Add an AI by Zapier action step after your trigger, choose a model tier, write the prompt with data inserted using the "/" key, define named output fields, and test it on real examples before switching on. It needs a Professional plan or above. The Standard tier uses one task per run; Advanced and Premium use three and five.
Most failures come from treating the AI's reply as free text. Ask for named output fields with a fixed list of allowed values, then put a Filter or Paths step between the AI and anything a customer sees or a record stores. That one habit is the difference between a clever demo and a Zap you can leave running for a year.
What AI steps do well inside a Zap, and what they don't
An AI step reads text from an earlier step and returns text, or structured fields, to later steps. Three jobs fit that shape well:
- Classify: read an email, form or ticket and pick one label from a list you define, so later steps can route it.
- Summarise: condense a long thread, call transcript or form answer into a fixed format for your CRM, a Slack channel or a spreadsheet.
- Draft: write a reply or a note that a person reviews before it goes anywhere.
What AI steps do badly: arithmetic across many records, looking things up in your systems (use a search step in the app instead), and anything where a wrong answer goes straight to a customer. Each run also starts fresh. Zapier's help pages are explicit that AI by Zapier doesn't learn from previous runs, so every rule and example it needs must be in the prompt or in a knowledge source you attach. If the task is really a fixed rule ("if the subject contains INVOICE"), a plain Filter step is cheaper and more reliable; AI versus rule-based automation covers where the line sits.
Work out the task cost before you build
Zapier bills per task, meaning each successful action step. Triggers, Filter, Paths and Formatter steps don't count. AI steps count as one, three or five tasks depending on tier. A quick sum stops unpleasant surprises.
Take an illustrative software reseller that sells and renews business software licences. Its sales inbox gets about 900 emails a month. A first design:
- Trigger on each new email: 0 tasks.
- AI step to classify it (Standard tier): 1 task.
- Gmail step to add the matching label: 1 task.
- For the roughly 300 renewal enquiries: an AI summary (1 task), a Google Sheets row (1 task) and a draft reply (AI step 1 task, Gmail draft 1 task).
That's 900 x 2 = 1,800 tasks, plus 300 x 4 = 1,200, so about 3,000 tasks a month. The entry Professional tier (from $29.99 a month billed monthly, or $19.99 billed annually) covers 750 tasks, so this Zap needs a higher tier; check the task slider on Zapier's pricing page for the exact figure. Two design changes cut the bill sharply. A Filter step before the AI that drops newsletters, delivery receipts and out-of-office replies removed about 280 emails, free of charge. And staying on the Standard tier matters: the same classification on Premium would cost 900 x 5 = 4,500 tasks on its own. If the numbers keep climbing, finding the point where Zapier gets too expensive helps you decide whether to redesign or move platform.
Adding the AI step, click by click
- Set up the trigger. For email, use Gmail's new-email trigger with a search string such as
to:sales@yourdomain.com -category:promotionsso only the right messages start the Zap. Test it to pull in a real sample. - Add a Filter. Drop anything you never want the AI to see: auto-replies, newsletters, your own outgoing mail.
- Click the + icon and choose AI by Zapier. The step opens with a Configure panel and a Preview panel.
- Pick a starting template. Zapier offers templates for summarising, writing, classifying and extracting. They're a starting point; you'll replace most of the text.
- Choose the model tier. Standard (1x tasks, no tools), Advanced (3x), Premium (5x), or Bring Your Own Key (1x task, with your provider billing you for tokens). Supported providers include OpenAI, Anthropic, Google Gemini, Azure OpenAI and Amazon Bedrock.
- Write the prompt. Type "/" wherever you want to insert data from earlier steps, such as the email's subject and plain-text body.
- Define output fields. Under Settings, open Output Fields and add each field you want back, or use "Generate from prompt" to have Zapier suggest them from your instructions. Named fields are what later steps map from.
- Attach knowledge if needed. You can connect up to 20 sources, such as a Google Doc, Notion page, SharePoint file or an uploaded file (up to 100 MB). Only add what the step genuinely needs.
- Test and read the preview. Then retest with three or four different real examples before adding the next step.
Classify: route every enquiry with a fixed label list
Classification is the best first AI step because it's cheap, easy to check and immediately useful. The prompt must name every allowed label and say what to do when none fits.
Classify this email to our sales inbox. We are a software reseller.
Allowed categories (use EXACTLY one of these words):
renewal - an existing customer asking about renewing licences
new_quote - a request for pricing on software they don't have yet
licence_issue - activation, keys, seats or access problems
invoice - billing, payment or invoice questions
supplier - a message from a vendor or distributor, not a customer
other - anything else
Return:
category: one of the words above
confidence: high, medium or low
reason: one sentence quoting the words that decided it
If the email fits two categories, pick the one that needs action
first. If you are unsure, use "other" with confidence "low".
Subject: [Subject]
Body: [Body Plain]An illustrative preview for a real-looking email ("Hi, our 25 seats run out on the 30th, can you send prices for another year, and do you do the version with the backup add-on?") came back as:
category: renewal
confidence: high
reason: "our 25 seats run out on the 30th, can you send prices for
another year"Next, add a Paths step with one path per category, each using the condition "category (Text) Exactly matches renewal" and so on, plus a final path for low confidence that applies a Needs human label. Paths and Filter steps use no tasks, so routing is free once the AI has done its job.
The failure to expect in the first week is label drift. Early on, the reseller's AI step occasionally returned "Renewal" or "renewal enquiry" instead of the exact word, and those emails fell through every path unlabelled. Two fixes solved it: the "use EXACTLY one of these words" line in the prompt, and a free Formatter step that lowercases and trims the category before Paths reads it. The second failure was subtler. A distributor's price-increase notice ("Renewal pricing changes from 1 November") was classified as a customer renewal. Adding the supplier category with a clear description fixed it.
Summarise: turn long threads into a fixed format
Summaries are useful when a person needs to act on a thread without reading all of it: a renewal that's bounced between three people, a support conversation, a call transcript. The trick is a fixed format, so every summary can be scanned in the same order and mapped into the same spreadsheet columns.
Summarise this email thread for our renewals tracker.
Return these fields:
customer: company name as written
product: product and edition mentioned
seats: number of seats ONLY if stated explicitly, else "not stated"
expiry: the expiry or renewal date EXACTLY as written, don't convert
relative dates like "end of next month"
ask: what the customer wants, in one sentence
blockers: anything stopping the renewal, or "none"
next_step: the single action we need to take
Use only what's in the thread. Don't infer numbers or dates.
Thread: [Body Plain]An illustrative output from a nine-message thread:
customer: Harbourside Dental Group
product: Endpoint security, business edition
seats: 25
expiry: "the 30th"
ask: Price for a 12-month renewal, with and without the backup add-on.
blockers: Their finance lead needs the quote before Thursday's meeting.
next_step: Send both quotes by Wednesday.That one is usable as it stands. Two early problems are worth knowing about. Before the "don't convert relative dates" instruction, the step turned "end of next month" into a specific calendar date, and got the month wrong because it had no reliable idea of today's date. And before the "only if stated explicitly" rule, it read "we're a team of 12" in an email signature as a seat count. Both errors looked perfectly plausible in the tracker, which is what made them dangerous. Map the fields into a Google Sheets row or your CRM note, and spot-check five rows a week against the original threads.
Draft: write replies that wait in Gmail, never send
Drafting is where AI steps save the most time and carry the most risk. The safe pattern is to use Gmail's Create Draft Reply action, so the reply lands in the thread as a draft and a person sends it. Put the rules in the prompt, including what the draft must never do.
Draft a reply to this renewal enquiry from our sales team.
Tone: friendly, brief, British English, no exclamation marks.
Structure: thank them, confirm what they asked for in one line,
say when they'll get the quote, ask any question needed to quote.
Never: state a price, offer a discount, promise a delivery date for
licences, or say anything is attached.
If the summary shows a blocker or deadline, acknowledge it.
Sign off as "The renewals team".
Summary: [fields from the summary step]
Original email: [Body Plain]An illustrative draft:
Thanks for getting in touch about your endpoint security renewal. We'll prepare two quotes for 25 seats over 12 months, one with the backup add-on and one without, and send them by Wednesday so you have them before Thursday's meeting. Could you confirm whether you'd like to keep the same licence administrator on the account? The renewals team
What was fixed during testing: the first version of the prompt, without the "never" list, produced "I've attached our latest pricing" and once offered "our usual 10% loyalty discount", which the reseller doesn't have. Neither appeared once the rules were explicit. Even so, drafts stay drafts. If you later want some replies sent automatically, put an approval step in front of the send; Zapier's Human in the Loop tool can pause the Zap until someone approves, though on the Professional plan requests can only go to yourself. Adding human approval steps to AI automations covers the setup and timeouts.
Giving the draft step your product facts
A draft that only sees the customer's email can be polite but can't be specific. The knowledge-source option lets the AI step read a document you maintain, which is how the reseller's drafts learned things like "the backup add-on is only available on the business edition" and "quotes for more than 100 seats go to the account manager". The reseller keeps this in a two-page Google Doc called Renewals facts, attached to the draft step only.
Three rules keep a knowledge source safe. Keep prices out of it, because prices change weekly and a stale figure in a draft is worse than none. Give the document an owner and a review date written at the top. And keep it short: the model follows a tight two-page document far more faithfully than a 40-page product catalogue. The reseller learned the second rule the hard way. Its facts document still listed a backup product the vendor had withdrawn two months earlier, and three drafts offered it to customers before a salesperson spotted the problem during the read-through. The fix took one minute; the lesson was to review the document whenever a vendor sends a product notice.
Choosing the model tier, and when to bring your own key
| Job | Start with | Move up when |
|---|---|---|
| Classify short emails or forms | Standard | Rarely needed; improve the label descriptions first |
| Summarise long threads | Standard | Spot-checks show missed details in long threads |
| Draft replies with rules | Standard or Advanced | Drafts ignore rules or misread the situation |
| Anything needing tools or multi-step reasoning | Advanced | Premium only after testing shows Advanced falls short |
Bring Your Own Key keeps the step at one task per run, and your AI provider bills you separately per token. Note that a ChatGPT or Claude subscription doesn't cover this; API use is billed on the provider's developer platform. The token cost for this kind of work is small. At OpenAI's list price for its cheapest current model, gpt-6-luna ($0.10 per million input tokens, $0.50 per million output), 900 classifications of roughly 1,500 input and 60 output tokens each come to about $0.16 a month, if that model is offered for your key. The saving is in tasks, not tokens: moving a high-volume step from Advanced (3x) to your own key (1x) cuts two tasks per run. The trade-off is another account to manage and another bill to watch.
Zapier also has separate ChatGPT (OpenAI) and Claude (Anthropic) app integrations that call the model with your API key. They work, but AI by Zapier's output fields make structured results easier, so start there unless you need a specific model setting.
Test with twenty real emails before switching on
A preview that looks right proves very little. Collect 20 recent, real examples covering every category, including awkward ones, and run each through the step. Record what came back against what a person would have decided. The reseller's first test run, illustratively:
| Result | Count | Action |
|---|---|---|
| Correct category, high confidence | 16 | None |
| Correct category, low confidence (sent to Needs human) | 2 | Acceptable |
| Wrong category | 2 | Supplier notice read as renewal; two-topic email |
After adding the supplier category and the "action needed first" rule, a second run of 20 fresh examples scored 19 correct with one sensible low-confidence flag. Set yourself a bar before you test (for example, at least 18 of 20 correct and no wrong answer with high confidence) and don't switch on until it's met.
The reseller's Zap after one month
Pulling the pieces together, the illustrative reseller's live Zap ran like this in its first month:
- About 900 emails arrived; the Filter dropped about 280 before any AI step.
- About 620 were classified on the Standard tier; 31 (5%) landed in Needs human.
- About 300 renewals got a summary row in the tracker and a draft reply in the thread.
- Monthly usage came to about 2,400 tasks, within the tier it chose.
- The sales administrator's sorting time fell from about 45 minutes a day to about 10, spent clearing the Needs human label and sending drafts after a read-through.
- A weekly check of 20 random emails found two misclassifications in the month, both low-stakes.
The biggest gain wasn't the minutes. Renewal replies went out the same day instead of whenever someone got to the inbox, which matters in a business where a missed expiry date means a lost customer. For the wider pattern of sorting a shared inbox across sales, support and invoices, see AI triage for shared inboxes.
Keeping an AI step honest after launch
- Sample weekly. Pull 20 runs from Zap history and check them against a human judgement. Ten minutes a week catches drift early.
- Watch the fallback rate. If Needs human or "other" climbs from 5% to 15%, something has changed: a new kind of enquiry, a supplier's new email format, or a model update behind the tier.
- Version your prompt. Keep the current prompt, its date and its test score in the Zap's description or a shared document. When results change, you'll know what changed with them.
- Make errors visible. An AI step that errors, times out or returns an empty field can stall a Zap without anyone noticing. Set up error notifications and a fallback, as described in stopping automations breaking silently.
Once one AI step is stable, the next is easier: the reseller's second Zap reused the same classification prompt for its support inbox, with different labels. If you're weighing whether Zapier is the right home for a growing set of AI workflows at all, comparing Zapier, Make and n8n for AI automation sets out the trade-offs.
Further reads
- How to Build Your First AI Automation in Make, Step by Step — The same classify-and-route idea built in Make instead.
- Power Automate for Small Businesses: When It Beats Zapier — Check whether Power Automate would cost you less.
- AI Ticket Triage: Tag, Route, and Prioritise Support Requests — Classification rules for support queues specifically.
- How to Draft Customer Email Replies With AI That Sound Like You — Make AI-drafted replies sound like your team.
- Automation Audit: Find the Zaps and Scenarios Nobody Owns — Keep track of who owns each Zap as the list grows.
- How to Tell If a Process Is Ready to Automate With AI — Test a task before you spend tasks on it.
- AI Quick Wins: 12 Things a Small Business Can Set Up This Week — Twelve small AI jobs sorted by setup time, each with a copyable prompt, a real business example and the catch to watch for.
- AI Use Cases by Department for Small Businesses (With Examples) — 21 practical AI use cases across seven departments, each shown in a small insurance brokerage, with how to start and what to watch.
- Generative AI vs Traditional AI: Which Does Each Task Need? — How generative and traditional AI differ in what they need, cost and get wrong, four questions that sort any task, and a print shop's six tasks sorted.
- How to Learn AI as a Business Owner: A 30-Day Self-Study Plan — A day-by-day 30-day plan for owners to learn AI on real work, from first prompts to one working automation, with free courses and a day-30 self-test.
- What Is a Webhook? Why Some Automations Run Instantly — Why some automations fire in seconds and others lag by minutes, how to tell which you have, and how to keep webhooks from failing silently.
- How Barbers Can Win Back Lapsed Clients With Automated Messages — A lapse rule based on each client's own visit rhythm, three short texts with timings, and how to handle chair renters and barbers who've left.
- How to Chase Missing Client Records With AI Before Deadlines — Build a missing-items list, send AI-written chasers on a ladder counted back from each deadline, and let AI sort the replies for you.
- Replying to Trade Enquiries in 60 Seconds With AI — Why most automations can't reply in 60 seconds, the three builds that can, and what a first reply to a roofing or plumbing enquiry should actually say.
- The First 30 Days of AI in a Trades Business — A 30-day plan for a small trades firm: one assistant, three jobs, a business brief, a scorecard and the rules to set before spending more.
- How to Automate Purchase Orders and Supplier Emails With AI — Turn rough purchase requests into approved POs, send them automatically, chase silent suppliers and let AI sort replies, without auto-accepting price changes.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Zapier help articles on AI by Zapier (analyse and return data), Human in the Loop, Paths and Gmail; Zapier pricing page; OpenAI API pricing page. Checked September 2026.