Add an AI step to a workflow

An AI step is just another action node in the visual flow — but instead of a fixed rule, it runs a real Claude model on the data the workflow already has. You can use it to summarize a record, draft an email, classify an input into a category, or suggest a next best action, and then use whatever it produces in the rest of the flow.

This guide shows where AI steps fit, how to give one good instructions, and how to wire its output into the actions and branches that come after — so the AI does the thinking and the workflow does the doing.

Who this is for

Anyone building a workflow who wants it to draft, summarize, classify, or recommend — rather than only move data around.

Before you start

• A workflow to add the step to (see "How to build your first workflow").
• A clear job for the AI — for example, "summarize this account's recent activity" or "classify this inbound message."
• Records with enough context for the model to work from (a deal with a timeline, a form with a message field, etc.).

These are real AI steps, not a gimmick
The four AI actions — summarize, draft email, classify, and next-best-action — are real Claude-powered steps that run on your workflow's data. They're live today and behave like any other action node: configured in the side drawer, and their output is available to later steps. We don't publish accuracy figures; treat the output as a strong first draft you can route, review, or act on automatically depending on how much you trust the task.

Add and wire an AI step

1. Add the AI action node
   In the flow, add an action and choose the AI step you want: summarize, draft email, classify, or next-best-action. It sits in the flow like any other node, after the trigger and any conditions you've set.
2. Point it at the right input
   In the side drawer, choose what the step reads from — the triggering record and its related data, like a deal's timeline or a form submission's message. The model only sees what you give it, so include the context the task needs.
3. Write a clear instruction
   Tell the step exactly what you want: "Summarize the last 30 days of activity on this account in three bullet points," or "Draft a friendly follow-up email referencing the demo." The more specific the instruction, the more useful and consistent the output.
4. Use the output downstream
   An AI step produces output the rest of the flow can use. A draft-email step can feed the body of a send-email action; a classify step can produce a category you branch on; a summarize step can populate a field or a Slack message.
5. Branch on a classification
   When the AI classifies an input, add a branch after it and route on the result — "billing" goes one way, "sales" another, with an else path for anything unexpected. This is how an AI step turns free text into a decision the workflow can act on.
6. Decide how much to automate
   Choose how much trust the task deserves. A low-risk summary can flow straight into a field; a customer-facing draft might be created as a draft for a human to review before sending. The workflow runs either way — you decide where a person stays in the loop.
7. Dry-run, then turn it on
   Dry-run the flow on a sample record to see what the AI step produces and which branch the result takes. When it looks right, turn the workflow on — the AI step then runs on every matching event as part of the async flow.

What you get

• Workflows that summarize, draft, classify, or recommend — not just move data.
• AI output wired straight into later actions, fields, or branches.
• Free-text inputs turned into decisions a workflow can route on.
• A clear choice, per step, about where a human reviews before the flow acts.

Frequently asked questions

What can the AI steps do today?

Four things, all live: summarize a record's recent activity, draft an email, classify an input into a category, and suggest a next best action. They're real Claude-powered steps that run on the data the workflow already has, and their output can flow into later actions, fields, or branches.

Is the AI step actually running a model, or is it a canned template?

It's a real Claude model running on your workflow's data — not a fixed template. You write the instruction, point it at the right input, and it produces output for that specific record. That's why a summarize step reads differently for every account, and a draft-email step references the actual deal.

Can I trust an AI step to act without me?

That's your call, per step. For low-risk tasks — summarizing activity into a field, classifying a message to route it — letting the flow act automatically is fine. For customer-facing output like a drafted email, many teams have the workflow create a draft for a human to review before it sends. The workflow supports both; you decide where to keep a person in the loop.

How do I use what an AI step produces?

Reference its output in the steps that follow. A draft-email step's text can become the body of a send-email action; a classify step's category can drive a branch; a summarize step's text can populate a field or a Slack post. The output is available to everything downstream in the flow.

Will the AI ever see data I didn't give it?

No — an AI step only works with the input you point it at, like the triggering record and the related data you choose in the side drawer. If you don't include something, the step doesn't see it. That's also why a good instruction names the context you want the model to use.

Learn more

• Workflows (feature) — The visual flow builder and its AI steps on the feature page.