Agentic AI for customer support reads a whole message, works out what the customer needs, and takes actions in your systems until the issue is resolved or handed to a person. A workflow chatbot follows an if-this-then-that flow someone drew in advance: it matches a keyword or button, walks the branch, and stops when the customer says something the flow did not expect. Most teams need both, with agentic reasoning for the conversation and deterministic steps for anything regulated, such as identity checks, payments and claims decisions.

This guide explains the difference, shows where each one breaks, and compares seven platforms on what their own pages say about actions, workflow builders, multi-part messages and testing. It does not rank them. It goes deeper on the comparison than our explainer What is agentic AI?, which covers the definitions.

Key takeaways

• "Agentic" is now a marketing word. Gartner describes "agent washing", the rebranding of chatbots and RPA without substantial agentic capabilities, and estimates only about 130 of the thousands of agentic AI vendors are real.

• The test is actions, not answers. A truly agentic system reads data, changes records and completes the task inside your systems, then escalates with context when it should not decide.

• Deterministic steps still matter. The platforms in this list that call themselves agentic all pair reasoning with some form of hard-coded control for sensitive steps.

• Insist on a test before launch. Ask every vendor to replay your own historical tickets, including messy multi-part ones, before any customer sees the agent.

Agentic AI vs workflow chatbots: what is the difference?

The difference is who decides the next step. In a workflow chatbot, the builder decided every step in advance. In an agentic system, the model decides the next step at runtime, inside limits you set. The CFPB describes rule-based chatbots as using "decision tree logic or a database of keywords to trigger preset, limited responses", and notes customers who filed complaints about getting stuck in "doom loops" when their issue fell outside what the bot could handle (CFPB, Chatbots in consumer finance).

When are deterministic workflows still the right tool?

Deterministic workflows are the right tool whenever a step must happen the same way every time and someone may later ask you to prove it did. Identity verification, payment capture, disclosures a regulator requires and claims decisions belong here. OWASP's guidance on prompt injection recommends handling privileged functions "in code rather than providing them to the model" and requiring human approval for high-risk actions (OWASP LLM01:2025). In practice, most production designs are hybrid: the conversation is agentic, the sensitive operation is not. Lorikeet, for example, describes "pockets of determinism" where payments and identity checks "run as code the agent can invoke but never alter" (How Lorikeet works).

How we chose and checked these platforms

We included platforms that describe themselves as AI agents for customer support, that publicly document how their agent takes actions in other systems, and that a support leader is likely to see on a 2026 shortlist. Every vendor fact below was checked on that vendor's own public pages on 22 Oct 2026, and each section links to them. Where a vendor does not publish something, the table says "Not published on pages checked". Order is not a ranking. Lorikeet publishes this guide and is one of the tools listed. We have tried to describe every product, including our own, using only what each vendor publishes, and we link to the source for each claim.

Which vendors are truly agentic? Comparison table

1. Lorikeet: for complex and regulated support that mixes agentic and deterministic steps

Lorikeet describes itself as "one agent that resolves issues end-to-end across chat, email, voice and SMS", which plugs into your existing ticketing system or its own chat widget and acts through your APIs, webhooks and MCP servers (How it works). The agent works inside both natural-language and deterministic workflows, and sensitive steps run as code the agent cannot change. In Lorikeet's documentation a concierge is assembled from a prompt plus shared skills, where a skill is a reusable standard operating procedure, and any decision "different people would answer the same way given the inputs" is pushed into a code tool rather than left to the model.

Testing is simulation-first: scenarios generated from your real tickets run in bulk, batch diffs show what a change broke, and you can author adversarial scenarios such as prompt injection and false authority claims (Simulations). Guardrails combine code-enforced limits with checks on incoming and outgoing messages. Carmoola, an FCA-regulated car finance lender, says its earlier automation on "a knowledge base and rigid, form-driven prompts" resolved about 30% of inbound questions, and Lorikeet now resolves 60% of all inbound on WhatsApp, chat and email end to end (Carmoola story).

Pricing: Start is $2,100/mo and Scale $5,100/mo, paid annually, plus $0.99 or $0.90 per chat, email or SMS resolution; escalations are not charged. Limit: by its own FAQ, setup is "more than a browser extension": you connect systems and train before go-live, and the plan fee is a fixed commitment that small teams should weigh.

2. Fin (now part of Salesforce): for teams that want procedures inside or alongside Intercom

Fin, which Salesforce completed its acquisition of on 10 Sep 2026 (Salesforce), uses Procedures to handle "queries with multiple steps, business logic and third-party systems from start to finish" (Fin Procedures). Its page states the agentic versus workflow question directly: "Procedures are linear, but conversations aren't", so Fin reasons at every turn, can skip to the right step or switch Procedures, and "combines agentic behavior with deterministic control" through if/else logic and code that enforces rules like eligibility checks. It connects to systems through Data Connectors or MCPs, and simulations run full conversations to catch regressions.

Best for: teams already on Intercom, though Fin's pricing page says it works with any helpdesk, including Salesforce and Zendesk. Pricing: $0.99 per outcome (Fin pricing). Trade-off: the deepest integration is with the Intercom helpdesk, so teams on another helpdesk should test the handoff carefully.

3. Sierra: for large enterprises that want a partner-led build

Sierra says its agents "securely connect to your systems of record" to handle tasks "from start to finish, like processing an insurance claim, returning an order, or originating a mortgage" (Sierra product). In Agent Studio, teams build Journeys from "composable building blocks" that adhere to goals and guardrails, and can generate step-by-step instructions with AI (Agent Studio). Testing covers AI-driven simulations, evaluations against defined outcomes, and regression suites before release. Agents run across voice, chat, email and WhatsApp in 59 languages, built once and deployed everywhere. Sierra pairs the platform with "an expert agent development team" and says 40% of the Fortune 50 partner with it.

Best for: enterprises that want a vendor team involved in building the agent. Pricing: outcome-based, with no public rate card on the pages checked. Trade-off: Sierra quotes deployment "in weeks", so plan for a project rather than a same-day trial.

4. Decagon: for CX teams that want to own agent logic as written SOPs

Decagon's Agent Operating Procedures let teams "define agent behavior in natural language, the same way you train human agents with SOPs" (Decagon AOPs). The page contrasts this with "rigid frameworks" that "break when conversations don't follow the script", which is the workflow chatbot failure mode in a sentence. Technical teams can version AOPs with Git-based tracking, guardrails cover brand voice, escalations and hallucinations across chat, voice and email, and teams can see an explanation of the agent's reasoning at any point in a conversation with runtime and latency numbers. The product menu lists simulations and live A/B testing.

Best for: teams where support operations, not engineering, should own day-to-day agent changes, with engineers keeping version control. Pricing: not published on the page checked. Trade-off: the AOP page focuses on authoring and visibility; ask in a demo exactly which steps can be forced to run as code.

5. Zendesk AI agents: for teams already running Zendesk Suite

Zendesk says its AI agents "reason through multi-intent requests, ask clarifying questions and adapt in real time, without rigid scripts guiding them", and that you can describe workflows in natural language and "generate procedures dynamically" (Zendesk AI agents). Agents can be given access to your tools and systems to act, including in legacy environments, and built-in QA evaluates every interaction. On messaging, agents clarify intent and complete complex resolutions within a single conversation; on voice, they handle complex call flows from start to finish. On email, Zendesk describes agents that "apply policies and execute multi-step resolutions end to end".

Best for: teams whose helpdesk is already Zendesk and want the AI inside the same platform. Pricing: AI agents are included in every Suite and Support plan, billed per automated resolution (Zendesk pricing). Trade-off: the pages checked do not describe pre-launch simulation, so ask how you test a procedure change before it goes live.

6. Salesforce Agentforce: for Salesforce-centred organisations

Agentforce starts with the Atlas Reasoning Engine, which "breaks down the initial prompt into smaller tasks, evaluating at each step and proposing a plan" until the answer or action is complete (Agentforce). Builders configure subagents, actions and instructions in Agent Builder, and Agent Script "gives builders precise control" for reliable results, which is Salesforce's route to deterministic behaviour inside an agent. Observability tools monitor agent performance in near real time, and Agentforce Voice brings AI-powered voice to every channel. Salesforce's FAQ sums up the pitch: reason through complex requests, retrieve knowledge and take action on your behalf.

Best for: organisations whose customer data and service desk already live in Salesforce. Pricing: Flex Credits, $2 per conversation, or per-user licensing (Agentforce pricing). Trade-off: the value depends heavily on how much of your data sits in Salesforce; teams on other stacks will do more integration work.

7. Gradient Labs: for financial services operations teams

Gradient Labs describes "a suite of specialist agents for lending, disputes, and KYC" that runs natural-language procedures mirroring your operations, learned from your SOPs and real conversations (Gradient Labs). It says it plugs into "the financial systems where the work actually happens", runs "20+ guardrails on every turn" specific to financial services, and lists coverage of FCA Consumer Duty, Reg E, Reg Z, PSD2, GDPR and the EU AI Act. Enterprise features include SSO, audit logs and role-based permissions, with automated quality checks. It runs on models from OpenAI, Anthropic and Google, with failover across multiple cloud and LLM providers.

Best for: lenders, card issuers and payment companies. Pricing: not published on the page checked. Trade-off: it is built for financial services, so insurers and other sectors should check how much of the guardrail set applies to them.

Insurance example: what does truly agentic look like for policyholders?

Take one message: "My car was hit last night, is my excess covered, and can you move my premium date?" A workflow chatbot picks one branch. An agentic system starts first notice of loss intake, looks up the policy to answer the excess question from the actual cover, and changes the premium date through a deterministic billing step, escalating with context if the claim needs an adjuster's judgement.

Regulators expect that control. The NAIC model bulletin says insurers should maintain a written program for AI systems that "make, or support decisions related to regulated insurance practices", with controls proportionate to factors including "the extent to which humans are involved in the final decision-making process" and reliance on third-party AI systems (NAIC Model Bulletin). The NIST AI Risk Management Framework, released on 26 January 2023, is a common reference for that program. Lorikeet's insurance page describes deterministic workflows for regulated steps and escalation when underwriting judgement or a sensitive claims decision is needed. For channel-specific detail see AI voice agents for FNOL intake and AI support for insurers with strict audit needs.

How do I test a vendor's agentic claim in a demo?

Bring your own tickets and run this checklist; a truly agentic platform passes most of it live.

1. Send a three-part message. Does it handle all three requests, or only the first?

2. Change your mind halfway. Switch goals mid-conversation and see whether it follows or restarts the flow.

3. Ask it to do something, not explain it. Watch it read and write a record in a sandbox system.

4. Show the hard boundary. Ask which steps run as code the model cannot alter, and try to talk it past one.

5. Replay history. Ask for a batch run on your past tickets and a diff after one instruction change.

6. Trigger an escalation. Check what context the human receives.

7. Read the bill. Ask how escalated and failed conversations are charged.

Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value or inadequate risk controls (Gartner), so the last two questions matter as much as the first five. For whether to add an agent on top of your helpdesk or use its built-in AI, see AI layer vs helpdesk built-in AI.

What still needs a human?

Even a truly agentic system should not own judgement calls with real consequences for the customer. Claims decisions that need an adjuster, coverage disputes, complaints a regulator may review, vulnerable customers and anything outside your written policy should escalate with full context. A person on your team should also sign off on every procedure that touches a regulated step, and someone should review guardrail events and escalations weekly. OWASP's advice to require human approval for high-risk actions is a good default for any vendor on this list.

If you want to run the demo checklist on your own tickets, Lorikeet has a 30-day free trial.