertise to complex ones. **Related terms:** [Claims leakage](https://www.lorikeetcx.ai/glossary/claims-leakage), [Straight-through processing rate](https://www.lorikeetcx.ai/glossary/straight-through-processing-rate), [Claims settlement cycle time](https://www.lorikeetcx.ai/glossary/claims-settlement-cycle-time) ## Match Rate URL: https://www.lorikeetcx.ai/glossary/match-rate **Learn what match rate measures in AI customer service, how to calculate it, and why matching accuracy determines automation effectiveness.** Match Rate measures the percentage of incoming tickets that trigger an AI workflow or automation rule. It answers: "How much of our ticket volume does our AI actually know how to handle?" A ticket "matches" when its content aligns with a defined workflow, intent, or automation rule. If you've built workflows for order status, returns, and password resets, your match rate reflects what percentage of incoming tickets fall into those categories. Match rate is a coverage metric, not a success metric. A matched ticket might still fail if the workflow is buggy, the customer provides unexpected information, or the AI misexecutes. But unmatched tickets definitely won't be automated—they go straight to human queues. Improving match rate means expanding automation coverage: adding new workflows for common ticket types, improving intent detection to catch variations, and refining rules to reduce false negatives. A 40% match rate means 60% of your volume can't even attempt automation. **Related terms:** [Success Rate](https://www.lorikeetcx.ai/glossary/success-rate), [Automation Rate](https://www.lorikeetcx.ai/glossary/automation-rate), [Natural Language Workflows (NLW)](https://www.lorikeetcx.ai/glossary/natural-language-workflows-nlw), [Resolution Rate](https://www.lorikeetcx.ai/glossary/resolution-rate) ## Model Drift URL: https://www.lorikeetcx.ai/glossary/model-drift **Understand what model drift is in AI, how production models degrade over time, and monitoring strategies to detect and correct drift early.** **Model drift** is the degradation of AI model performance over time as the statistical properties of production data diverge from training data. AI models learn patterns from historical data. When reality shifts—new products, changed policies, evolving customer language, emerging issues—the model's learned patterns become less accurate. Drift happens gradually and often invisibly until performance degrades noticeably. Two types matter for customer service. Data drift: the inputs change (customers ask about new things, phrase requests differently). Concept drift: the right answers change (policies update, processes evolve). Both cause previously-accurate models to produce incorrect results. A model trained on pre-pandemic support patterns struggled with "can I cancel my event" queries when that intent spiked dramatically. Detecting drift requires ongoing monitoring: tracking confidence distributions, comparing current inputs to training data distributions, measuring performance metrics over time. Addressing drift typically means retraining with updated data—but for grounded AI systems, it might mean updating knowledge sources rather than the model itself. The key is having observability infrastructure that surfaces drift before customers complain. **Related terms:** [AI observability](https://www.lorikeetcx.ai/glossary/ai-observability), [Hallucination detection](https://www.lorikeetcx.ai/glossary/hallucination-detection), [Knowledge base](https://www.lorikeetcx.ai/glossary/knowledge-base) ## Multi-brand customer support URL: https://www.lorikeetcx.ai/glossary/multi-brand-customer-support **Learn what multi-brand customer support is, challenges of supporting multiple brands, and how AI enables consistent cross-brand service.** Multi-brand customer support is the capability to serve customers across multiple brands, sub-brands, or product lines from a single platform — with each brand maintaining its own identity, policies, tone of voice, and knowledge base. This is a common requirement for holding companies, multi-product businesses, and companies operating across regions with different brand identities. The complexity of multi-brand support is often underestimated. Each brand may have: - Different products, pricing, and policies - Different tone of voice and communication style guidelines - Different regulatory requirements (especially across regions) - Different knowledge bases and escalation paths - Different SLAs and operating hours - Shared backend systems but separate customer-facing identities For traditional contact centers, multi-brand support typically means separate agent teams (or at minimum, separate training) for each brand. This creates operational overhead: duplicate staffing, duplicate QA processes, and difficulty sharing best practices across brands. AI changes this equation. A well-architected AI agent platform can serve multiple brands from a single deployment — switching between brand personalities, policy sets, and knowledge bases based on which brand the customer is interacting with. This delivers the operational efficiency of a single platform with the customer experience of dedicated brand support. The key architectural requirement is true separation of brand contexts within the AI system, not just cosmetic customization. The AI needs distinct knowledge bases, distinct guardrails, and distinct personality configurations per brand — not a single model with brand-name tokens swapped in. This capability is particularly valuable for insurance companies (multiple product brands), financial services groups (multiple financial products), and consumer companies operating across regions. **Related terms:** [AI concierge](https://www.lorikeetcx.ai/glossary/ai-concierge), [omnichannel customer support](https://www.lorikeetcx.ai/glossary/omnichannel-customer-support), [AI guardrails](https://www.lorikeetcx.ai/glossary/ai-guardrails) ## Multi-Turn Conversation URL: https://www.lorikeetcx.ai/glossary/multi-turn-conversation **Discover what a multi-turn conversation is in AI, how context is maintained across exchanges, and why multi-turn capability defines advanced AI.** **Multi-turn conversation** is a dialogue that extends across multiple exchanges between user and AI system, requiring the system to maintain context and coherence throughout the interaction. Single-turn interactions (question in, answer out) are easy. Multi-turn conversations are hard. The customer says "I want to return my order." The AI asks which order. The customer says "the one from last week." The AI needs to track that they're discussing a return, identify the correct order, and continue the conversation with that context intact. Multi-turn capabilities require: context management (tracking what's been said), reference resolution (understanding "it," "that one," "the second thing"), state tracking (where are we in the conversation flow), and coherent continuation (responses that acknowledge conversation history). Quality drops as conversations lengthen—context gets lost, references become ambiguous, the AI repeats itself or contradicts earlier statements. For support operations, multi-turn performance determines whether AI handles real customer conversations or just simple queries. Most real support interactions span multiple exchanges. Test your AI with realistic multi-turn scenarios: topic switches, clarifications, corrections, compound requests. If it falls apart after three turns, it's not production-ready for complex interactions. **Related terms:** [Context window](https://www.lorikeetcx.ai/glossary/context-window), [AI agent memory](https://www.lorikeetcx.ai/glossary/ai-agent-memory), [Conversational AI](https://www.lorikeetcx.ai/glossary/conversational-ai) ## Natural language processing (NLP) URL: https://www.lorikeetcx.ai/glossary/natural-language-processing-nlp **Learn what natural language processing (NLP) is, how NLP enables AI to understand human language, and its role in modern customer service.** Natural language processing (NLP) is the field of AI focused on enabling machines to understand, interpret, and generate human language. It encompasses the technologies that allow AI systems to read text, understand speech, extract meaning, and produce natural-sounding responses. NLP has evolved through several generations: - **Rule-based (1960s-2000s):** Manually coded grammar rules and pattern matching. Brittle and expensive to maintain. - **Statistical (2000s-2010s):** Machine learning models trained on labeled data. Better at handling variation but limited by training data. - **Neural/transformer-based (2017-present):** Deep learning models (BERT, GPT, etc.) that learn language patterns from massive datasets. Dramatic improvement in understanding context, nuance, and ambiguity. In customer service, NLP is the foundation for: - **Understanding customer messages:** Parsing intent, extracting entities (dates, account numbers, product names), and interpreting sentiment - **Generating responses:** Producing natural, contextually appropriate replies - **Processing unstructured data:** Analyzing free-text feedback, survey responses, and social media mentions - **Multilingual support:** Handling customer interactions across languages For CX teams, the practical distinction that matters is between NLP as a component and NLP as a solution. Having strong NLP capabilities doesn't automatically translate to effective customer service AI — the NLP needs to be combined with system integrations, business logic, guardrails, and operational workflows to deliver value. The best NLP model in the world is useless if it can't access the customer's account or execute a refund. **Related terms:** [intent detection](https://www.lorikeetcx.ai/glossary/intent-detection), [sentiment analysis](https://www.lorikeetcx.ai/glossary/sentiment-analysis), [large language model](https://www.lorikeetcx.ai/glossary/large-language-model), [conversational AI](https://www.lorikeetcx.ai/glossary/conversational-ai) ## Natural language workflows URL: https://www.lorikeetcx.ai/glossary/natural-language-workflows **Understand what natural language workflows are, how plain-language instructions configure AI agents, and why they simplify automation setup.** Natural language workflows are automated business processes defined and configured using plain language descriptions rather than traditional programming, flowcharts, or visual rule builders. A team member describes the workflow in natural language — "When a customer requests a refund for an order placed more than 60 days ago, check if they have a loyalty status, and if so, offer a partial credit" — and the system translates this into executable logic. In customer service, natural language workflows bridge the gap between the CX team's domain expertise and the technical implementation of automation. The people who understand customer needs and business policies best (support managers, CX leaders) can directly create and modify the logic that governs AI behavior, without depending on engineering resources. Key characteristics of natural language workflows: - **Human-readable:** Anyone can understand what the workflow does by reading its description - **Auditable:** The logic is expressed in business terms, making it reviewable by compliance, legal, and management teams - **Iterative:** Workflows can be refined through conversation — "also check if the product is in the excluded category" — rather than rebuilt from scratch - **Deterministic where needed:** Despite being expressed in natural language, the resulting logic can enforce strict rules for compliance-sensitive operations Natural language workflows are particularly valuable in regulated industries where the people who understand compliance requirements (legal, compliance officers) need to verify that the AI's behavior matches policy. When the workflow logic is expressed in plain English rather than code, verification becomes a conversation rather than a code review. The key architectural challenge is ensuring that natural language descriptions translate into reliable, consistent execution. Ambiguity in language must be resolved before the workflow processes real customer interactions. **Related terms:** [conversational configuration](https://www.lorikeetcx.ai/glossary/conversational-configuration), [AI guardrails](https://www.lorikeetcx.ai/glossary/ai-guardrails), [AI compliance](https://www.lorikeetcx.ai/glossary/ai-compliance) ## Natural Language Workflows (NLW) URL: https://www.lorikeetcx.ai/glossary/natural-language-workflows-nlw **Learn what Natural Language Workflows (NLW) are, how they let teams build AI automations using everyday language without code.** Natural Language Workflows (NLW) are AI automation rules written in plain English rather than code, decision trees, or flowcharts. They describe what the AI should do in conversational language that non-technical teams can read and modify. Traditional support automation requires building rigid decision trees: if the customer says X, then do Y. This approach breaks on the infinite variability of human language. NLWs flip the model—you describe the intent, conditions, and desired outcome in prose, and the AI interprets customer messages against that description. An NLW might read: "When a customer asks about their order status, look up their most recent order. If it shipped, share the tracking number and expected delivery date. If it hasn't shipped, check if it's delayed and explain why." This replaces dozens of branching rules with a single readable policy. The power of NLWs is maintainability. When your return policy changes, you update one document rather than tracing through a flowchart. When edge cases emerge, you add clarifying sentences rather than new branches. Product teams can write and modify workflows without engineering support. NLWs work because modern language models can parse intent from prose—but they require clear, complete descriptions. Ambiguous NLWs produce ambiguous AI behavior. **Related terms:** [Match Rate](https://www.lorikeetcx.ai/glossary/match-rate), [Effective Automation Rate](https://www.lorikeetcx.ai/glossary/effective-automation-rate), [Resolution Rate](https://www.lorikeetcx.ai/glossary/resolution-rate) ## Net Collection Rate URL: https://www.lorikeetcx.ai/glossary/net-collection-rate **Discover what net collection rate measures in healthcare revenue cycle, how to calculate it, and strategies to maximize reimbursement collection.** **Net collection rate** is the percentage of expected collectible revenue actually collected, calculated as payments received divided by charges minus contractual adjustments. This is the metric that tells you whether your revenue cycle is working. A 95% net collection rate means you're leaving 5% of money you've already earned on the table through write-offs, untimely filing, and collection failures. Top performers achieve 98%+ net collection rates. For telehealth, patient responsibility portions often represent a larger share of revenue due to higher deductible plans and out-of-network scenarios. This means your patient collections process—payment at time of service, payment plan options, follow-up on balances—has outsized impact on net collection rate compared to traditional practices heavily weighted toward payer revenue. Calculate this monthly and track the trend. Declining net collection rate is an early warning sign of revenue cycle problems before they hit your cash position. **Related terms:** [Patient payment collection rate](https://www.lorikeetcx.ai/glossary/patient-payment-collection-rate), [Days in Accounts Receivable](https://www.lorikeetcx.ai/glossary/days-in-accounts-receivable), [Cost to collect](https://www.lorikeetcx.ai/glossary/cost-to-collect) ## Net Promoter Score (NPS) URL: https://www.lorikeetcx.ai/glossary/net-promoter-score-nps **Learn what Net Promoter Score (NPS) is, how to calculate it, and why NPS is the most widely used metric for measuring customer loyalty.** Net Promoter Score (NPS) measures customer loyalty by asking a single question: "How likely are you to recommend [company] to a friend or colleague?" on a 0-10 scale. Respondents are categorized as Promoters (9-10), Passives (7-8), or Detractors (0-6). **NPS = % Promoters - % Detractors** NPS scores range from -100 to +100. A positive score means more promoters than detractors; scores above 50 are considered excellent. NPS differs from CSAT in an important way: it measures the overall relationship, not a single interaction. A customer might rate an individual support interaction as satisfactory (high CSAT) while still being unlikely to recommend the company (low NPS) due to cumulative frustrations. NPS captures the bigger picture. The connection between customer service and NPS is well-documented but often indirect. Customer service interactions rarely create Promoters on their own — but they frequently create Detractors. A single terrible support experience can turn a satisfied customer into someone who actively warns others away from the company. This asymmetry means the primary goal of customer service in NPS terms is preventing Detractors, not creating Promoters. For CX teams, NPS is a lagging indicator — by the time NPS moves, the underlying experience changes happened weeks or months ago. More actionable approaches combine NPS tracking with operational metrics (FCR, CES, CSAT) that can be influenced in real-time. AI customer service impacts NPS through consistency and availability. When every customer interaction meets a baseline quality standard (no dropped conversations, no incorrect information, no multi-day response times), the Detractor creation rate drops, and NPS gradually improves. **Related terms:** [customer satisfaction score](https://www.lorikeetcx.ai/glossary/customer-satisfaction-score), [customer effort score](https://www.lorikeetcx.ai/glossary/customer-effort-score), [customer churn rate](https://www.lorikeetcx.ai/glossary/customer-churn-rate) ## Next-Best Action URL: https://www.lorikeetcx.ai/glossary/next-best-action **Understand what next-best action is, how AI recommends optimal actions in real time, and why proactive suggestions improve customer outcomes.** **Next-best action** is an AI-driven recommendation system that determines the optimal action to take with a customer at any given moment, based on context, history, and predicted outcomes. Next-best action synthesizes available information to recommend what to do next. For an incoming support request, this might mean: route to specialist, offer proactive discount, suggest relevant article, escalate to retention, or auto-resolve. The recommendation considers customer value, issue type, emotional state, historical patterns, and business objectives. In practice, next-best action powers both automation decisions (what should the AI do) and agent assistance (what should the human do). The AI might surface "this customer has churned previously—consider retention offer" or "this account has an open billing dispute—prioritize resolution." Recommendations adapt in real-time as conversations unfold. Effective next-best action requires integration: connecting customer data (who is this), interaction history (what's happened before), current context (what's happening now), and outcome data (what worked in similar situations). Without these inputs, recommendations are generic. With them, next-best action enables personalized, optimized handling at scale—the right action for the right customer at the right moment. **Related terms:** [AI personalization](https://www.lorikeetcx.ai/glossary/ai-personalization), [AI agent orchestration](https://www.lorikeetcx.ai/glossary/ai-agent-orchestration), [AI workflow automation](https://www.lorikeetcx.ai/glossary/ai-workflow-automation) ## NLU (Natural Language Understanding) URL: https://www.lorikeetcx.ai/glossary/nlu-natural-language-understanding **Learn what NLU is, how natural language understanding extracts meaning from text, and why NLU accuracy is critical for AI customer service.** **NLU (natural language understanding)** is the AI capability to extract meaning, intent, and relevant information from human language input, going beyond surface text to comprehend what users actually mean. NLU is the comprehension layer of conversational AI. It transforms raw user input ("yeah I need to change that address from the order I placed yesterday") into structured understanding: intent=address_change, entity=order_date:yesterday, sentiment=neutral. This structured interpretation enables downstream processing—routing, response generation, action execution. Core NLU tasks include: intent classification (what does the user want), entity extraction (what specific information did they provide), sentiment analysis (how do they feel), coreference resolution (what does "it" refer to), and semantic parsing (understanding relationships in complex sentences). Quality NLU handles the mess of human communication: typos, slang, run-on sentences, implicit meaning. NLU capabilities have improved dramatically with large language models. Earlier NLU required extensive training data for each intent and entity type. Modern approaches leverage pre-trained models that understand language generally, requiring less domain-specific training. This shifts effort from building NLU to designing what to do with the understanding—workflow execution, response generation, escalation logic. **Related terms:** [Natural language processing](https://www.lorikeetcx.ai/glossary/natural-language-processing), [Intent detection](https://www.lorikeetcx.ai/glossary/intent-detection), [Intent recognition](https://www.lorikeetcx.ai/glossary/intent-recognition) ## Omnichannel customer support URL: https://www.lorikeetcx.ai/glossary/omnichannel-customer-support **Discover what omnichannel customer support is, how it unifies conversations across channels, and why seamless experiences drive customer loyalty.** Omnichannel customer support is an approach where customers can interact with a business across multiple channels (chat, email, phone, social media, messaging apps) with a seamless, consistent experience — meaning the customer's context and history carry over regardless of which channel they use or switch to. This is distinct from multichannel support, where a company offers multiple channels but each operates independently. In a multichannel setup, a customer who starts on chat and follows up via email has to re-explain their issue. In an omnichannel setup, the email agent sees the full chat history and picks up where the conversation left off. True omnichannel is hard to implement because it requires: - **Unified customer identity:** Recognizing the same customer across channels (email address, phone number, account ID) - **Shared conversation history:** A single view of all interactions regardless of channel - **Consistent policies and capabilities:** The same actions available whether the customer is on chat, email, or phone - **Intelligent routing:** Directing customers to the right channel and agent based on their issue and preferences AI agents simplify certain aspects of omnichannel while making others more complex. An AI agent can inherently maintain context across a conversation — it doesn't forget the first half when the customer switches channels. However, ensuring consistent AI behavior across channels (chat is instant and interactive; email is asynchronous and longer-form) requires thoughtful design. The practical question for most CX teams is not "should we be omnichannel?" (yes) but "which channels matter for our customers?" and "where should we invest in AI-first experiences vs. human-first experiences?" **Related terms:** [customer journey mapping](https://www.lorikeetcx.ai/glossary/customer-journey-mapping), [AI concierge](https://www.lorikeetcx.ai/glossary/ai-concierge), [ticketing system](https://www.lorikeetcx.ai/glossary/ticketing-system) ## Open Banking URL: https://www.lorikeetcx.ai/glossary/open-banking **Learn what open banking is, how API-driven financial data sharing works, and the impact of open banking on customer service in fintech.** **Open banking** is a regulatory and technical framework that requires banks to share customer financial data with authorized third parties through secure APIs, enabling new financial services built on direct account access. Enabled by regulations like PSD2 in Europe and emerging frameworks in Australia, the US, and elsewhere, open banking changes what's possible in fintech CX. Instead of asking customers to manually enter account details or upload statements, you can pull verified data directly. Instead of card payments with their 2-3% fees and chargeback risk, you can initiate account-to-account transfers. For support teams, open banking creates new workflows. Customers may need help connecting accounts, understanding permissions, or troubleshooting failed connections. The consent model—where customers explicitly authorize data sharing—means support agents need to explain what data you access and why. Transparency builds trust. The CX opportunity is significant. Open banking payments typically have higher approval rates than cards (no CVV errors, no expired cards) and near-instant settlement. Account aggregation lets you personalize experiences based on a customer's complete financial picture. The teams that build support processes around open banking now will have an advantage as adoption accelerates. **Related terms:** [Real-time payments](https://www.lorikeetcx.ai/glossary/real-time-payments), [Transaction approval rate](https://www.lorikeetcx.ai/glossary/transaction-approval-rate), [Payment failure rate](https://www.lorikeetcx.ai/glossary/payment-failure-rate) ## Patient Acquisition Cost URL: https://www.lorikeetcx.ai/glossary/patient-acquisition-cost **Understand what patient acquisition cost measures in healthcare, how to calculate it, and strategies to reduce cost while growing patient volume.** **Patient acquisition cost** is the total marketing and sales expense required to acquire one new patient, calculated as total acquisition spending divided by new patients acquired in a period. Telehealth economics often feature higher patient acquisition costs than traditional practices because digital marketing is expensive and patients have more options. Offset this with higher lifetime value through geographic reach (patients who couldn't access your specialty locally) and retention strategies that keep patients in your ecosystem. Calculate this by channel to understand which acquisition sources deliver efficiently. Paid search might cost $150 per patient while referral programs cost $30. This intelligence should drive budget allocation, but also consider patient quality—a $150 patient who completes a care plan is worth more than a $30 patient who churns after one visit. For telehealth startups and growth-stage companies, patient acquisition cost relative to lifetime value determines unit economics viability. If you're spending more to acquire patients than they'll ever generate in margin, you don't have a business, you have a burning pile of cash. **Related terms:** [Virtual visit no-show rate](https://www.lorikeetcx.ai/glossary/virtual-visit-no-show-rate), [Care plan adherence rate](https://www.lorikeetcx.ai/glossary/care-plan-adherence-rate), [Patient satisfaction score](https://www.lorikeetcx.ai/glossary/patient-satisfaction-score) ## Patient Payment Collection Rate URL: https://www.lorikeetcx.ai/glossary/patient-payment-collection-rate **Learn what patient payment collection rate measures, how to improve point-of-service collections, and reduce outstanding patient balances.** **Patient payment collection rate** is the percentage of patient-responsibility amounts (copays, deductibles, coinsurance, self-pay) that are actually collected. As patient financial responsibility has grown—high deductible plans now cover roughly half of commercially insured Americans—this metric has become increasingly material to telehealth revenue. Organizations that can't collect from patients are effectively subsidizing patient care. Time-of-service collection dramatically outperforms post-visit billing. For telehealth, this means payment capture during scheduling or immediately after visit completion, while patients are still engaged. Payment plans and financing options improve collection rates for larger balances. Track collection rate by payment timing (pre-visit, time of service, post-visit) and payment method. If credit card on file at scheduling collects at 95% and post-visit billing collects at 40%, you know where to focus. **Related terms:** [Net collection rate](https://www.lorikeetcx.ai/glossary/net-collection-rate), [Cost to collect](https://www.lorikeetcx.ai/glossary/cost-to-collect), [Days in Accounts Receivable](https://www.lorikeetcx.ai/glossary/days-in-accounts-receivable) ## Patient Satisfaction Score URL: https://www.lorikeetcx.ai/glossary/patient-satisfaction-score **Discover what patient satisfaction score measures, how it relates to HCAHPS, and strategies to improve the patient experience.** **Patient satisfaction score** is a standardized metric measuring patient-reported experience with telehealth services, typically captured through post-visit surveys using scales like NPS, CSAT, or CG-CAHPS. In telehealth, satisfaction scores correlate strongly with technical experience. A brilliant diagnosis delivered over a choppy video connection with three disconnections will tank your scores regardless of clinical quality. This means CX and IT are inextricably linked in ways traditional practices never faced. Survey timing matters enormously. Surveys sent immediately after visits capture technical frustrations and wait time complaints. Surveys sent 24-48 hours later better reflect clinical outcomes but suffer from lower response rates. Most telehealth operations use immediate post-visit surveys for operational metrics and periodic relationship surveys for strategic planning. Benchmark targets vary by specialty, but top telehealth performers achieve NPS scores of 60+ and CSAT above 90%. If you're below those thresholds, segment your data by visit type, provider, and time of day—the variance often reveals specific fixable problems rather than systemic issues. **Related terms:** [Average virtual wait time](https://www.lorikeetcx.ai/glossary/average-virtual-wait-time), [Technical connection success rate](https://www.lorikeetcx.ai/glossary/technical-connection-success-rate), [Virtual visit completion rate](https://www.lorikeetcx.ai/glossary/virtual-visit-completion-rate) ## Payment Failure Rate URL: https://www.lorikeetcx.ai/glossary/payment-failure-rate **Learn what payment failure rate measures in financial services, common causes of failed payments, and how to reduce transaction failures.** **Payment failure rate** is the percentage of initiated payments that fail to complete successfully, encompassing declines, timeouts, processing errors, and insufficient funds scenarios. Unlike transaction approval rate (which focuses on the payment processor's decision), payment failure rate captures the full customer experience of trying to move money. A payment can fail for reasons that never reach the processor: network timeouts, validation errors, 3D Secure abandonment, or your own system issues. For CX teams, payment failures are high-emotion moments. The customer tried to do something important—pay a bill, fund an account, complete a purchase—and couldn't. How you handle these moments defines your brand. Proactive notification beats silence. Clear error messages beat generic ones. Offering alternative payment methods beats asking customers to "try again later." Segment your failure rate by cause. Processor declines, fraud blocks, and technical errors each require different interventions. If your fraud rules are declining 5% of transactions but only 0.1% of those would have been actual fraud, you're creating more customer pain than you're preventing. Review your decline codes monthly and tune accordingly. **Related terms:** [Transaction approval rate](https://www.lorikeetcx.ai/glossary/transaction-approval-rate), [Chargeback rate](https://www.lorikeetcx.ai/glossary/chargeback-rate), [Customer effort score](https://www.lorikeetcx.ai/glossary/customer-effort-score) ## PCI DSS Compliance URL: https://www.lorikeetcx.ai/glossary/pci-dss-compliance **Understand what PCI DSS compliance is, key requirements for handling payment data, and how to maintain compliance in customer service operations.** **PCI DSS compliance** is adherence to the Payment Card Industry Data Security Standard, a set of security requirements that any organization handling cardholder data must meet to process, store, or transmit credit card information. This isn't optional. If you touch card data, you comply with PCI DSS or you don't process cards. The standard has four levels based on transaction volume, with Level 1 merchants (over 6 million transactions annually) requiring the most rigorous external audits. For CX teams, PCI DSS shapes what you can and can't do. Agents shouldn't take full card numbers over the phone without proper call recording controls. Screen sharing sessions shouldn't capture payment pages. Customer data exports need encryption and access controls. The security requirements that feel like friction often exist because someone, somewhere, failed an audit or got breached. The smart approach is building PCI-compliant processes that don't punish the customer experience. Use tokenization so customers can reference cards without agents seeing full numbers. Implement secure payment links for phone transactions. Design support tools that mask sensitive fields by default. Compliance and good CX aren't mutually exclusive—they just require intentional design. **Related terms:** [Account takeover rate](https://www.lorikeetcx.ai/glossary/account-takeover-rate), [Data security](https://www.lorikeetcx.ai/glossary/data-security), [Compliance training](https://www.lorikeetcx.ai/glossary/compliance-training) ## Per-resolution pricing URL: https://www.lorikeetcx.ai/glossary/per-resolution-pricing **Learn what per-resolution pricing is, how outcome-based AI pricing works, and why paying per resolution aligns vendor incentives with results.** Per-resolution pricing is a commercial model for AI customer service where the business pays only when the AI successfully resolves a customer issue — rather than paying per seat, per interaction, or a flat platform fee. It aligns the vendor's incentives with the customer's outcomes: the vendor only earns when they deliver value. The model works by defining what constitutes a "resolution" — typically a customer issue that is fully addressed without requiring follow-up, escalation to a human agent, or a repeat contact within a defined window. Interactions that are escalated, abandoned, or unresolved are not billed. Per-resolution pricing changes the economics of AI customer service in several ways: - **De-risked adoption:** Companies can deploy AI without committing to a large upfront investment. If the AI doesn't resolve issues, they don't pay. - **Aligned incentives:** The vendor is motivated to improve resolution quality, not just deflection volume. A "deflected" interaction that doesn't actually help the customer generates no revenue. - **Transparent ROI:** The cost per AI-resolved issue is explicit and directly comparable to the cost of human resolution, BPO rates, or other alternatives. - **Scalable costs:** As volume increases, costs scale linearly with value delivered rather than requiring additional seat licenses or infrastructure investment. The counterpoint to per-resolution pricing is that not all resolutions are equal — a password reset and a complex billing dispute represent very different levels of effort. Some vendors address this through tiered pricing based on interaction complexity. For CX leaders evaluating pricing models, per-resolution pricing is worth examining because it forces the vendor to stand behind their AI's performance. If the AI can't resolve, the vendor doesn't get paid. **Related terms:** [cost per resolution](https://www.lorikeetcx.ai/glossary/cost-per-resolution), [resolution rate](https://www.lorikeetcx.ai/glossary/resolution-rate), [deflection rate](https://www.lorikeetcx.ai/glossary/deflection-rate) ## Policy Renewal Rate URL: https://www.lorikeetcx.ai/glossary/policy-renewal-rate **Discover what policy renewal rate measures in insurance, how to calculate it, and strategies to improve policyholder retention and renewals.** **Policy renewal rate** is the percentage of policies that renew at the end of their term rather than lapsing or being replaced by a competitor. Renewal rate is the compound interest of insurance economics. A 90% renewal rate versus an 85% renewal rate doesn't sound like much, but over five years it's the difference between retaining 59% of your book versus 44%. Acquisition costs mean the lifetime value of a renewed policy far exceeds a new policy of equivalent premium. The claims experience is the single largest driver of renewal rate. Policyholders who never file claims renew based on price and inertia. Policyholders who file claims renew—or don't—based on how that experience felt. A smooth, fast, empathetic claims process creates fierce loyalty; a frustrating one sends customers to competitors who haven't yet failed them. Measuring renewal rate alone isn't enough. Segment by profitability: are you retaining your best risks or your worst? A high renewal rate on unprofitable segments accelerates losses. Pair renewal rate with retention of premium weighted by loss ratio to understand whether your retention is actually valuable. **Related terms:** [Policyholder satisfaction](https://www.lorikeetcx.ai/glossary/policyholder-satisfaction), [Quote-to-bind ratio](https://www.lorikeetcx.ai/glossary/quote-to-bind-ratio), [Claims settlement cycle time](https://www.lorikeetcx.ai/glossary/claims-settlement-cycle-time) ## Policyholder Satisfaction URL: https://www.lorikeetcx.ai/glossary/policyholder-satisfaction **Learn what policyholder satisfaction measures, key drivers of satisfaction in insurance, and how AI improves the policyholder experience.** **Policyholder satisfaction** is a measure of how well an insurer meets or exceeds customer expectations across the policy lifecycle—typically captured through surveys like NPS, CSAT, or transactional feedback following key interactions. Satisfaction matters because insurance is a grudge purchase that becomes emotionally significant only at moments of need. Policyholders don't think about their insurer until they file a claim or receive a renewal notice. These moments of truth disproportionately shape perception and loyalty. The claims experience dominates satisfaction scores. An NPS survey after premium payment tells you little; one after claims settlement tells you everything. Insurers with high satisfaction scores typically share common traits: fast acknowledgment, clear communication throughout, fair outcomes, and empathetic human contact when it matters. Measuring satisfaction is necessary but not sufficient. The value comes from linking satisfaction scores to operational metrics and outcomes. Which touchpoints drive satisfaction? How does satisfaction correlate with renewal rate? Which segments are underserved? Satisfaction data without action is vanity measurement. **Related terms:** [Policy renewal rate](https://www.lorikeetcx.ai/glossary/policy-renewal-rate), [Claims settlement cycle time](https://www.lorikeetcx.ai/glossary/claims-settlement-cycle-time), [FNOL response time](https://www.lorikeetcx.ai/glossary/fnol-response-time) --- # Telehealth-Specific Terms ## Prescription Fulfillment Rate URL: https://www.lorikeetcx.ai/glossary/prescription-fulfillment-rate **Understand what prescription fulfillment rate measures, how to track medication dispensing accuracy, and strategies to improve fulfillment.** **Prescription fulfillment rate** is the percentage of prescriptions written during telehealth visits that are successfully filled by patients at a pharmacy. This metric sits at the intersection of clinical outcomes and patient experience. A prescription that isn't filled represents a clinical intervention that didn't happen—the visit succeeded technically but failed therapeutically. For chronic condition management and medication-dependent specialties, fulfillment rates directly predict health outcomes. Telehealth creates unique fulfillment barriers: e-prescribing to unfamiliar pharmacies, insurance coverage verification challenges, and the lack of a physical handoff moment where patients ask clarifying questions. Some telehealth organizations have responded by integrating pharmacy services directly, while others focus on follow-up workflows to catch unfilled prescriptions. Target fulfillment rates of 80%+ for acute prescriptions and 70%+ for chronic medications. Track unfilled prescriptions by medication type and follow up proactively—both because it improves outcomes and because it generates billable clinical touchpoints. **Related terms:** [Care plan adherence rate](https://www.lorikeetcx.ai/glossary/care-plan-adherence-rate), [Virtual visit completion rate](https://www.lorikeetcx.ai/glossary/virtual-visit-completion-rate), [Patient satisfaction score](https://www.lorikeetcx.ai/glossary/patient-satisfaction-score) ## Prior Authorization Success Rate URL: https://www.lorikeetcx.ai/glossary/prior-authorization-success-rate **Learn what prior authorization success rate measures in healthcare, how to reduce denials, and strategies to streamline the auth process.** **Prior authorization success rate** is the percentage of prior authorization requests approved by payers on initial submission. Prior auth is the revenue cycle bottleneck that most directly impacts patient experience. A denied prior auth doesn't just delay revenue—it delays treatment, frustrates patients, and creates rework for clinical and administrative staff. For telehealth services with payer-specific coverage rules, prior auth success rates can vary dramatically across payers. Track success rates by payer and service type. Build institutional knowledge about which services require prior auth for which payers, and front-load that work before scheduling. Nothing destroys patient trust faster than a scheduled appointment canceled due to prior auth denial. Some telehealth organizations have achieved significant improvements by using automated prior auth tools that integrate with payer systems. If you're processing high volumes of prior auths manually, this technology investment often pays back quickly in both approval rates and staff time. **Related terms:** [Claims denial rate](https://www.lorikeetcx.ai/glossary/claims-denial-rate), [Clean claim rate](https://www.lorikeetcx.ai/glossary/clean-claim-rate), [Days in Accounts Receivable](https://www.lorikeetcx.ai/glossary/days-in-accounts-receivable) ## Proactive customer service URL: https://www.lorikeetcx.ai/glossary/proactive-customer-service **Discover what proactive customer service is, how anticipating issues before they arise improves CX, and strategies for proactive support.** Proactive customer service is the practice of reaching out to customers before they contact you — anticipating needs, preventing problems, and creating positive touchpoints rather than waiting for issues to arise. It's the opposite of reactive support, where the customer must initiate every interaction. Examples of proactive customer service include: - Notifying a customer about a shipping delay before they check the tracking page - Alerting a customer that their credit card is about to expire before a payment fails - Following up after onboarding to check if the customer needs help with features they haven't used - Reaching out when usage patterns suggest the customer may be considering churning - Proactively communicating about known issues or outages before customers experience them The business impact of proactive service is significant. It reduces inbound ticket volume (fewer customers need to contact you when you've already addressed their concern), improves CSAT (customers appreciate being looked after), and strengthens retention (proactive outreach signals that the company values the relationship). Historically, proactive service was difficult to scale because it required human agents to identify opportunities and initiate outreach. AI changes this constraint. AI systems can monitor customer data in real-time, identify situations that warrant proactive outreach, and execute that outreach at scale — sending personalized messages to thousands of customers simultaneously. The shift from reactive to proactive service represents a broader transformation in CX: from "support as a cost center that responds to problems" to "customer experience as a value driver that builds relationships." AI is the enabling technology that makes this shift economically viable. **Related terms:** [AI concierge](https://www.lorikeetcx.ai/glossary/ai-concierge), [customer retention rate](https://www.lorikeetcx.ai/glossary/customer-retention-rate), [customer journey mapping](https://www.lorikeetcx.ai/glossary/customer-journey-mapping) ## Prompt Engineering URL: https://www.lorikeetcx.ai/glossary/prompt-engineering **Learn what prompt engineering is, how crafting effective prompts improves AI output, and best practices for prompt design in customer service.** **Prompt engineering** is the practice of designing and refining the instructions, context, and examples provided to large language models to achieve desired outputs. With generative AI, how you ask matters as much as what you ask. Prompt engineering is the discipline of crafting inputs that reliably produce accurate, appropriate, and useful outputs. For customer service AI, this means designing prompts that guide the model to respond helpfully, stay within policy bounds, use appropriate tone, and acknowledge limitations. Key prompt engineering techniques: system prompts (defining role, constraints, and persona), few-shot examples (showing desired input-output patterns), chain-of-thought (encouraging step-by-step reasoning), and retrieval augmentation (providing relevant context). Prompts also define guardrails: what topics to avoid, when to escalate, how to handle edge cases. Prompt engineering is iterative. Initial prompts produce unexpected results, require refinement, and need ongoing adjustment as edge cases emerge. It's also fragile—small prompt changes can produce large output differences. For production systems, prompt engineering requires version control, testing infrastructure, and monitoring. The skill is becoming as important as traditional software engineering for AI-powered customer service. **Related terms:** [Generative AI for customer service](https://www.lorikeetcx.ai/glossary/generative-ai-for-customer-service), [AI grounding](https://www.lorikeetcx.ai/glossary/ai-grounding), [Retrieval augmented generation](https://www.lorikeetcx.ai/glossary/retrieval-augmented-generation) ## Provider Utilization Rate URL: https://www.lorikeetcx.ai/glossary/provider-utilization-rate **Understand what provider utilization rate measures in healthcare, how to calculate it, and strategies to optimize clinician scheduling.** **Provider utilization rate** is the percentage of available provider appointment slots that are filled with patient visits, measuring how efficiently your clinical capacity is being used. High utilization rates indicate strong patient demand and effective scheduling. Low rates mean you're paying for clinical capacity that isn't generating revenue. The target range is 80-90%—high enough to maximize revenue, low enough to absorb same-day scheduling and avoid excessive patient wait times. Telehealth enables utilization optimization strategies impossible in physical practices: providers can fill gaps with patients from different time zones, appointment durations can flex more easily, and scheduling can be more dynamic. Organizations that exploit these advantages can achieve utilization rates 10-15 points higher than their in-person benchmarks. Track utilization by provider, time of day, and day of week. Patterns reveal optimization opportunities—if Tuesday mornings run at 95% while Thursday afternoons run at 60%, you have a demand shaping problem, not a capacity problem. **Related terms:** [Appointment completion rate](https://www.lorikeetcx.ai/glossary/appointment-completion-rate), [Virtual visit no-show rate](https://www.lorikeetcx.ai/glossary/virtual-visit-no-show-rate), [Average virtual wait time](https://www.lorikeetcx.ai/glossary/average-virtual-wait-time) ## Quality assurance (QA) in customer service URL: https://www.lorikeetcx.ai/glossary/quality-assurance-qa-in-customer-service **Learn what QA means in customer service, how quality assurance programs work, and why consistent QA drives better support outcomes.** Quality assurance (QA) in customer service is the systematic process of evaluating customer interactions to ensure they meet defined standards for accuracy, professionalism, policy adherence, and customer experience. QA identifies where agents (human or AI) are performing well and where they need improvement. Traditional QA in customer service involves: - **Sampling:** Selecting a percentage of interactions (typically 2-5%) for review - **Scorecard evaluation:** Rating each interaction against predefined criteria (greeting, accuracy, empathy, resolution, compliance) - **Calibration:** Ensuring QA evaluators apply standards consistently - **Coaching:** Using QA findings to coach agents on improvement areas - **Trending:** Tracking quality metrics over time to identify patterns The fundamental limitation of traditional QA is the sample size. Reviewing 2-5% of interactions means 95-98% go unexamined. Quality issues that affect a small percentage of interactions — but a large absolute number of customers — go undetected. A systematic policy violation occurring in 3% of interactions won't show up reliably in a 2% sample. This is where automated QA transforms the function. By using AI to evaluate 100% of interactions, QA shifts from a sampling-based estimation to a comprehensive assessment. Every interaction is scored, every policy violation is flagged, and every quality trend is visible. For organizations deploying AI agents, QA takes on an additional dimension: it's not just evaluating human performance but monitoring AI behavior in production. Auto QA becomes the safety net that catches issues guardrails missed, and the feedback loop that drives continuous AI improvement. **Related terms:** [automated quality assurance](https://www.lorikeetcx.ai/glossary/automated-quality-assurance), [customer satisfaction score](https://www.lorikeetcx.ai/glossary/customer-satisfaction-score), [AI observability](https://www.lorikeetcx.ai/glossary/ai-observability) ## Quality Assurance (QA) Score URL: https://www.lorikeetcx.ai/glossary/quality-assurance-qa-score **Discover what a QA score measures, how to calculate quality assurance scores for agents, and how to use scores to improve performance.** **Quality Assurance (QA) score** is a composite rating of agent performance based on evaluated interactions, typically assessing accuracy, compliance, communication quality, and adherence to process. QA programs exist because customer satisfaction scores alone don't tell you why interactions succeed or fail. A CSAT of 4/5 doesn't reveal that the agent gave incorrect policy information that will cause problems later. QA catches process failures before they compound into churn, complaints, or regulatory issues—critical for fintech and insurance where errors have legal consequences. The challenge is calibration. Inconsistent scoring across QA reviewers makes the data useless. Establish clear rubrics, run regular calibration sessions, and track inter-rater reliability. Modern QA increasingly uses AI to score 100% of interactions rather than sampling 2-5%—this catches outliers that random sampling misses and removes the "performance on reviewed calls" vs. "performance on all calls" gap. Just ensure AI scoring is validated against human judgment before trusting it for coaching decisions. **Related terms:** [Agent quality score](https://www.lorikeetcx.ai/glossary/agent-quality-score), [Escalation rate](https://www.lorikeetcx.ai/glossary/escalation-rate), [First reply time](https://www.lorikeetcx.ai/glossary/first-reply-time) ## Quote-to-Bind Ratio URL: https://www.lorikeetcx.ai/glossary/quote-to-bind-ratio **Learn what quote-to-bind ratio measures in insurance, how to calculate conversion rates, and strategies to improve quote-to-policy conversion.** **Quote-to-bind ratio** is the percentage of insurance quotes that convert into bound policies, measuring the effectiveness of the sales process from pricing through purchase completion. Quote-to-bind ratio reflects both pricing competitiveness and customer experience friction. A low ratio might mean your rates are uncompetitive—or it might mean your application process is too cumbersome, your website abandons users at key moments, or your follow-up is inadequate. Diagnosing the root cause requires segmenting: where in the process are prospects dropping off? The metric varies dramatically by channel and line of business. Direct-to-consumer auto insurance might see 10-15% quote-to-bind; commercial lines sold through brokers might exceed 50%. Compare against your own historical performance and direct competitors, not across product lines. Quote-to-bind is increasingly a technology metric. Insurers with real-time quoting, pre-filled applications, and embedded purchase flows convert at higher rates than those requiring callbacks or manual processes. For high-consideration purchases like life insurance, quote-to-bind also depends on follow-up cadence and agent responsiveness. **Related terms:** [Policy renewal rate](https://www.lorikeetcx.ai/glossary/policy-renewal-rate), [Underwriting automation](https://www.lorikeetcx.ai/glossary/underwriting-automation), [Policyholder satisfaction](https://www.lorikeetcx.ai/glossary/policyholder-satisfaction) ## Real-Time Payments URL: https://www.lorikeetcx.ai/glossary/real-time-payments **Understand what real-time payments are, how instant payment rails work, and the impact on customer service in banking and financial services.** **Real-time payments** (RTP) are payment transactions that clear and settle within seconds, 24/7/365, providing immediate confirmation to both sender and receiver. Unlike traditional payment rails that batch process overnight or take days to settle, RTP networks—FedNow in the US, NPP in Australia, Faster Payments in the UK—move money instantly. For customers, this means no more "2-3 business days" waiting periods. For support teams, it means different problems to solve. The good news: real-time payments eliminate timing-related support tickets. "When will my payment arrive?" becomes irrelevant when arrival is immediate. The challenging news: there's no recall window. Once an RTP transaction settles, it's final. Misdirected payments, fat-finger errors, and authorized push payment fraud all become harder to remediate. Build your support processes for this finality. Confirmation screens should be crystal clear about where money is going. Error messaging needs to catch mistakes before submission, not after. Your fraud detection must work in real-time—you can't review transactions post-settlement. And your agents need clear escalation paths for the inevitable "I sent money to the wrong person" calls, even when the answer is uncomfortable. **Related terms:** [Open banking](https://www.lorikeetcx.ai/glossary/open-banking), [Transaction approval rate](https://www.lorikeetcx.ai/glossary/transaction-approval-rate), [Payment failure rate](https://www.lorikeetcx.ai/glossary/payment-failure-rate) --- # Insurtech-Specific Terms ## Reimbursement Cycle Time URL: https://www.lorikeetcx.ai/glossary/reimbursement-cycle-time **Learn what reimbursement cycle time measures in healthcare, how to reduce time from claim submission to payment, and optimize revenue cycle.** **Reimbursement cycle time** is the elapsed time between claim submission and payment receipt, measuring how quickly payers process and pay your claims. This differs from Days in AR by starting the clock at submission rather than date of service. The gap between the two metrics reveals internal billing delays—if your Days in AR is 45 and your reimbursement cycle time is 25, you're taking 20 days to submit claims after visits. Track reimbursement cycle time by payer. Variations reveal which payers process quickly versus slowly, informing both cash flow forecasting and payer contracting strategy. A payer that pays 15% more but takes 30 days longer might be a worse deal depending on your cash position. Telehealth-specific factors that extend reimbursement cycle time include payer requests for additional telehealth documentation, place-of-service clarifications, and review of out-of-state provider credentials. Address these proactively by including required documentation in initial submissions. **Related terms:** [Days in Accounts Receivable](https://www.lorikeetcx.ai/glossary/days-in-accounts-receivable), [Clean claim rate](https://www.lorikeetcx.ai/glossary/clean-claim-rate), [First-pass resolution rate](https://www.lorikeetcx.ai/glossary/first-pass-resolution-rate) ## Resolution rate URL: https://www.lorikeetcx.ai/glossary/resolution-rate **Discover what resolution rate measures, how to calculate it for AI and human support, and why resolution rate is the definitive CX metric.** Resolution rate measures the percentage of customer issues that are fully resolved — the customer's problem is solved, their question is answered, or their request is completed. It is the outcome-focused alternative to deflection rate and one of the most important metrics for evaluating AI customer service performance. **Resolution rate = (Issues fully resolved / Total issues handled) x 100** The critical word is "fully." A ticket marked as closed is not necessarily resolved. A customer who stops responding after a confusing chatbot interaction is not resolved. A customer who is given a partial answer and doesn't follow up is not resolved. Genuine resolution means the customer's need was met. Measuring resolution accurately requires defining clear criteria: - **No recontact:** The customer doesn't raise the same issue again within a defined window (7-30 days) - **Action completion:** If the issue required an action (refund, account change, appointment), the action was successfully executed - **Customer confirmation:** The customer explicitly confirmed the issue was resolved (via survey, follow-up message, or in-conversation confirmation) Resolution rate is the metric that distinguishes AI systems that genuinely help customers from those that merely process interactions. Two AI platforms can both report handling 70% of incoming volume — but if one resolves 90% of those and the other resolves 50%, the customer experience and business impact are radically different. For CX leaders evaluating AI vendors, asking "what is your resolution rate?" and "how do you define and measure resolution?" is more revealing than asking about automation rate or deflection rate. The definition is where vendors differentiate. **Related terms:** [deflection rate](https://www.lorikeetcx.ai/glossary/deflection-rate), [first contact resolution](https://www.lorikeetcx.ai/glossary/first-contact-resolution), [cost per resolution](https://www.lorikeetcx.ai/glossary/cost-per-resolution) ## Retrieval-augmented generation (RAG) URL: https://www.lorikeetcx.ai/glossary/retrieval-augmented-generation-rag **Learn what RAG is, how retrieval-augmented generation combines search with AI to produce accurate answers grounded in your knowledge base.** Retrieval-augmented generation (RAG) is an AI architecture that improves the accuracy of language model responses by retrieving relevant information from external data sources before generating a response. Instead of relying solely on what the model learned during training, RAG pulls current, verified information from a knowledge base or document store and uses it to ground the response. In customer service, RAG addresses one of the biggest risks of using large language models: hallucination. Without RAG, an LLM might generate plausible-sounding but incorrect information about a company's policies, pricing, or procedures. With RAG, the LLM references the company's actual documentation when formulating responses. The RAG process works in three steps: - **Retrieval:** When a customer asks a question, the system searches the knowledge base for relevant documents, articles, or data - **Augmentation:** The retrieved information is included in the LLM's context alongside the customer's message - **Generation:** The LLM generates a response grounded in the retrieved information rather than its parametric knowledge RAG quality depends heavily on: - **Knowledge base quality:** RAG can only retrieve what exists. Gaps, outdated content, or conflicting articles in the knowledge base directly degrade response quality. - **Retrieval accuracy:** The system must find the right documents for the given question. Poor retrieval means the LLM generates responses grounded in irrelevant information. - **Chunking strategy:** How documents are split and indexed affects whether the retrieval system can find the specific paragraph that answers the customer's question. For CX teams, RAG is one of those architectural decisions that's invisible when it works and obvious when it doesn't. The practical question is not whether the AI uses RAG (most do), but how well the RAG pipeline is implemented and maintained. **Related terms:** [knowledge base](https://www.lorikeetcx.ai/glossary/knowledge-base), [large language model](https://www.lorikeetcx.ai/glossary/large-language-model), [AI hallucinations](https://www.lorikeetcx.ai/glossary/ai-hallucinations) ## Self-service rate URL: https://www.lorikeetcx.ai/glossary/self-service-rate **Understand what self-service rate measures, how to calculate it, and strategies to increase self-service adoption to reduce support costs.** Self-service rate measures the percentage of customer issues resolved through self-service channels — help centers, FAQ pages, community forums, or automated tools — without any interaction with a human agent or AI agent. **Self-service rate = (Issues resolved via self-service / Total issues) x 100** Self-service is the lowest-cost resolution channel by a significant margin. When a customer finds the answer in a help article, the marginal cost is essentially zero. This makes self-service rate an important efficiency metric, particularly for high-volume operations. However, self-service has inherent limitations: - **Complexity ceiling:** Self-service works for straightforward, common questions but breaks down for complex, multi-step, or account-specific issues - **Measurement difficulty:** It's hard to know if a customer who visited a help article actually found their answer. Page views don't equal resolutions. - **Customer preference:** Many customers prefer to ask their question directly rather than search through documentation, especially for urgent issues - **Maintenance burden:** Self-service content requires continuous updates as products, policies, and procedures change AI is blurring the line between self-service and assisted service. An AI agent that resolves a customer's issue through a conversational interface shares characteristics with both: it's automated like self-service but interactive like agent assistance. This is why many organizations are moving from "self-service rate" to "automation rate" or "AI resolution rate" as their primary metric for non-human-handled interactions. The most effective self-service strategies are not standalone — they're integrated with AI and human channels so customers can seamlessly escalate from a help article to an AI conversation to a human agent without losing context. **Related terms:** [knowledge base](https://www.lorikeetcx.ai/glossary/knowledge-base), [deflection rate](https://www.lorikeetcx.ai/glossary/deflection-rate), [resolution rate](https://www.lorikeetcx.ai/glossary/resolution-rate) ## Self-Service Ratio URL: https://www.lorikeetcx.ai/glossary/self-service-ratio **Learn what self-service ratio is, how it compares automated resolutions to total volume, and benchmarks for AI-driven customer self-service.** **Self-service ratio** is the percentage of customer support needs resolved through automated or self-directed channels—knowledge bases, FAQs, chatbots, community forums—without agent involvement. This metric measures how well you've productized your support. High self-service ratios (60-80% for mature operations) mean you've anticipated common questions and built answers into the experience. Low ratios mean agents are manually handling issues that could be deflected. For fintech and insurance, where regulatory requirements often mandate certain disclosures anyway, self-service is a double win: better compliance documentation and lower support costs. The catch: don't celebrate self-service ratio in isolation. If customers attempt self-service, fail, and then contact support anyway, you've added friction without reducing load. Track self-service success rate (resolved without escalation to agent) alongside the ratio. Also watch for "dark" self-service failure—customers who hit your FAQ, don't find answers, and churn silently without ever contacting you. Survey data and session analytics catch what ticket metrics miss. **Related terms:** [Contact rate](https://www.lorikeetcx.ai/glossary/contact-rate), [Ticket volume](https://www.lorikeetcx.ai/glossary/ticket-volume), [Escalation rate](https://www.lorikeetcx.ai/glossary/escalation-rate) ## Sentiment analysis URL: https://www.lorikeetcx.ai/glossary/sentiment-analysis **Discover what sentiment analysis is, how AI detects customer emotion from text and voice, and why real-time sentiment drives better support.** Sentiment analysis is the use of AI to detect the emotional tone of customer communications — classifying messages as positive, negative, neutral, or more granularly (frustrated, confused, satisfied, angry, anxious). In customer service, sentiment analysis enables real-time understanding of how customers feel during interactions. Applications in customer service include: - **Real-time escalation:** Detecting rising customer frustration and triggering escalation to a human agent before the situation deteriorates - **Routing:** Prioritizing interactions from distressed or angry customers for faster handling - **Quality assurance:** Scoring agent (human or AI) empathy and tone across interactions - **Trend analysis:** Tracking aggregate customer sentiment over time, by product, by issue type, or following product launches and outages - **Voice of customer:** Aggregating sentiment across all channels to understand overall customer health Modern sentiment analysis goes beyond simple positive/negative classification. Nuanced systems can detect: - Sarcasm ("Oh great, another update that breaks everything") - Mixed sentiment ("I love the product but your support is terrible") - Urgency without negative sentiment ("I need this resolved before my trip tomorrow") - Escalating frustration across a conversation (customer started calm but is getting increasingly terse) For CX teams, sentiment analysis is most valuable as an input to other systems rather than a standalone metric. Sentiment-informed routing, sentiment-triggered escalation, and sentiment-based QA scoring are all more actionable than a sentiment dashboard alone. The goal is to act on emotional signals in real-time, not just report on them after the fact. **Related terms:** [auto-tagging](https://www.lorikeetcx.ai/glossary/auto-tagging), [intent detection](https://www.lorikeetcx.ai/glossary/intent-detection), [customer satisfaction score](https://www.lorikeetcx.ai/glossary/customer-satisfaction-score) ## Simulation testing URL: https://www.lorikeetcx.ai/glossary/simulation-testing **Learn what simulation testing is for AI agents, how synthetic conversations validate AI before deployment, and why testing reduces production errors.** Simulation testing in customer service AI is the practice of testing AI agent behavior against realistic customer scenarios before deploying changes to production. Instead of testing with real customers and real consequences, teams run simulated conversations to verify that the AI handles various scenarios correctly. Simulation testing addresses a fundamental risk of AI customer service: unlike traditional software where behavior is deterministic, AI responses can vary based on conversation context, phrasing, and model updates. A change that improves handling of one scenario might degrade performance on another. Without comprehensive testing, these regressions go undetected until customers experience them. Key aspects of simulation testing include: - **Scenario coverage:** Testing across the full range of interaction types — common cases, edge cases, compliance-sensitive scenarios, hostile inputs, multi-turn conversations - **Regression detection:** Comparing AI performance before and after changes (model updates, knowledge base edits, workflow modifications) to catch unintended degradation - **Guardrail validation:** Verifying that safety constraints hold under adversarial inputs — can the AI be tricked into taking unauthorized actions or providing prohibited information? - **Scale testing:** Running hundreds or thousands of simulated conversations to build statistical confidence in AI performance For regulated industries, simulation testing is particularly important. Deploying AI changes that haven't been tested against compliance scenarios — and discovering the issue through a customer interaction — creates regulatory risk. Simulation testing provides an audit-ready record that the AI was validated before deployment. The most effective simulation testing combines automated scenario generation (testing at scale) with human-designed edge cases (testing the scenarios that matter most). Neither alone is sufficient. **Related terms:** [AI guardrails](https://www.lorikeetcx.ai/glossary/ai-guardrails), [automated quality assurance](https://www.lorikeetcx.ai/glossary/automated-quality-assurance), [AI compliance](https://www.lorikeetcx.ai/glossary/ai-compliance) ## Straight-Through Processing Rate URL: https://www.lorikeetcx.ai/glossary/straight-through-processing-rate **Understand what straight-through processing rate measures in insurance, how automation enables touchless claims, and industry benchmarks.** **Straight-through processing rate (STP rate)** is the percentage of claims that are received, processed, and paid without any human intervention. STP rate is the clearest measure of claims automation maturity. Leading insurers now achieve 40-60% STP on simple, high-volume claim types like auto glass, travel delay, and pet wellness. The theoretical ceiling is higher, but real-world factors—fraud detection triggers, documentation gaps, coverage ambiguity—mean some human review will always be necessary. Calculating STP requires precision about what counts as "no human intervention." Does a supervisor approval count as human touch? What about a customer service inquiry that doesn't change the claim outcome? Most insurers define STP as claims that flow from FNOL to payment without entering an adjuster's queue. Be consistent in your definition over time. The business case for STP is compelling: claims processed straight-through cost 50-90% less than manually adjusted claims and resolve in hours instead of days. But STP done poorly—paying claims without adequate fraud detection or coverage verification—can be catastrophic. The goal is confident automation, not reckless automation. **Related terms:** [FNOL response time](https://www.lorikeetcx.ai/glossary/fnol-response-time), [Claims triage](https://www.lorikeetcx.ai/glossary/claims-triage), [Underwriting automation](https://www.lorikeetcx.ai/glossary/underwriting-automation) ## Subrogation Recovery Rate URL: https://www.lorikeetcx.ai/glossary/subrogation-recovery-rate **Learn what subrogation recovery rate measures, how insurers recover costs from third parties, and strategies to improve recovery outcomes.** **Subrogation recovery rate** is the percentage of claims costs recovered from third parties who were legally responsible for the loss, measured against the total recoverable amount identified. Subrogation is found money—claims already paid that you can partially or fully recover from at-fault parties, their insurers, or product manufacturers. Most insurers leave significant recovery on the table because subrogation requires proactive identification, documentation, and pursuit. The challenge is identifying recovery opportunities at the right time. Subrogation potential is often clearest immediately after a claim is reported—when evidence is fresh and third-party liability is obvious—but most claims teams are focused on the policyholder, not the recovery. By the time claims close, subrogation opportunities have often degraded or expired. Top performers build subrogation identification into claims triage rather than treating it as an afterthought. They flag potential recovery during FNOL, preserve evidence proactively, and measure adjusters on recovery identification alongside resolution speed. Recovery rates above 70% of identified potential are achievable; many insurers operate below 50%. **Related terms:** [Claims triage](https://www.lorikeetcx.ai/glossary/claims-triage), [Loss Adjustment Expense ratio](https://www.lorikeetcx.ai/glossary/loss-adjustment-expense-ratio), [Claims leakage](https://www.lorikeetcx.ai/glossary/claims-leakage) ## Success Rate URL: https://www.lorikeetcx.ai/glossary/success-rate **Discover what success rate measures in AI customer service, how to define and calculate it, and why it is the ultimate measure of AI effectiveness.** Success Rate measures the percentage of matched tickets where the AI workflow executed correctly and achieved its intended outcome. It answers: "When AI attempts a ticket, how often does it succeed?" Success rate is calculated as successful executions divided by matched tickets. A ticket is "successful" when the workflow ran without errors, the AI took the correct actions, and the outcome matched the workflow's intent. This is narrower than resolution rate—a ticket might reach a successful workflow outcome but the customer might still be unsatisfied. Low success rate despite high match rate indicates workflow problems: edge cases the workflow doesn't handle, integration failures, unclear workflow logic, or AI misinterpreting customer inputs. Diagnosing success rate drops means reviewing failed executions to identify patterns. Success rate multiplied by match rate gives you your theoretical automation ceiling—the maximum percentage of tickets that could be fully automated if everything worked. If match rate is 60% and success rate is 80%, your ceiling is 48%. Improving either metric raises the ceiling. **Related terms:** [Match Rate](https://www.lorikeetcx.ai/glossary/match-rate), [Effective Automation Rate](https://www.lorikeetcx.ai/glossary/effective-automation-rate), [Resolution Rate](https://www.lorikeetcx.ai/glossary/resolution-rate), [Total Quality Score (TQS)](https://www.lorikeetcx.ai/glossary/total-quality-score-tqs) ## Technical Connection Success Rate URL: https://www.lorikeetcx.ai/glossary/technical-connection-success-rate **Learn what technical connection success rate measures in telehealth, how to reduce failed connections, and improve virtual visit reliability.** **Technical connection success rate** is the percentage of attempted telehealth visits where both patient and provider successfully establish a stable audio/video connection sufficient to complete the clinical encounter. This is the foundational metric for telehealth operations. Everything else—clinical quality, satisfaction, revenue—depends on patients and providers actually being able to communicate. Technical failures create the worst patient experience: time invested, problem unresolved, frustration high, trust destroyed. Track both connection establishment (did the call connect?) and connection stability (did it stay connected?). A visit with three disconnections and choppy audio technically "succeeded" but delivered a terrible experience. Define quality thresholds—minimum audio clarity, video resolution, latency—and measure against them. High-performing telehealth platforms achieve 98%+ connection success rates. Below 95% indicates infrastructure problems requiring immediate investment. Segment failures by device type, browser, network type, and geography to identify whether problems are systemic or concentrated in specific technical scenarios. **Related terms:** [Virtual visit completion rate](https://www.lorikeetcx.ai/glossary/virtual-visit-completion-rate), [Patient satisfaction score](https://www.lorikeetcx.ai/glossary/patient-satisfaction-score), [Appointment completion rate](https://www.lorikeetcx.ai/glossary/appointment-completion-rate) ## Ticket Backlog URL: https://www.lorikeetcx.ai/glossary/ticket-backlog **Understand what ticket backlog is, how backlog accumulates in support operations, and proven strategies to clear and prevent ticket backlogs.** **Ticket backlog** is the cumulative count of unresolved customer support requests awaiting agent action at any point in time. Backlog is a lagging indicator of capacity problems. When inbound volume exceeds resolution capacity over sustained periods, backlog grows. Small backlogs are normal—work in progress—but growing backlogs indicate structural understaffing, efficiency problems, or a spike you haven't recovered from. For SLA-driven teams, backlog age distribution matters more than total count: 50 tickets under 4 hours old is fine; 50 tickets over 24 hours is a crisis. Monitor backlog trends, not snapshots. A Monday morning backlog of 200 tickets that clears by Wednesday is healthy absorption of weekend volume. A backlog of 200 that's been 200 for three weeks means you're exactly at capacity with no buffer. Fintech and insurance companies with regulatory response time requirements should set backlog alerts tied to SLA aging—you need to know when tickets are approaching breach, not after they've breached. **Related terms:** [Ticket volume](https://www.lorikeetcx.ai/glossary/ticket-volume), [First reply time](https://www.lorikeetcx.ai/glossary/first-reply-time), [Agent occupancy](https://www.lorikeetcx.ai/glossary/agent-occupancy) --- # Fintech-Specific Terms ## Ticket complexity URL: https://www.lorikeetcx.ai/glossary/ticket-complexity **Learn what ticket complexity means in customer service, how to classify ticket difficulty, and why complexity scoring improves routing and staffing.** Ticket complexity refers to the difficulty level of a customer service interaction, determined by factors like the number of systems involved, the judgment required, the regulatory sensitivity, and the number of steps needed to resolve the issue. Understanding and categorizing ticket complexity is essential for effective AI deployment and workforce planning. Not all customer service tickets are equal. A spectrum exists: - **Low complexity:** Single-system, single-step, clear-cut resolution. Password resets, order status checks, address updates. These are highly automatable. - **Medium complexity:** Multi-step processes with some decision points. Processing a return with restocking conditions, updating a subscription with prorated billing, handling a straightforward insurance claim. - **High complexity:** Multi-system, judgment-intensive, often emotionally charged or compliance-sensitive. Billing disputes involving multiple transactions, medical inquiries requiring clinical context, insurance claims with coverage ambiguity, complaints involving regulatory implications. The distribution of ticket complexity in a given business determines the value of AI automation. If 80% of tickets are low-to-medium complexity, AI can handle the majority of volume. If 50% are high complexity (common in regulated industries), AI's role shifts from volume handling to triage and augmentation. The "hard 20%" — the most complex tickets that require deep product knowledge, regulatory awareness, and human judgment — is where AI vendor differentiation becomes most apparent. Many AI systems can handle the easy cases; the question is what happens when the interaction gets complex. Does the AI escalate cleanly? Can it handle multi-step resolution? Does it understand regulatory constraints? For CX leaders planning AI deployments, mapping ticket complexity before selecting a vendor prevents mismatched expectations. A vendor promising 90% automation on a ticket mix that's 40% high-complexity is either underestimating the challenge or redefining "automation" to include deflection. **Related terms:** [escalation rate](https://www.lorikeetcx.ai/glossary/escalation-rate), [resolution rate](https://www.lorikeetcx.ai/glossary/resolution-rate), [AI agent](https://www.lorikeetcx.ai/glossary/ai-agent) ## Ticket Volume URL: https://www.lorikeetcx.ai/glossary/ticket-volume **Discover what ticket volume measures, how to track and forecast support demand, and strategies to manage volume spikes without adding headcount.** **Ticket volume** is the total count of customer support requests received across all channels within a defined time period. Volume is the foundation of capacity planning. Without accurate volume data—and forecasts—you can't staff appropriately, budget correctly, or measure the impact of deflection initiatives. Track volume by channel, time of day, day of week, and seasonality patterns. Fintech companies see spikes around market volatility; insurance companies around renewal periods and catastrophic events. Build these into your models. Raw volume trends are less useful than normalized metrics like contact rate. Volume rising 20% while your customer base grew 30% is actually good news. Volume flat while customers declined 10% is a warning sign. Segment volume by issue type to identify product problems, knowledge gaps, or process failures driving contacts. A sudden spike in password reset tickets might indicate a UX issue worth fixing; a gradual rise in billing questions might mean your invoices need clearer formatting. **Related terms:** [Contact rate](https://www.lorikeetcx.ai/glossary/contact-rate), [Ticket backlog](https://www.lorikeetcx.ai/glossary/ticket-backlog), [Self-service ratio](https://www.lorikeetcx.ai/glossary/self-service-ratio) ## Ticketing system URL: https://www.lorikeetcx.ai/glossary/ticketing-system **Learn what a ticketing system is, how support ticketing platforms organize and track customer issues, and key features to look for.** A ticketing system (also called a helpdesk or service desk) is the software platform that manages customer service interactions by creating, tracking, and organizing support requests as individual "tickets." Each ticket represents a customer issue and progresses through a lifecycle: created, assigned, in-progress, pending, resolved, closed. Major ticketing systems include Zendesk, Intercom, Freshdesk, Salesforce Service Cloud, Help Scout, and Jira Service Management. These platforms provide the operational backbone for customer service teams: agent workspaces, routing rules, SLA tracking, reporting, and customer communication tools. Core ticketing system capabilities include: - **Ticket creation and tracking:** Capturing customer requests from multiple channels and assigning unique identifiers - **Routing and assignment:** Distributing tickets to agents or teams based on rules (round-robin, skill-based, load-balanced) - **SLA management:** Tracking response and resolution times against defined service level agreements - **Workflow automation:** Triggering actions based on ticket events (auto-assign, auto-close, escalation rules) - **Reporting:** Providing operational metrics (volume, backlog, resolution time, CSAT) The relationship between ticketing systems and AI is evolving rapidly. Initially, AI was embedded within ticketing platforms as an add-on (Zendesk AI, Intercom Fin). Increasingly, AI-native platforms operate alongside or in front of ticketing systems — handling conversations directly and creating tickets only for cases that require human attention. For CX leaders, the strategic question is whether the ticketing system remains the center of gravity for their operation or whether AI becomes the primary interaction layer, with the ticketing system serving as a backend for human-escalated cases. The answer depends on what percentage of volume AI can genuinely resolve. **Related terms:** [omnichannel customer support](https://www.lorikeetcx.ai/glossary/omnichannel-customer-support), [first response time](https://www.lorikeetcx.ai/glossary/first-response-time), [escalation rate](https://www.lorikeetcx.ai/glossary/escalation-rate) ## Total quality score (TQS) URL: https://www.lorikeetcx.ai/glossary/total-quality-score-tqs **Understand what total quality score measures, how TQS aggregates multiple QA dimensions, and why holistic quality scoring improves service.** Total quality score (TQS) is a composite metric that evaluates the overall quality of a customer service interaction by combining multiple quality dimensions into a single score. Rather than tracking accuracy, tone, policy adherence, and resolution quality as separate metrics, TQS rolls them into one number that represents the holistic quality of the interaction. A typical TQS framework might weight dimensions like: | Dimension | Weight | What it measures | |-----------|--------|------------------| | Accuracy | 30% | Was the information provided correct? | | Resolution completeness | 25% | Was the customer's issue fully addressed? | | Policy adherence | 20% | Were required procedures followed? | | Communication quality | 15% | Was the interaction clear, professional, empathetic? | | Efficiency | 10% | Was the interaction handled without unnecessary steps? | TQS is particularly useful when comparing quality across different channels, agents, or handling methods (human vs. AI). A single composite score makes it possible to answer: "Is our AI agent delivering the same quality as our best human agents?" — a question that's difficult to answer when tracking five separate metrics. For AI customer service, TQS serves as the unified quality gate. Rather than setting separate thresholds for accuracy, tone, and policy adherence, teams set a TQS threshold that the AI must maintain. If TQS drops below the threshold, the system can automatically adjust — escalating more interactions to humans, tightening guardrails, or flagging specific interaction types for review. The main risk with composite scores is that they can mask problems. An AI that scores perfectly on accuracy and efficiency but poorly on empathy might still achieve a passing TQS — even though the empathy gap creates real customer friction. Teams using TQS should track component dimensions alongside the composite score. **Related terms:** [quality assurance in customer service](https://www.lorikeetcx.ai/glossary/quality-assurance-in-customer-service), [automated quality assurance](https://www.lorikeetcx.ai/glossary/automated-quality-assurance), [customer satisfaction score](https://www.lorikeetcx.ai/glossary/customer-satisfaction-score) ## Transaction Approval Rate URL: https://www.lorikeetcx.ai/glossary/transaction-approval-rate **Learn what transaction approval rate measures in financial services, how to optimize approval rates, and reduce false declines that frustrate customers.** **Transaction approval rate** is the percentage of payment attempts that successfully complete, measuring how often customers can actually use your product to move money. This is arguably the most important operational metric in payments. A 95% approval rate sounds high until you realize that's 1 in 20 customers hitting a wall when they try to pay. Those failed transactions generate support tickets, erode trust, and often result in lost customers. Approval rates vary by payment method, geography, and transaction type. Card-not-present transactions typically see 85-95% approval. Bank transfers and open banking payments often hit 98%+. Cross-border transactions are lower due to additional fraud screening. Your job is to understand your baseline by segment and systematically improve the laggards. Common culprits for low approval rates include outdated card credentials (use account updater services), insufficient funds (let users retry with different payment methods), and overly aggressive fraud rules (tune your risk thresholds). Your support team should have visibility into decline reason codes and be empowered to help customers resolve issues on first contact. **Related terms:** [Payment failure rate](https://www.lorikeetcx.ai/glossary/payment-failure-rate), [Chargeback rate](https://www.lorikeetcx.ai/glossary/chargeback-rate), [First contact resolution](https://www.lorikeetcx.ai/glossary/first-contact-resolution) ## Underwriting Automation URL: https://www.lorikeetcx.ai/glossary/underwriting-automation **Discover what underwriting automation is in insurance, how AI accelerates risk assessment, and the benefits of automated underwriting decisions.** **Underwriting automation** is the use of rules engines, machine learning models, and data integrations to assess risk, price policies, and make accept/decline decisions without manual underwriter review. Underwriting automation exists on a spectrum. At one end: simple rules that auto-decline applications outside appetite (too young, too risky, wrong geography). At the other: sophisticated ML models that ingest third-party data, predict loss ratios, and price policies dynamically—with no human touch unless exceptions trigger. The business case is straightforward: automated underwriting decisions cost pennies; manual ones cost dollars to tens of dollars. Speed matters too. A quote delivered in seconds converts better than one delivered in days, especially in direct-to-consumer channels. Personal lines insurers now automate 80-90% of new business; commercial lines lag significantly. The risk is model drift and adverse selection. Automated underwriting codifies assumptions about risk that may become stale as markets change. Insurers with mature automation practices continuously monitor model performance, comparing predicted to actual loss ratios and retraining when gaps emerge. **Related terms:** [Quote-to-bind ratio](https://www.lorikeetcx.ai/glossary/quote-to-bind-ratio), [Straight-through processing rate](https://www.lorikeetcx.ai/glossary/straight-through-processing-rate), [Fraud detection](https://www.lorikeetcx.ai/glossary/fraud-detection) ## Virtual Visit Completion Rate URL: https://www.lorikeetcx.ai/glossary/virtual-visit-completion-rate **Learn what virtual visit completion rate measures in telehealth, how to improve completion, and reduce patient drop-off during virtual care.** **Virtual visit completion rate** is the percentage of initiated telehealth sessions that conclude with a completed clinical encounter, including documentation and any necessary prescriptions or referrals. This differs from appointment completion rate by focusing on visits that start but don't finish—technical failures mid-visit, patient disconnections, provider emergencies, or sessions ended due to scope issues requiring in-person care. Track the reasons for incomplete visits. Technical failures point to infrastructure investments. Scope issues might indicate better pre-visit screening or patient education needs. Patient disconnections mid-visit could signal UX problems with your platform. Target 95%+ completion rate for visits that successfully connect. Below 90% suggests problems worth investigating—you're successfully getting patients to show up and connect, then losing them before delivering care. **Related terms:** [Appointment completion rate](https://www.lorikeetcx.ai/glossary/appointment-completion-rate), [Technical connection success rate](https://www.lorikeetcx.ai/glossary/technical-connection-success-rate), [Patient satisfaction score](https://www.lorikeetcx.ai/glossary/patient-satisfaction-score) ## Virtual Visit No-Show Rate URL: https://www.lorikeetcx.ai/glossary/virtual-visit-no-show-rate **Understand what virtual visit no-show rate is, common causes of telehealth no-shows, and strategies to reduce missed appointments.** **Virtual visit no-show rate** is the percentage of scheduled telehealth appointments where the patient fails to appear or connect within the appointment window. No-shows are the silent killer of telehealth unit economics. Unlike in-person practices where a missed appointment might be partially recaptured through walk-ins, an empty telehealth slot is pure waste—provider time paid for, zero revenue collected. The typical telehealth no-show rate runs 20-30%, roughly double the 10-15% seen in traditional practices. Contributing factors include lower perceived commitment (no travel investment), competing distractions at home, and technical anxiety. The good news: these are solvable problems. Effective interventions include SMS reminders 24 hours and 1 hour before appointments, pre-visit technical checks, and overbooking models calibrated to historical no-show patterns. Some organizations charge no-show fees, though this can backfire with patient satisfaction and is harder to enforce for telehealth. **Related terms:** [Appointment completion rate](https://www.lorikeetcx.ai/glossary/appointment-completion-rate), [Patient acquisition cost](https://www.lorikeetcx.ai/glossary/patient-acquisition-cost), [Provider utilization rate](https://www.lorikeetcx.ai/glossary/provider-utilization-rate) ## Voice of the customer (VoC) URL: https://www.lorikeetcx.ai/glossary/voice-of-the-customer-voc **Learn what Voice of the Customer (VoC) is, how to build a VoC program, and why customer feedback drives continuous experience improvement.** Voice of the customer (VoC) is the practice of systematically collecting, analyzing, and acting on customer feedback to understand their needs, expectations, and experiences. It encompasses structured data (surveys, ratings, NPS) and unstructured data (support conversations, social media posts, reviews, community forums). VoC programs aim to close the loop between what customers experience and what the business does about it. The most mature VoC programs don't just collect feedback — they route insights to the teams that can act on them: product receives feature requests and pain points, operations receives process friction signals, and leadership receives strategic patterns. In customer service, every interaction is a VoC signal — whether or not a formal survey is sent afterward. The content of support tickets reveals: - **Product friction:** What features cause confusion, what processes break, what documentation is missing - **Competitive pressure:** What alternatives customers mention, what capabilities they expect from competitors - **Emerging issues:** New problems that don't fit existing categories, which may indicate product bugs, market shifts, or unmet needs - **Sentiment trends:** Whether customers are generally getting happier or more frustrated over time AI transforms VoC from a periodic reporting function to a real-time intelligence system. Instead of quarterly surveys analyzed by a research team, AI can analyze every customer interaction as it happens — tagging themes, detecting sentiment shifts, surfacing emerging issues, and routing actionable insights to the right teams automatically. The companies that get the most value from VoC treat customer service as a strategic listening post, not just an operational cost center. Every resolved ticket contains data that can improve the product, the process, or the experience. **Related terms:** [sentiment analysis](https://www.lorikeetcx.ai/glossary/sentiment-analysis), [auto-tagging](https://www.lorikeetcx.ai/glossary/auto-tagging), [customer satisfaction score](https://www.lorikeetcx.ai/glossary/customer-satisfaction-score), [Net Promoter Score](https://www.lorikeetcx.ai/glossary/net-promoter-score) ## Warm handoff URL: https://www.lorikeetcx.ai/glossary/warm-handoff **Discover what a warm handoff is in customer service, how it differs from cold transfers, and why warm handoffs preserve context and satisfaction.** A warm handoff (also called a warm transfer) is the process of transferring a customer from one agent (or from an AI agent to a human agent) while preserving the full context of the conversation — so the customer doesn't have to repeat their issue. The receiving agent sees the conversation history, understands what's already been attempted, and picks up where the previous interaction left off. This is in contrast to a cold handoff, where the customer is transferred without context and must re-explain their situation. Cold handoffs are consistently rated as one of the top customer frustrations in service interactions. In AI-to-human handoffs, warm handoff quality is critical because it occurs at the moment of highest customer sensitivity. The customer has already interacted with AI — possibly for several turns — and the escalation to a human typically happens because the issue is complex, the customer is frustrated, or the AI reached its limits. Starting over at this point compounds the frustration. An effective warm handoff includes: - **Full conversation transcript:** Everything the customer has said and the AI has responded - **Issue summary:** A concise description of the customer's problem and what's been attempted - **Customer context:** Account details, history, sentiment indicators - **Reason for escalation:** Why the AI couldn't handle the interaction, helping the human agent approach it appropriately - **Attempted actions:** What the AI already tried, so the human agent doesn't repeat steps For CX teams evaluating AI platforms, testing the handoff experience is essential. The smoothness of the handoff — from the customer's perspective — is often more important than the AI's resolution rate, because it determines the experience for every customer who needs human help. **Related terms:** [escalation rate](https://www.lorikeetcx.ai/glossary/escalation-rate), [human-in-the-loop](https://www.lorikeetcx.ai/glossary/human-in-the-loop), [customer effort score](https://www.lorikeetcx.ai/glossary/customer-effort-score) ## WISMO (Where is my order?) URL: https://www.lorikeetcx.ai/glossary/wismo-where-is-my-order **Learn what WISMO means, why 'where is my order' is the most common support query, and how AI automates order tracking to reduce ticket volume.** WISMO — "Where is my order?" — is one of the highest-volume inquiry types in e-commerce and retail customer service. WISMO queries typically account for 20-40% of inbound support volume for companies that ship physical products. Despite their volume, WISMO inquiries are relatively straightforward to resolve: check the order status, look up tracking information, and communicate the expected delivery date. This makes WISMO one of the first interaction types that companies automate with AI. However, WISMO queries are more nuanced than they appear: - **Simple WISMO:** "Where's my order?" → Check tracking, provide status. Fully automatable. - **Complex WISMO:** "My order says delivered but I never received it" → Requires investigation, potential re-shipment, possible fraud check. Needs more sophisticated handling. - **Emotional WISMO:** "I ordered this as a birthday gift for tomorrow and it's not here yet" → Requires empathy and potentially expedited resolution. The customer needs more than a tracking link. - **Multi-order WISMO:** "I have three orders outstanding and only got one" → Requires looking up multiple orders, comparing tracking, and potentially splitting the response. For CX leaders, WISMO is a useful benchmark for AI capability evaluation. Every AI customer service vendor can handle simple WISMO. The test is whether the system handles the complex variants — the lost package investigation, the time-sensitive gift, the multi-order inquiry — without breaking down or escalating unnecessarily. WISMO automation also has a direct financial impact: reducing each WISMO inquiry from a $5-10 human-handled cost to $0.50-1 AI-handled cost saves significant money at scale, while simultaneously providing faster responses (instant vs. minutes or hours in a queue). **Related terms:** [AI agent](https://www.lorikeetcx.ai/glossary/ai-agent), [ticket complexity](https://www.lorikeetcx.ai/glossary/ticket-complexity), [resolution rate](https://www.lorikeetcx.ai/glossary/resolution-rate), [cost per resolution](https://www.lorikeetcx.ai/glossary/cost-per-resolution) ## Workforce management (WFM) URL: https://www.lorikeetcx.ai/glossary/workforce-management-wfm **Understand what workforce management (WFM) is, how WFM tools forecast demand and schedule agents, and why WFM is critical for contact centers.** Workforce management (WFM) in customer service encompasses the processes and tools used to forecast interaction volume, schedule agents, manage staffing levels, and ensure service level targets are met. It's the operational discipline of having the right number of agents available at the right time. Core WFM functions include: - **Forecasting:** Predicting future interaction volume based on historical patterns, seasonality, marketing campaigns, product launches, and known events - **Scheduling:** Creating agent schedules that match forecasted demand while respecting labor laws, agent preferences, and skill requirements - **Real-time management:** Adjusting staffing in response to unexpected volume spikes or drops during the day - **Performance tracking:** Monitoring adherence (are agents working their scheduled hours?) and productivity (are they handling the expected volume?) AI fundamentally changes the WFM equation. When AI handles a significant portion of interaction volume, WFM shifts from "how many agents do we need?" to "how many agents do we need for the interactions AI can't handle?" This is a smaller, more specialized workforce handling more complex interactions — which changes scheduling patterns, skill requirements, and staffing ratios. The practical implications for CX leaders: - **Reduced baseline staffing:** AI handles predictable volume; humans handle exceptions and complexity - **Flattened peaks:** AI absorbs volume spikes that previously required overtime, temporary staff, or degraded service levels - **Skill-shifted workforce:** Remaining human agents need stronger problem-solving and judgment skills, not just speed - **Changed metrics:** Traditional WFM metrics like utilization and adherence matter less; quality metrics and escalation handling matter more WFM doesn't disappear with AI — it evolves. The discipline of matching capacity to demand remains critical; the inputs and outputs change. **Related terms:** [average handling time](https://www.lorikeetcx.ai/glossary/average-handling-time), [business process outsourcing](https://www.lorikeetcx.ai/glossary/business-process-outsourcing), [escalation rate](https://www.lorikeetcx.ai/glossary/escalation-rate) --- # KYC Automation in 2026: How Automated KYC Verification Works URL: https://www.lorikeetcx.ai/articles/kyc-automation-fintech **The average bank spends $1,500-3,000 to complete a single client's KYC review - while automated systems do it in under 30 seconds.** KYC automation is the use of artificial intelligence and digital verification tools to simplify Know Your Customer checks in financial services - replacing manual document review, identity verification, and compliance screening with automated workflows. The global KYC market is projected to grow from $6.73 billion in 2025 to $16.31 billion by 2031, according to Mordor Intelligence. - 54% of banks spend $1,500-3,000 per manual KYC review, with 21% spending over $3,000 per client, per Fenergo research - Automated KYC reduces processing costs by up to 70% and cuts verification time by 78%, according to AU10TIX - 70% of financial firms lost clients last year due to slow onboarding - up from 48% in 2023, per Fenergo - AI adoption in KYC operations jumped from 42% in 2024 to 82% in 2025, yet only 4% have fully automated workflows *Last updated: March 2026* Know Your Customer checks are the single most expensive compliance activity for financial institutions. According to [Fenergo's 2025 KYC Trends Report](https://resources.fenergo.com/newsroom/arduous-client-onboarding-and-kyc-costing-banks-millions), the average firm now spends $72.9 million annually on AML and KYC operations alone. That cost is not just financial - 70% of firms lost clients last year because their onboarding process was too slow. The gap between AI adoption and actual automation is striking. While 82% of firms now use AI somewhere in their KYC workflows, only 4% have automated the majority of checks. This article explains how KYC automation works, what fintech teams should expect from implementation, and where the biggest cost and time savings come from. ## What Is KYC Automation? KYC automation is the use of AI, machine learning, and digital verification technologies to replace manual Know Your Customer processes in financial services. Automated KYC systems handle identity verification, document validation, sanctions screening, and ongoing monitoring without requiring human analysts to manually review every case. Traditional KYC involves analysts manually checking identity documents, screening against sanctions lists, verifying addresses, and cross-referencing multiple databases. A single manual review takes 18+ minutes and requires specialist staff. According to [Veriff](https://www.veriff.com/onboarding/learn/kyc-automation-enhancing-customer-onboarding), a human KYC specialist performs roughly 3 checks per hour. Automated systems verify up to 50 users in the same timeframe. **Know Your Customer (KYC)**: The regulatory process of verifying a customer's identity and assessing their risk profile before and during a business relationship with a financial institution. Lorikeet is an AI customer support platform that resolves tickets end-to-end - processing refunds, updating accounts, and handling complex multi-step workflows across chat, email, and voice. For fintech teams running KYC operations, Lorikeet handles the customer-facing side of verification workflows, guiding applicants through document submission and status updates while maintaining full audit trails. ## How Does Automated KYC Verification Work? Automated KYC verification works by combining optical character recognition, biometric matching, and AI-driven risk scoring to validate customer identities in real time. The system extracts data from identity documents, cross-references it against sanctions and PEP databases, and assigns a risk score - all without manual analyst intervention. ### Document Extraction and Validation AI systems scan passports, driver's licenses, and utility bills using OCR technology. They extract name, date of birth, address, and document numbers, then validate authenticity by checking security features, detecting tampering, and comparing against known document templates. This step alone eliminates hours of manual data entry. ### Biometric and Liveness Checks Facial recognition compares the applicant's selfie against their identity document photo. Liveness detection confirms the person is physically present - not using a printed photo or deepfake. According to [AU10TIX](https://www.au10tix.com/blog/fintech-trends-and-best-practices/), digital verification enhances fraud detection accuracy by 61% compared to manual review. "Banks will continue to increase spending on agentic systems that can investigate, trace funds, flag risky behaviors, and take automated action. Fraud teams won't just get dashboards; they'll get autonomous agents that can handle the first layer of triage." - Subramanian Narayanaswamy, Executive Director, Wells Fargo ## What Are the Biggest Cost Savings from KYC Automation? The biggest cost savings from KYC automation come from eliminating manual analyst time, reducing rework from human error, cutting client abandonment during onboarding, and lowering ongoing monitoring costs. Firms implementing automation report up to 70% cost reduction in KYC operations. - **Analyst time reduction.** Manual KYC reviews cost $1,500-3,000 per client according to [Fenergo](https://resources.fenergo.com/blogs/kyc-compliance-for-banks-addressing-the-cost). Automated systems complete the same verification in under 30 seconds, freeing analysts to focus on genuinely complex risk cases. - **Client retention.** 70% of financial firms lost clients due to slow onboarding in 2025. Automated KYC eliminates the 18+ minute manual processing bottleneck that drives applicants to competitors. - **Error and rework reduction.** [Reducing manual touchpoints](/articles/how-to-reduce-customer-service-costs) cuts the data entry errors and missed screening flags that trigger costly compliance remediation. - **Ongoing monitoring efficiency.** Continuous KYC monitoring through AI replaces periodic manual reviews, catching changes in customer risk profiles in real time rather than during annual check cycles. ## What Results Are Fintech Teams Seeing with KYC Automation? Fintech teams implementing KYC automation report verification time dropping from 18+ minutes to under 30 seconds, cost reductions of 48-70% in onboarding operations, and measurably improved fraud detection accuracy. According to [AU10TIX's 2026 research](https://www.au10tix.com/blog/fintech-trends-and-best-practices/), digital verification reduces manual processing time by 78%, enhances fraud detection accuracy by 61%, and lowers onboarding costs by 48%. [McKinsey](https://www.mckinsey.com/capabilities/risk-and-resilience/our-insights/how-agentic-ai-can-change-the-way-banks-fight-financial-crime) reports that agentic AI applied to KYC and AML processes can deliver productivity uplifts of 200% to 2,000%, with one human able to supervise 20 or more AI agents simultaneously. The total market spend on KYC and KYB systems is projected to reach $35.5 billion in 2026, growing to $53 billion by 2030 according to [Juniper Research](https://www.juniperresearch.com/research/fintech-payments/identity/kyc-kyb-research-report/). That spending reflects the scale of the problem - and the ROI available to teams that automate effectively. **Firms automating KYC cut verification time by 78% and onboarding costs by up to 48%. **[**See how Lorikeet handles customer-facing verification workflows**](https://www.lorikeetcx.ai/get-a-demo)**.** ## What Should Fintech Teams Consider Before Automating KYC? Fintech teams should evaluate KYC automation platforms on regulatory compliance coverage, integration depth with existing systems, accuracy of identity verification, audit trail completeness, and the ability to handle edge cases that require human review. The adoption gap is real. According to [Fenergo's survey of 450+ C-suite executives](https://resources.fenergo.com/reports/kyc-trends-2024-banking), AI adoption in KYC surged from 42% to 82% in a single year - but only 4% of banks have automated the majority of their workflows. Most firms are stuck in partial automation, handling the easy cases digitally while analysts still manually process complex reviews. The risk of not automating is equally important to consider. [AI guardrails](/articles/what-are-ai-guardrails-for-customer-service) and proper governance frameworks matter because McKinsey warns that banks failing to adopt agentic AI in compliance could put $170 billion in global profits at risk. The competitive pressure is not just about cost savings - it is about institutional survival. ## Lorikeet's Take on KYC Automation At Lorikeet, we see KYC automation as two problems, not one. The backend verification is one challenge. The customer experience during verification is another entirely. Most teams automate the document checks but leave applicants waiting in a communication black hole - no status updates, no guidance on failed checks, no way to ask questions without calling a phone line. Lorikeet handles the customer-facing side of KYC workflows. When a document upload fails or additional verification is needed, Lorikeet's AI agents guide the applicant through next steps immediately across [voice](https://www.lorikeetcx.ai/product/voice), chat, and email. That combination of backend automation and customer communication is what drives the onboarding completion rates that matter. If KYC drop-off is costing you clients, [see how Lorikeet's Resolution Loop closes the gap](https://www.lorikeetcx.ai/product/resolution-loop). ## Key Takeaways - Manual KYC costs $1,500-3,000 per review while automated systems verify in under 30 seconds - a 78% reduction in processing time - 70% of financial firms lost clients due to slow onboarding in 2025, making KYC automation a retention imperative - AI adoption in KYC jumped from 42% to 82% in one year, but only 4% of banks have fully automated their workflows - Agentic AI can deliver 200-2,000% productivity uplifts in KYC operations according to McKinsey - The global KYC market is projected to reach $16.31 billion by 2031 at 15.88% CAGR ## Frequently Asked Questions ### How much does KYC automation cost to implement? KYC automation implementation costs vary by platform and scope, but most fintech firms see positive ROI within 6-12 months. With manual KYC costing $1,500-3,000 per client review and firms spending an average $72.9 million annually on compliance, even a 48% cost reduction through automation generates substantial savings in the first year. ### How long does automated KYC verification take? Automated KYC verification typically completes in under 30 seconds for straightforward cases, compared to 18+ minutes for manual review. Complex cases involving enhanced due diligence may take longer but still resolve significantly faster than manual processes because the AI pre-assembles all relevant data for analyst review. ### Can automated KYC meet regulatory compliance requirements? Yes - automated KYC systems are designed to meet regulatory requirements including AML directives, FATF guidelines, and local financial regulations. They maintain complete audit trails of every verification step, which regulators require. The key is choosing a platform with configurable rules that adapt to your specific regulatory jurisdiction. ### What happens when automated KYC cannot verify a customer? When automated KYC cannot verify a customer, the system flags the case for manual review and provides the analyst with all collected data, the specific verification failures, and a risk assessment. This hybrid approach ensures edge cases receive human judgment while routine verifications stay fully automated. ### Is KYC automation suitable for small fintech companies? KYC automation is particularly valuable for small fintech companies that cannot afford large compliance teams. Cloud-based KYC platforms offer pay-per-verification pricing models that scale with volume, eliminating the need for upfront infrastructure investment while maintaining the same compliance standards as large institutions. ### What is the difference between KYC automation and eKYC? eKYC refers specifically to electronic Know Your Customer processes using digital identity verification. KYC automation is broader - it encompasses eKYC plus automated sanctions screening, ongoing monitoring, risk scoring, and compliance reporting. eKYC is one component within a fully automated KYC workflow. ### How does KYC automation improve fraud detection? KYC automation improves fraud detection by applying consistent AI-driven analysis to every application. According to AU10TIX, digital verification enhances fraud detection accuracy by 61% compared to manual review. Automated systems catch document tampering, synthetic identities, and cross-application patterns that human analysts frequently miss. KYC automation is no longer a competitive advantage in fintech - it is becoming table stakes. The firms that automated early are already seeing 48-70% cost reductions and dramatically faster onboarding. Those still relying on manual processes are losing clients at an accelerating rate. The 4% full-automation figure reveals both the challenge and the opportunity. Most firms are stuck in partial automation. The teams that push through to end-to-end automated workflows will capture the full productivity and retention benefits that partial automation only hints at. **Ready to automate your KYC customer communications and cut onboarding drop-off? **[**Get started with Lorikeet**](https://www.lorikeetcx.ai/get-a-demo)** and see how AI handles the customer-facing side of verification.** ## FAQ ### What is automated KYC verification? Automated KYC (Know Your Customer) verification uses AI and document recognition to verify identity documents, perform liveness checks, and screen against sanctions and PEP lists in seconds. Where manual KYC review takes 2-5 days and costs $1,500-3,000 per file, automated KYC verification completes in under 30 seconds at $1-3 per check. Modern fintechs run automated KYC for 90%+ of applicants and reserve human review for edge cases. ### How does KYC automation work? KYC automation works in four steps: document capture (driver's licence, passport), AI document authentication and OCR, biometric matching against a live selfie, and parallel database screening (sanctions, PEP, adverse media). Decision logic combines all signals into a risk score; low-risk applicants approve instantly, high-risk applicants escalate to a compliance officer. The whole pipeline runs in under a minute end-to-end. ### What are the cost savings from automated KYB solutions? Automated KYB (Know Your Business) cuts cost per business verification from $1,500-3,000 manual to $5-50 automated - a 95%+ reduction. For a fintech onboarding 1,000 businesses per month, that is $1.5M+ in annual savings on review costs alone. Add the abandonment-rate gains from cutting onboarding time from 7 days to 1 day, and the compounding revenue impact often exceeds the cost saving. ### What KYC automation solutions are available for fintechs? KYC automation solutions for fintechs split into three categories: identity verification specialists (Onfido, Persona, Veriff) that handle document and biometric checks, full onboarding orchestrators (Alloy, Sumsub, Plaid) that combine KYC with KYB, fraud, and account opening, and AI customer support platforms like Lorikeet that handle the human-in-the-loop conversations when KYC reviews escalate. Most fintechs combine a verification specialist with an orchestration layer, then add an AI agent to handle applicant questions during the wait. ## AI Customer Support for Peptide Telehealth Companies (2026 Guide) URL: https://www.lorikeetcx.ai/articles/peptide-telehealth-ai-support Updated: 2026-09-02 Keywords: AI customer support for peptide telehealth companies, peptide clinic customer service, GLP-1 telehealth support, compounded peptide support automation Peptide telehealth companies run support queues that most helpdesks were never built for: a single FDA list change can invalidate every open peptide prescription overnight. AI customer support for peptide telehealth companies is the use of AI agents to resolve patient-facing administrative tickets, such as shipping, refills, subscriptions, pharmacy transfers and regulatory notices, while routing anything clinical to a licensed provider. In 2026, with 12 peptides removed from the FDA's Category 2 list in April and 7 more under review, ticket volume in this category is growing faster than headcount. - Fortune sized the addressable peptide market at roughly $30 billion in July 2026, and Google Keyword Planner shows about 1,900 US searches a month for "peptide clinic". - Healthcare support rarely costs less than $30 per ticket, per LiveChatAI's 2025 benchmarks, against $2.70 to $5.60 per ticket in e-commerce. - The FDA warned 30 telehealth companies in one wave over compounded GLP-1 marketing, so every AI-written message needs the same compliance review as a landing page. - Administrative tickets (delivery, refills, billing, cancellations) are automatable today; dosing and side-effect questions must reach a clinician, not a bot. Last updated: September 2026 Peptides went from a niche longevity topic to a mainstream telehealth category in under 18 months. Hims & Hers, Ro and dozens of smaller clinics now sell protocols priced from $150 to more than $1,500 a month, and Google Keyword Planner shows about 4,400 US searches a month for "peptide companies". Support teams inherited the mess: subscription billing disputes, warm vials on doorsteps, patients asking what to do about a missed dose, and a regulatory calendar that changes what can legally be shipped. Traditional chatbots make this worse because they answer everything. The teams doing well are drawing a hard line: AI resolves the administrative work end-to-end, and every clinical question reaches a human clinician with full context attached. ## What Is Customer Support for a Peptide Telehealth Company? Customer support for a peptide telehealth company covers every patient interaction that is not a clinical consult: order tracking, cold-chain delivery problems, refill timing, subscription changes, refunds, pharmacy transfers, payment failures and questions about regulatory changes. In 2026 most of these tickets arrive by chat, email and SMS, and a growing share arrive by phone. The mix skews heavily toward billing and access. Better Business Bureau complaint files for several peptide sellers show repeating patterns of unexpected recurring charges, subscriptions that were hard to cancel and refunds refused on unopened product, according to BBB records. Persistence data explains the volume: Blue Health Intelligence found that 30% of GLP-1 patients stop in the first month and 58% stop within 12 weeks, and every one of those exits generates a cancellation or refund ticket. 503A compounding pharmacy: a state-licensed pharmacy that prepares patient-specific medications from bulk substances, subject to FDA rules on which substances may be used. Category 2 bulk drug substance: an FDA designation for a substance with identified safety concerns, which blocks 503A pharmacies from compounding it for human use. Lorikeet is an AI customer support platform that resolves tickets end-to-end, processing refunds, updating subscriptions and handling multi-step workflows across chat, email and voice. For peptide and GLP-1 telehealth companies, Lorikeet resolves administrative tickets inside the pharmacy, billing and EHR systems already in use, and hands clinical questions to licensed staff with the conversation attached. ## Why Is Peptide Telehealth Support Harder Than Standard Telehealth? Peptide telehealth support is harder than standard telehealth support because 3 pressures land on the same queue: a regulatory calendar that changes what can be compounded, a medical-advice line that runs through routine questions, and a physical cold chain that fails on doorsteps. Each pressure creates tickets that a generic FAQ bot cannot safely close. ### Regulatory volatility drives ticket spikes The FDA declared the semaglutide shortage over on February 21, 2025, with compounding enforcement discretion ending April 22 for pharmacies and May 22 for outsourcing facilities. Then on February 27, 2026, HHS Secretary Robert F. Kennedy Jr. said publicly that most restricted peptides would leave Category 2, and the FDA formally removed 12 of them effective April 23, 2026, per Orrick. Every date produced a wave of "can I still get my medication?" tickets. ### The medical-advice line runs through routine questions "When do I take my next dose?" is a scheduling question until the patient adds that they skipped 2 weeks. California's AB 3030, effective January 1, 2025, exempts AI messages about appointments and billing but requires disclaimers and a human contact path once clinical information enters the message. Support automation has to detect that shift mid-conversation. ### Cold chain failures arrive as support tickets Semaglutide and most injectable peptides must stay between 2°C and 8°C, and the FDA has flagged compounded GLP-1s arriving warm with degraded ice packs. Unlike branded pens, compounded vials carry no temperature indicator, so the patient is the only quality checkpoint, and the support agent is the first person they tell. "Everyone wins when there's compliance. Consumers get products that are actually what the label says: sterile, verified, correctly dosed." - Jon Keidan, Founder and Managing Partner, Torch Capital, in Fortune ## Which Peptide Telehealth Support Tickets Can AI Resolve End-to-End? AI can resolve the administrative majority of peptide telehealth support tickets end-to-end: delivery and cold-chain claims, refill timing, subscription pauses and cancellations, refunds, payment failures and pharmacy-transfer notices. The rule is simple: if the resolution is an action in a system rather than a clinical judgment, an AI agent rather than a chatbot can own it. - Warm or late delivery claims. The agent checks carrier scans, confirms the excursion window, files a replacement with the pharmacy and tells the patient not to use the vial. Replacement-first policies close these in one contact instead of 3. - Refill and reorder timing. The agent reads the active prescription, the last ship date and the titration schedule, then confirms the next ship date or moves it. No dosing advice is given, only logistics. - Subscription pause, cancel and refund. The agent applies the published policy, processes the pro-rata refund or pause, and sends written confirmation. This is the single largest complaint category in BBB files, and it is fully automatable. - Payment failures and HSA or FSA receipts. Retry the card, update billing details, issue an itemised receipt. Our guide to automating failed payment recovery covers the workflow. - Pharmacy transfer and regulatory notices. When a substance changes FDA status, the agent explains the change in approved language, confirms the patient's options and books the clinician follow-up if a new prescription is needed. - Side-effect intake and escalation. The agent collects structured symptoms, flags red-line terms, and routes to a licensed clinician with the full transcript. The AI never interprets the symptoms. ## What Results Can Peptide Telehealth Companies Expect From AI Support? Peptide telehealth companies that automate administrative tickets can expect resolution rates in line with the broader AI support market, materially lower cost per ticket, and faster response on the time-sensitive tickets that drive refunds. The gains come from resolving actions, not from deflecting patients to articles. According to Intercom, its Fin agent averages a 76% resolution rate across 12,000 customers, while Ada reports about 52% ... --- ## Can a Support Chatbot Give Medical Advice? Telehealth Rules (2026) URL: https://www.lorikeetcx.ai/articles/chatbot-medical-advice-telehealth Updated: 2026-09-02 Keywords: can a support chatbot give medical advice, AI chatbot medical advice liability, telehealth chatbot compliance, California AB 3030 chatbot A telehealth support bot that answers "can I double my dose to catch up?" has just practised medicine without a licence, and the company owns the answer. A customer support chatbot cannot give medical advice. In 2026, US law and professional rules reserve clinical judgement for licensed providers, and AI-specific statutes in California, Utah and Illinois now regulate what an automated system may say to a patient. A support agent may handle scheduling, billing, shipping and account tasks, and must route dosing, symptom and treatment questions to a clinician. - California AB 3030 (effective January 1, 2025) requires a disclaimer and a human contact path whenever generative AI communicates clinical information to a patient. - Illinois fines up to $10,000 per violation for AI delivering therapy, while permitting administrative and supplementary support. - The FTC has already extracted $1.5 million from GoodRx and $7.8 million from BetterHelp over health-data handling, before any chatbot-specific case. - 32% of consumers used AI chatbots for health information in 2025 per Rock Health, so patients will ask your bot clinical questions whether you invite them or not. Last updated: September 2026 Telehealth companies, and peptide and GLP-1 providers in particular, feel this problem more sharply than most. Their patients self-inject at home, titrate doses on a schedule, and message support at 11pm when a vial arrives warm or a dose is missed. Every one of those messages sits close to the clinical line. A generic chatbot trained on the help centre will try to be helpful, and "helpful" is exactly the failure mode regulators are now writing rules about. The fix is not a disclaimer at the bottom of the chat window. It is an agent architecture that knows which questions it may resolve, which it must hand to a clinician, and how to make that handoff in a way that satisfies the law and the patient. ## Can a Customer Support Chatbot Give Medical Advice? A customer support chatbot cannot legally or safely give medical advice. Medical advice means interpreting a patient's symptoms, history or medication to recommend a course of action, and in the US that activity is reserved for licensed clinicians. A support chatbot may provide logistics, account and policy information, and may relay clinician-approved instructions verbatim. The distinction matters because the same sentence can be either. "Store the vial between 2°C and 8°C" is product information. "Your vial was warm for 6 hours, so it is still fine to inject" is a clinical judgement. Patients will ask anyway. According to Rock Health's 2025 consumer survey, 32% of US consumers used AI chatbots for health information, double the 16% in 2024, so patients arrive expecting the bot to answer. Gartner's July 2024 survey found 64% of customers would prefer companies did not use AI in customer service at all, which makes the trust cost of a wrong answer even higher. Medical advice: a recommendation about diagnosis, treatment, dosing or symptom management that depends on an individual patient's circumstances and requires clinical judgement. Administrative support: help with scheduling, orders, shipping, billing, subscriptions, account access and policy explanations that does not require clinical judgement. Lorikeet is an AI customer support platform that resolves tickets end-to-end, processing refunds, updating subscriptions and completing multi-step workflows across chat, email and voice. In telehealth, Lorikeet is configured to resolve administrative tickets fully and to stop at the clinical line, handing dosing, symptom and treatment questions to a licensed provider with the full conversation attached. ## Where Is the Line Between Support and Clinical Advice in Telehealth? The line between support and clinical advice in telehealth sits at individual judgement: a support agent may state facts, policies and clinician-approved instructions, but the moment an answer depends on this patient's symptoms, history or dose, it becomes clinical. Most telehealth tickets are administrative, and a minority cross the line mid-conversation. ### Clearly administrative Order status, cold-chain replacement claims, refill ship dates, subscription pauses, refunds, payment retries, appointment booking, insurance receipts and "how do I upload my ID" questions. California AB 3030 explicitly exempts messages about appointments, billing and business matters from its clinical-AI rules. These tickets can be resolved end-to-end by an AI agent, and our telehealth use-case guide walks through each one. ### Clearly clinical Dose adjustments, missed-dose instructions, side-effect interpretation, drug interactions, whether to continue after a warm shipment, and any "is this normal?" question. These require a licensed clinician. The AI's job is intake and routing, not answering. ### The grey zone that trips up bots "When is my next dose?" is administrative if the answer is a date from the prescription record, and clinical if the patient adds that they skipped 2 weeks. Good agents detect that shift and change behaviour mid-conversation, which is where AI guardrails earn their keep. "This is one of the most significant open questions in healthcare law right now, and the honest answer is that liability allocation is going to be highly fact-dependent and will likely take years of litigation to clarify." - Meghan O'Connor, Health Law Partner and Co-chair of the AI team, Quarles & Brady, in MedCity News ## What Laws Govern AI Chatbots in Patient Communication in 2026? In 2026, AI chatbots in patient communication are governed by a patchwork: HIPAA for protected health information, the FTC Act and Health Breach Notification Rule for consumer health data, FDA guidance on clinical decision support, and state AI statutes in California, Utah, Illinois and others. No single law covers a telehealth support bot, so compliance means satisfying all of them at once. - California AB 3030 (January 1, 2025). Health facilities, clinics and practices using generative AI to communicate clinical information must include a disclaimer and instructions for reaching a human. In chat, the disclaimer must persist throughout the interaction. Messages reviewed by a licensed provider are exempt. - Utah AI Policy Act and SB 226 (2025). Disclosure is required for "high-risk" AI interactions, which Future of Privacy Forum notes include collecting health data and providing medical advice. HB 452 adds rules for mental health chatbots. - Illinois Wellness and Oversight for Psychological Resources Act (August 1, 2025). Prohibits AI from providing therapy, with fines up to $10,000 per violation per the Illinois IDFPR. Administrative and supplementary support for licensed professionals remains allowed. - FTC Health Breach Notification Rule (updated 2024). The FTC extended the rule to health apps, after fining GoodRx $1.5 million and settling with BetterHelp for $7.8 million over sharing health data with advertisers. - FDA Clinical Decision Support guidance (final, January 2026). Per Manatt Health's policy tracker, software that makes patient-specific recommendations a clinician does not independently review can be a regulated device. A support bot that interprets symptoms drifts toward that definition. - HIPAA. Any AI vendor that touches protected health information needs a signed Business Associate Agreement. Our BAA and HIPAA guide lists what to verify. ## What Happens When a Support Bot Crosses the Line? When a support bot gives medical advice, the company inherits the clinical, regulatory and product-liability exposure that the bot cannot carry itself. Enforcement to date has focused on data handling and therapy bots, but 2025 court rulings and state attorneys general have moved chatbots squarely into existing liability ... --- ## GLP-1 and Peptide Telehealth Support Tickets: 7 to Automate (2026) URL: https://www.lorikeetcx.ai/articles/glp-1-peptide-support-tickets Updated: 2026-09-02 Keywords: GLP-1 telehealth support tickets, peptide telehealth customer support, GLP-1 patient support automation, telehealth ticket automation 30% of GLP-1 patients quit in the first month, and every one of them opens a ticket before they go. GLP-1 and peptide telehealth support tickets are the patient requests that follow an injectable prescription home: delivery and cold-chain problems, refill timing, subscription changes, refunds, payment failures, side-effect reports and questions triggered by FDA rule changes. In 2026 these tickets cluster around 7 types, and most of them are administrative actions that AI agents can complete end-to-end. - Blue Health Intelligence found 30% of GLP-1 patients stop in month 1 and 58% within 12 weeks, each exit producing a cancellation or refund ticket. - Prime Therapeutics reports 1-year persistence nearly doubled from 33.2% in 2021 to 60.9% in 2024, so the surviving cohort now generates years of refill tickets. - The FDA changed the status of 12 compounded peptides on a single day in April 2026, producing a same-day wave of "can I still get this?" tickets. - Healthcare tickets cost $30 or more each per LiveChatAI, and 5 of the 7 common ticket types need no clinical judgement. Last updated: September 2026 GLP-1 and peptide telehealth is a subscription business wrapped around a cold-chain pharmacy wrapped around a regulated prescription. Each layer generates its own ticket type, and the layers interact: a warm delivery becomes a refund request, a refund request becomes a cancellation, and a cancellation becomes a churn statistic. Meanwhile the regulatory layer resets the whole queue whenever the FDA moves a substance on or off a list. Support teams that treat this as one undifferentiated inbox burn out. Teams that classify the 7 ticket types, automate the administrative ones and route the clinical ones cleanly get faster resolutions, fewer chargebacks and clinicians who only see what needs them. ## What Are the Most Common GLP-1 and Peptide Telehealth Support Tickets? The most common GLP-1 and peptide telehealth support tickets are delivery and cold-chain claims, refill and reorder timing, subscription pause or cancellation, refunds and billing disputes, payment failures, side-effect intake, and regulatory-change questions. Delivery, billing and subscription tickets typically dominate volume; side-effect and dosing tickets are fewer but carry clinical risk. The billing share is unusually high for healthcare. Better Business Bureau complaint files for multiple peptide sellers show repeating patterns: recurring charges the patient did not expect, cancellations buried in marketing emails, and refunds refused on unopened product. Persistence data explains the volume. Blue Health Intelligence, analysing nearly 170,000 commercial members, found 30% stop within the first month and 58% within 12 weeks, and each exit is a conversation with support. Titration schedule: the clinician-set plan for increasing an injectable dose over weeks, which determines when each refill must ship. Cold-chain excursion: any period during shipping when a temperature-sensitive medication leaves its required range, for semaglutide 2°C to 8°C. Lorikeet is an AI customer support platform that resolves tickets end-to-end, processing refunds, updating subscriptions and completing multi-step workflows across chat, email and voice. For GLP-1 and peptide telehealth companies, Lorikeet resolves the administrative ticket types in this article inside existing pharmacy and billing systems and routes side-effect and dosing questions to licensed clinicians with full context. ## Why Do GLP-1 and Peptide Support Volumes Spike? GLP-1 and peptide support volumes spike for 3 reasons: early-treatment attrition concentrates cancellations in the first 12 weeks, regulatory events change what can be shipped on a fixed date, and the physical cold chain fails in heat waves and holiday shipping. None of these spikes align with a staffing calendar. ### Attrition is front-loaded Blue Health Intelligence's 30% first-month discontinuation rate means a clinic onboarding 1,000 patients a month should expect roughly 300 cancellation or pause conversations from that cohort within 30 days. Prime Therapeutics data shows 1-year persistence rising from 33.2% in 2021 to 60.9% in the first half of 2024, so the patients who stay now generate refill tickets for far longer. ### Regulatory dates move the whole queue The FDA declared the semaglutide shortage resolved on February 21, 2025, ending compounding discretion on April 22 and May 22, 2025. On April 23, 2026, 12 peptides left the Category 2 list per Orrick, with 7 more reviewed on July 23 and 24, 2026. Every date is a same-day ticket surge. ### The cold chain breaks at the doorstep The FDA has flagged compounded GLP-1s arriving warm with degraded ice packs. Compounded vials carry no temperature indicator, so the patient reports the problem and support decides what happens next. Fortune's July 2026 reporting sizes the addressable peptide market at roughly $30 billion and ties its growth to exactly this kind of compliance and fulfilment discipline. ## Which 7 GLP-1 and Peptide Support Tickets Should You Automate First? The 7 GLP-1 and peptide support tickets to automate first are delivery claims, refill timing, subscription changes, refunds, payment failures, regulatory notices and side-effect intake. The first 6 can be resolved end-to-end by an AI agent; the 7th is automated intake and routing, never automated advice. - Warm, late or missing delivery. Verify carrier scans and excursion window, instruct the patient not to use the vial, file the pharmacy replacement and confirm the new date. One contact instead of the 3 that manual back-and-forth usually takes. - Refill and reorder timing. Read the active prescription, last ship date and titration schedule; confirm or move the next shipment. This is logistics from the clinician's plan, not a dosing decision. - Subscription pause, skip or cancel. Apply the published policy, process the change, send written confirmation. Given 30% first-month attrition, this is the highest-volume ticket in the queue and the one BBB complaints most often cite. - Refunds and billing disputes. Check eligibility against policy and order history, issue the refund or credit, and document it. Our guide to AI for refund status covers the pattern. - Payment failures. Retry, update card details, prevent the involuntary cancellation. See automating failed payment recovery for the workflow. - Regulatory-change notices. Explain a status change in clinician-approved language, confirm what the patient can still receive, and book a provider consult if a new prescription is needed. Never imply a compounded product is the same as an approved one; the FDA warned 30 telehealth companies for exactly that. - Side-effect intake and escalation. Collect structured symptoms, detect red-line terms, and route to a licensed clinician with the transcript. Under California AB 3030 the agent must disclose AI use once clinical content appears. ## What Results Do Telehealth Teams See From Ticket Automation? Telehealth teams that automate the administrative ticket types see resolution rates comparable to the wider AI support market, lower cost per ticket and faster handling of the time-sensitive tickets that otherwise become refunds and chargebacks. The gains depend on resolving actions in systems, not deflecting patients to help articles. According to Intercom, its Fin agent averages a 76% resolution rate across 12,000 customers, and Ada reports around 52% automated resolution across more than 550 deployments. LiveChatAI's 2025 benchmarks put phone contacts at $17 to $25, chat at $10 to $16 and self-service at $1 to $4, with healthcare rarely under $30 per ticket. Gartner predicts agentic AI will autonomously resolve 80% of common service issues by 2029. For a telehealth company handling 8,000 tickets a month ... ---