Adding AI to Your Zapier Automations

Once you have built a basic Zap, the next step that makes the biggest difference is adding an AI action. This is what turns a simple data-moving automation into something that can actually read, understand, and respond to content.

This tutorial walks through exactly how to do that, with examples you can use right away.

What an AI step does inside a Zap

Zapier has a built-in AI action called "AI by Zapier" that lets you write a prompt and have it run as part of your automation. Every time the Zap fires, the AI reads the incoming data and produces an output you can use in the next step.

For example:

• A new email arrives (trigger)
• The AI reads the email body and classifies it as a question, complaint, or general inquiry (AI step)
• The classification gets added as a tag in your CRM (action)

Or:

• A form gets submitted with customer feedback (trigger)
• The AI summarizes the feedback into two sentences and scores the sentiment (AI step)
• The summary and score get added to a Notion database row (action)

The AI step sits between the trigger and the action, transforming or enriching the data before it moves on.

Adding an AI step to an existing Zap

Here is how to add an AI step inside Zapier's editor:

1. Open the Zap you want to edit, or start a new one
2. Click the plus button between your trigger and your final action
3. Search for "AI by Zapier" in the app list
4. Choose the action type: most of the time you want "Single Prompt"
5. Write your prompt in the text field (more on this below)
6. Map the output of the AI step to your final action

The output of the AI step becomes a piece of data you can use anywhere else in the Zap, just like any other field.

Writing prompts that work inside a Zap

The prompt is where people get stuck. The challenge is that you are writing a prompt that will run on different inputs every time. You need to write it so it works whether the email is long or short, formal or casual, about billing or about a product question.

A few principles that help:

Be specific about the output format. If you need a single word (like a category label), say so. If you need a JSON object, describe the exact keys. Vague prompts produce inconsistent outputs that are hard to use in the next step.

Bad:  Classify this email.
Good: Classify this email into exactly one of these categories: billing,
      technical-support, feature-request, or general. Reply with only
      the category name, nothing else.

Reference the incoming data using Zapier's variables. In the prompt field, click the curly brace button or type to insert a variable from the previous step. The email body becomes something like {{1. Body Plain}} in your prompt.

Summarize the following customer feedback in exactly two sentences.
Focus on the main issue they are describing and their emotional tone.

Feedback: {{1. Message Body}}

Tell it what to do if the input is unclear. Automation runs without you watching, so the AI will occasionally receive messy or incomplete input. Give it a fallback instruction.

Extract the company name from this email signature.
If you cannot find a company name, reply with the word UNKNOWN.

Signature: {{1. Email Signature}}

Keep outputs short and consistent. A long, varied response is hard to use in the next step. Aim for outputs you could put in a spreadsheet cell or a database field without editing.

Three workflows worth building right now

Workflow 1: Email triage

Trigger: New email in Gmail (filter to a specific label or sender type if needed) AI step prompt:

Read the following email and classify it:
- Category: one of billing, technical, sales, or general
- Priority: high (needs reply today) or normal
- One-sentence summary of what the person needs

Reply in this exact format:
Category: [category]
Priority: [priority]
Summary: [summary]

Email: {{1. Body Plain}}

Action: Add a row to a Google Sheet with the category, priority, and summary

Workflow 2: Form submission enrichment

Trigger: New Typeform or Tally form submission AI step prompt:

The following is a response to our customer survey. Summarize it
in two sentences and rate the overall sentiment on a scale of 1 to 5
where 1 is very negative and 5 is very positive.

Respond in this format:
Summary: [two sentence summary]
Sentiment: [number]

Response: {{1. Answer}}

Action: Create a Notion database entry with the summary and sentiment score

Workflow 3: Meeting notes to action items

Trigger: New file added to a Google Drive folder (your meeting notes folder) AI step: Read the file content using the Google Drive "Get File Contents" step first, then pass it to AI:

The following are notes from a meeting. Extract all action items
as a numbered list. Each action item should name the person responsible
(if mentioned) and the task. If no person is mentioned, write "Unassigned".

Notes: {{2. File Content}}

Action: Send the action items list to a Slack channel or create tasks in your project management tool

Debugging when an AI step produces unexpected output

The most common problem is inconsistent output format. The AI produces "billing" in one run and "Billing Issue" in the next, which breaks the filtering logic downstream.

Fix this by being more explicit in your prompt:

Reply with ONLY one of these exact words: billing, technical, sales, general
Do not include punctuation, capitals, or explanation.

If the output is blank or the Zap is failing, test the step on its own using Zapier's "Test" button. Paste a sample input manually and see what comes back before relying on a live trigger.

A quick note on costs

AI steps in Zapier use your Zapier task quota, the same as any other action in a Zap. A two-step Zap with an AI step in the middle uses two tasks per run. Check your plan's task limit before building high-volume automations, especially for email-based triggers that might fire dozens of times per day.