The AI platforms that execute multi-step resolution, instead of just deflecting, are the ones that can verify a customer, read and write in your systems, follow a procedure in order and hand off with context when they reach a limit. In 2026 that list includes Fin (now part of Salesforce), Lorikeet, Zendesk AI agents, Decagon, Sierra, Ada, Salesforce Agentforce and Gradient Labs, each documented on its own site. They differ in how you write the steps, how you force the steps that must never vary, and how you test before customers see it.

Key takeaways

• Deflection measures whether the customer went away; resolution measures whether the problem was fixed. The CFPB found that bank chatbots' "effectiveness wanes as problems become more complex" and has received complaints about chatbot "doom loops".

• Check three things in a demo: write access to your systems, a way to force fixed steps (code, if/else rules or scripted flows), and simulations you can rerun after every change.

• If you already run Zendesk Suite, its AI agents are included in every Suite and Support plan and billed per automated resolution. Start there before buying a second vendor.

• Gartner predicts agentic AI will autonomously resolve 80% of common customer service issues by 2029, and also that over 40% of agentic AI projects will be canceled by the end of 2027. Buy for the hard tickets, and prove it in a pilot.

Which AI platforms execute multi-step resolution tasks?

Lorikeet runs sensitive steps such as payments and identity checks as code the agent can invoke but never alter, while the conversation around them stays in natural language. Fin's Procedures let support teams write multi-step processes in plain language, then add if/else logic and code to enforce rules like verifying eligibility or updating records. Zendesk says its AI agents handle complex, multi-step workflows across channels and connect to the systems you already run. Decagon's Agent Operating Procedures combine natural language with what Decagon calls "the reliability of code". The full list, with limits, follows.

How we chose and checked these

We included platforms that publish, on their own site, a way for the AI agent to take actions in other systems and to follow a multi-step procedure. Every vendor fact was checked on that vendor's public pages on 30 October 2026 and is linked where it appears. We did not use analyst estimates of contract value, review scores or private conversations. Where a vendor does not publish something, the table says so. The order reflects fit for the question in the title, meaning how clearly the vendor's own pages show finished multi-step work and how it is controlled. It is not a quality score.

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.

What does multi-step resolution mean in practice?

Multi-step resolution means the AI agent completes a dependent chain of actions for one request and can prove it did. Take "my card was declined and I think someone used it." A resolving agent verifies identity, pulls recent transactions, freezes the card through the processor's API, opens a dispute, tells the customer what happens next and passes the case to a person if the dispute rules need a human decision. Each step can fail on its own, so the agent also needs a plan for the unhappy path.

Anthropic's engineering guide draws a useful line: "workflows" orchestrate models and tools through predefined code paths, while "agents" dynamically direct their own process and tool use. Every platform below mixes the two. The model handles the conversation; fixed code, rules or scripted flows handle the steps that must not vary.

Comparison table

1. Fin: support teams that want to write procedures themselves, on almost any helpdesk

Fin's Procedures are built for queries "with multiple steps, business logic and third-party systems from start to finish." You write them like a document, and Fin reasons at every turn, skipping to the right step or switching Procedures when the customer changes direction. Where precision matters you add if/else branches and code, and you connect Stripe, Shopify or internal systems through Data Connectors or MCP. Simulations run full conversations against success criteria, and you can rerun saved tests whenever a Procedure changes.

Fin runs on the Intercom helpdesk and, per the same page, with Zendesk, Salesforce or Freshdesk. Pricing is $0.99 per outcome; you are not charged when a conversation is simply passed to your team without an outcome.

Trade-off: a completed Procedure handoff to a human also counts as a $0.99 outcome, so track resolutions and handoffs separately when you work out cost per solved problem. Salesforce completed its acquisition of Fin on 10 September 2026; if you run a non-Salesforce stack, ask how that affects the integrations you rely on.

2. Lorikeet: regulated teams whose tickets need identity checks, payments and an audit trail

Lorikeet acts in your systems through APIs, webhooks and MCP servers, and its how-it-works page says a team of specialist agents can coordinate with third parties by phone, SMS, email and Slack to resolve multi-party issues. Strict steps run as deterministic code; the product docs call code tools the main primitive for deterministic, multi-step action-taking, such as checking order status and then issuing a refund. Before launch you replay real tickets and adversarial scenarios in bulk. One agent covers chat, email, voice, SMS and WhatsApp.

On its customer pages, Carmoola, a UK car finance lender, moved from automation built on a knowledge base and rigid, form-driven prompts that resolved about 30% of inbound questions to 60% of all inbound on WhatsApp, chat and email resolved end to end. Hnry reached around 70% automation in the peak week of Australian tax season.

Pricing: Start $2,100 a month and Scale $5,100 a month, paid annually, plus $0.99 or $0.90 per chat, email or SMS resolution and $1.50 or $1.20 per voice resolution (3 minute average). Escalations and unresolved tickets are not charged. There is a 30-day free trial.

Limits: the plan fee makes it poor value for a team with a few hundred simple FAQ tickets a month, where helpdesk-native AI costs less. There is no built-in Freshdesk connector; webhook-based helpdesks connect through Flexible Ticketing instead.

3. Zendesk AI agents: teams already running Zendesk Suite

Zendesk says its AI agents resolve "multi-step, multi-intent requests" by asking clarifying questions and taking action across systems. Its help center lists the building blocks as goal-oriented generative procedures, scripted dialogues, authorized actions and API integrations. A generative procedure can insert actions, API integrations and links to other procedures or dialogues, so a flexible flow can hand a step to a scripted one.

The commercial case is strong for existing customers: AI agents come with every Suite and Support plan and are billed per automated resolution, which Zendesk defines as a request resolved by the AI agent without escalation to a human. For a Zendesk shop with mostly order, account and policy tickets, this is the lowest-friction way to test multi-step automation.

Trade-off: the native AI agents live inside Zendesk. For other helpdesks Zendesk points to Forethought AI agents, which it says work on any platform, and which are sold through its sales team.

4. Decagon: enterprises that want natural-language procedures with engineering controls

Decagon's AOPs let teams "define agent behavior in natural language, the same way you train human agents with SOPs," and combine that with code for reliability. Engineers can version AOPs with Git-based tracking, and Decagon shows the reasoning behind each agent decision, with runtime and latency numbers. AOPs pull live data and trigger workflows across connected systems. Decagon covers voice, chat and email, and offers simulations at scale for testing.

Decagon's Duet assistant can draft AOPs from uploaded conversation transcripts, and guardrails for brand voice, escalations and hallucinations apply across channels. It suits enterprises with an engineering team that wants procedures reviewed like code.

Trade-off: pricing is not published on the pages we checked, so cost per resolution only becomes clear in a sales process.

5. Sierra: large brands that want outcome-based pricing

Sierra's Agent Studio builds agents from "composable building blocks," lets you define step-by-step workflows and puts integrations to work inside journeys, "from taking real-time action on behalf of customers." It lists 40+ pre-built integrations, simulations and a regression test suite. On its trust page, Sierra says that when the agent accesses your systems of record, "those interactions are deterministic and controlled." Sierra covers voice, chat, email and WhatsApp, and its product page describes outcome-based pricing where you pay "only when the software achieves specific, valuable outcomes."

Agent Studio is pitched as no-code, with real-time traces of every decision, tool call and response while you build, and Sierra says it wraps models in supervisory layers to manage their non-deterministic nature. It suits large consumer brands that want one vendor across many channels and languages.

Trade-off: per-outcome rates are not published, so model your cost from a quote, not the pricing model alone.

6. Ada: high-volume consumer brands that want actions run outside the model

Ada's platform page describes Playbooks as "structured workflows for handling complex tasks" that follow multi-step SOPs using real-time data. Its Ada Computer layer carries out the steps the reasoning engine chooses, and Ada states that "refunds, entitlements and eligibility come from your sandboxed Python, not from the model." Tools from a connected MCP server stay off until you turn them on. Ada deploys across voice, messaging and email.

Ada describes the loop as decide, act, check, decide, running until the task is done, and connects to systems such as Zendesk, Salesforce and Twilio. That design suits high-volume brands that want the model to choose the next step but never compute a refund amount itself.

Trade-off: pricing is not published on the pages we checked.

7. Salesforce Agentforce: organisations whose support already runs on Salesforce

Agentforce agents are built in Agent Builder from subagents and actions, using Flows for automation, MuleSoft connectors to reach other systems, and Apex or JavaScript for custom logic. Salesforce's Agentforce guide describes hybrid reasoning, which combines LLM reasoning with deterministic, rules-based execution through Agent Script. Pricing is consumption-based through Flex Credits or Conversations, or per-user licensing.

Salesforce says the Atlas Reasoning Engine breaks a request into smaller tasks, evaluating each step and proposing a plan until the answer or action is complete, and that Agent Script's text format lets teams version, review and govern agent behaviour like code. Agents escalate to human agents when an issue is beyond their scope.

Trade-off: the strongest fit is when your customer data and processes already live in Salesforce; outside that, you are building on a platform rather than buying a support product.

8. Gradient Labs: financial services operations teams

Gradient Labs focuses on customer operations in finance. It says its agent executes natural-language procedures, runs financial-services guardrails on every turn and can automate disputes end to end: it gathers and verifies evidence, decides the outcome, escalates to a human for sign-off and files the chargeback with the card scheme. It cites SOC 2 Type 2 certification.

It also says it ships with 20+ guardrails out of the box and learns from real conversations, not just the help centre, and it extends the same agent to collections outreach. It suits banks, lenders and payment companies that want a finance-only specialist.

Trade-off: it is built for finance only, and pricing is not published.

What still needs a human?

A multi-step agent should stop and escalate when the next step needs judgement, authority or accountability that your policy gives to a person. In practice that means hardship and vulnerability cases, disputes above a value threshold, complaints that may become regulatory matters, and any request where the customer contests what the system shows. OWASP's guidance on excessive agency recommends human-in-the-loop approval for high-impact actions before they are taken. NIST's AI Risk Management Framework, released 26 January 2023 for voluntary use, is a sensible structure for writing down who signs off on which actions. Compliance sign-off on the procedures themselves stays with your team.

Questions to ask before you pick

1. Show one real ticket where the agent ran four or more dependent actions and finished. Which steps were model-driven and which were fixed code or rules?

2. What happens when the third system call in a chain fails: retry, roll back or escalate with context? What does the customer see?

3. Can we replay our own historical tickets and adversarial prompts before go-live, and rerun them after every change?

4. What exactly is billed: resolutions only, or handoffs and other outcomes too?

5. Can we set read-only or write access per system, and see an audit trail of every action?

For more on measuring this, see what AI resolution rates hide, per-resolution versus deflection pricing, agentic AI versus workflow chatbots and regression testing an AI support agent. To test multi-step workflows on your own tickets, start a 30-day free trial.