How to Add an AI Step to Your Zapier Workflow

Introduction

Zapier used to keep its agentic AI in a separate product called Zapier Agents. That product is gone. Its capability now lives inside the regular Zap editor as a single action, AI by Zapier, which means you can drop reasoning, summarising, classifying, and tool use into a workflow you already run, without rebuilding anything. This tutorial walks through adding that step to an existing Zap, from choosing a model tier to publishing.

Quick answer

Open your Zap, add an action step, search for AI by Zapier, pick a model tier, write a prompt with fields mapped from earlier steps, define output fields, preview, publish. Standard costs 1 task per run, Advanced 3, Premium 5.

Evidence

Zapier describes the feature in its own help center in one sentence: “AI by Zapier is a built-in tool that lets you add AI-powered steps to your Zaps, powered by popular AI models.” The pricing is tiered by model strength, and the numbers matter because they multiply fast. A Standard step costs 1 task per run, Advanced costs 3 tasks, and Premium costs 5 tasks, and each tool call the AI makes is billed again at the same tier rate. The help center also sets hard limits on reference material: you can attach up to 20 knowledge sources to one AI step, each source syncs at most 96,000 words, and a directly uploaded file can be at most 100 MB. The official AI by Zapier help article documents the tiers, the Free plan limits, and the knowledge source caps, and Zapier’s own AI by Zapier guide walks through the same setup from the product side.

Prerequisites

• A Zapier account. A Free account is enough to try the step, but publishing a Zap that contains it needs a paid plan.
• An existing Zap with a working trigger, or a clear idea of the trigger you want. This tutorial uses a Gmail trigger as the running example, but any trigger works.
• The app accounts the AI step will act on, if you plan to give it tools, such as Slack or Google Sheets. Tools are only available on the Advanced and Premium tiers, or with Bring Your Own Key.
• A rough idea of what the AI should return. You will turn that into output fields in step 5, so decide the field names before you start clicking.

Step 1: Pick the Zap and the spot for the AI step

Open the Zap you want to make smarter and look at where the decision happens. The AI step belongs at the point where a fixed rule stops being enough: classifying an email, summarising a long message, drafting a reply from context, or extracting structured data from free text. In the running example, the Zap triggers when a Gmail message gets a specific label, and the AI step sits between that trigger and a Slack message that posts the summary.

Step 2: Add the AI by Zapier action

In your Zap, click the + icon to add an action step. In the search box, type AI by Zapier and select it. The step opens in the Zap editor with two panels: Configure on the left and Preview on the right. Configure is where you write the prompt, pick a model, and set up output fields, tools, and knowledge. Preview is where you test the step against real sample data from your trigger before you save.

If you do not see AI by Zapier in the search results, check that you are adding an action step and not a trigger step. The AI step is an action, so it can only sit after a trigger.

Step 3: Choose a model tier before you write anything

Click the model dropdown at the bottom right of the Prompt field. Zapier offers four tiers, and the choice is a cost decision as much as a quality decision:

• Standard (1x tasks): best for high-volume, simple tasks such as short summaries and classification. Tools are not available on this tier.
• Advanced (3x tasks): stronger reasoning for complex prompts and tool usage.
• Premium (5x tasks): complex analysis, tool usage, and multi-step reasoning. This is the default selection for new steps, so change it deliberately if you do not need it.
• Bring Your Own Key (1x tasks): connect your own AI provider account and pay that provider directly, while Zapier bills the step at the 1x task rate, including tool calls.

You can connect your own provider from OpenAI, Anthropic, Google, Azure OpenAI, or Amazon Bedrock. Pick the tier first because it changes what the step can do: a Standard step cannot call tools, so a prompt that says “send the summary to Slack” will fail at the tool stage if you stay on Standard. For the Gmail example, a short summary fits Standard, while a step that also files the email or updates a sheet needs Advanced or Premium.

Step 4: Write the prompt in plain language

In the Prompt field, write what you want the step to do as if you were explaining the task to a colleague. A prompt that works well names the input, the job, and the shape of the answer. For the Gmail example:

Summarise the email below in two sentences: one sentence for what the sender wants, one for the deadline or action needed. If there is no deadline, say “no deadline”. Use this email: /

Type a forward slash (/) inside the prompt to insert a field from a previous step, such as the email body or subject from the Gmail trigger. Mapping real fields is what makes the step reusable: the same prompt runs on every new email instead of being written for one message.

Zapier offers pre-built prompt templates under Use a template, in the Summarize, Write, Classify, and Extract categories. A template is a good starting point, but edit it to name your fields and your output shape before you test. Two details from the help center are easy to miss: prompt versions are only saved within a session, so only the current version survives after you close the editor, and the AI does not learn from previous runs, so anything it needs must be in the prompt, the mapped fields, or a knowledge source every time.

Step 5: Define output fields so later steps can use the result

Under Settings, open Output Fields and click + Add field. For each field, enter a name, pick a type, and add a short description. For the Gmail example, add summary (text, required) and deadline (text, optional). You can also click Generate from prompt and let Zapier propose fields from your prompt, then edit the names to match what your later steps expect.

This step is the difference between a demo and a working automation. Without output fields, the AI step returns one combined block of text, and the Slack step after it has to parse that text by luck. With fields, the Slack step maps summary and deadline as separate values, and a missing deadline shows up as an empty field instead of a sentence your team has to read. Mark a field required when a later step breaks without it. If you need several results from one prompt, turn on Return a list of results under Advanced to get line items instead of a single object.

Step 6: Add tools and knowledge, and gate anything dangerous

Open Tools and Knowledge under Settings only if your tier supports it. Tools let the AI step act in other apps, for example looking up a customer in a CRM or adding a row to a sheet, and knowledge sources give it reference documents such as a pricing page or a support playbook. Keep the tool list short: every tool you allow is an action the model can take without asking, and every tool call is billed again at your tier rate.

Two guardrails matter here. First, the Require approval before running toggle is off by default for each tool, so a tool that sends email or deletes records will run without pausing unless you turn approval on for that specific tool. Our guide to setting up per-tool approval for AI by Zapier steps walks through turning it on for write actions. Second, pair tools with AI Guardrails when the step touches customer data, so blocked topics and sensitive content are stopped before the model acts.

For knowledge, note the caps: at most 20 sources per step, 96,000 synced words per source, and uploaded files capped at 100 MB. A source larger than that may time out when you add it, and the help center warns that sources with 1,000 or more pages or files can time out on add. Add the smallest document that answers the question, not your whole drive.

Step 7: Preview, fix the prompt, then publish

Click Preview in the right panel to run the step against your trigger sample. Read the output the way the next step will read it: are the fields filled, is the deadline actually a date, does the summary mention the sender’s request? If the answer is vague, tighten the prompt with one concrete instruction, such as “quote the deadline exactly as written in the email”, and preview again. If the output is right but a field is empty, check that the field is marked required and that its description tells the model where the value comes from.

When the preview is correct, click Finish, then publish the Zap. On a Free plan this is where you hit the wall: the plan only lets you preview results with the Standard tier, and you cannot publish a Zap containing an AI by Zapier step. Upgrade to a paid plan, or switch the step to Bring Your Own Key, before you expect the Zap to run on live triggers.

What it costs in practice

Task math decides whether the AI step is cheap or ruinous, so work it out before you publish. A Standard step that only summarises costs 1 task per run. At 200 labelled emails a month, that is 200 tasks, a rounding error on most paid plans. The same step on Premium costs 5 tasks per run, or 1,000 tasks for those 200 emails, and if it calls two tools per run, the bill is 5 for the step plus 5 per tool call, which is 15 tasks per run, or 3,000 tasks a month.

The practical rule is simple: use Standard for summarise, classify, and extract jobs with no tools; use Advanced or Premium only when the step must reason or call tools; and use Bring Your Own Key when you run high volume and already pay an AI provider, since it holds the 1x rate including tool calls. Check the current Zapier pricing page for the task allowance on your plan before you scale a Zap that runs hundreds of times a day.

Troubleshooting

The AI step is missing from the action search. You are probably adding a trigger step, or your account admin has restricted the app. AI by Zapier is an action, so add it after a trigger. If it is still missing, check with your workspace admin, because admins can restrict which models and apps accounts are allowed to use.

Preview works, but the published Zap never runs the AI step. On a Free plan, a Zap containing an AI by Zapier step cannot be published at all, so the Zap stays off. Upgrade the plan, or confirm the Zap actually published after the upgrade. Also check Zap History: a trigger that never fires means the problem is the trigger, not the AI step.

The AI step runs but returns an empty field. The model did not find the value in the mapped input, or the field description is too vague. Open Output Fields, rewrite the description to say exactly where the value lives, for example “the date the sender asks for a reply, copied exactly from the email body”, mark the field required, and preview again with a sample that contains a real deadline.

The step is burning far more tasks than expected. Check the tier first: Premium is the default for new steps, and it costs 5 tasks per run before any tool calls. Drop to Standard if the job is a simple summary with no tools, or switch to Bring Your Own Key for high volume. Then count tool calls in the run history, since each call is billed again at the tier rate.

The AI step fails with a content error. AI by Zapier uses large language models from Zapier’s AI subprocessors, and a step can fail if its content violates their usage policies. Content moderation is applied by default on Zapier-provided models and cannot be switched off; with Bring Your Own Key you can switch it off. If the failure is a false positive on ordinary business text, try BYOK, or reword the prompt so the model is not asked to reproduce the flagged content.

Conclusion

Adding an AI step to a Zapier workflow is a ten-minute job once you know the three decisions that matter: the tier, the prompt, and the output fields. Pick Standard for simple text jobs and Advanced or Premium only when the step must reason or call tools, write the prompt around mapped fields instead of pasted text, and define output fields before you publish so later steps get clean data. Keep tools few, turn on approval for anything that writes or sends, and do the task math before the Zap goes live. If you came from the old standalone product, start with our guide to migrating Zapier Agents to AI by Zapier, and if you want the same AI step pattern in a self-hosted stack, compare it with the Zapier MCP versus n8n MCP walkthrough.

Frequently asked questions