AI automation

How to use ChatGPT in Zapier and Make without breaking your workflow

By Sauravjit Singh · Updated 2026-11-04 · 7 min read

Short answer: Add an AI step between your trigger and your next action: AI by Zapier or Make's built-in AI tools need no API key, while the OpenAI, Claude or Gemini modules use your own key. The key to reliability is a prompt that returns a fixed format, a short list of allowed answers and an 'unknown' fallback.

Putting ChatGPT inside an automation is easy. Keeping it reliable is the hard part. A chat window can give a slightly different answer each time and you simply read it; an automation feeds that answer straight into a sheet, a filter or a WhatsApp message, so one unexpected word can send a lead down the wrong path.

This guide is about prompts for machines, not people. For general prompting skills, the AI course for working professionals guide and our which AI tool to use for what explainer are better starting points.

What can an AI step actually do inside a workflow?

An AI step is good at four jobs: classifying, summarising, extracting and drafting. It is poor at anything that needs exact facts it was not given, such as prices or stock.

  • Classify: tag an enquiry as hot, warm or cold; route a support email to billing or delivery.
  • Summarise: turn a long form answer or call note into two lines for the sales sheet.
  • Extract: pull the city, budget and product from a free-text message.
  • Draft: write a first reply for a human to approve. See our guide to automating email replies with AI for that use.

AI by Zapier, Make's AI tools or your own API key?

Start with the built-in AI step. Move to your own API key when you need a specific model, longer inputs, structured JSON output, or tighter control over cost.

OptionNeeds an API key?Good forWatch out for
AI by Zapier (built-in step)NoQuick classification, summaries and extraction inside a ZapUses Zapier's own allowance and tasks; less control over the model
Make AI tools (built-in modules)NoCommon tasks such as categorise, summarise, translateCan use more credits than a standard module; check the pricing page
OpenAI (ChatGPT) app or moduleYes, from the OpenAI platformChoosing the model, JSON output, longer promptsBilled separately from a ChatGPT Plus subscription
Anthropic (Claude) or Google Gemini modulesYes, from each providerLong documents, an alternative model, matching tools your team already usesModule options differ by platform; check the current integration list

A point that confuses many people: a ChatGPT Plus or business workspace subscription does not include API usage. The API is a separate account with its own billing, as of November 2026.

The automation-safe prompt pattern

Write the AI step's prompt like a form the model has to fill, not a conversation. Five rules do most of the work.

  1. Give the role and the single task in one line: 'You classify sales enquiries for an interior design firm in Delhi.'
  2. List the only allowed answers: 'Category must be one of: kitchen, wardrobe, full_home, office, unknown.'
  3. Fix the output format: 'Reply with JSON only, with keys category, city, budget_inr, summary. No other text.'
  4. Add the fallback: 'If the message does not say, use unknown or null. Never guess.'
  5. Put the input last, clearly marked: 'Enquiry: {{message}}'.

Then, in the next step, use a filter or router on the category field, never on the summary text. Free text is for humans; fixed labels are for machines.

Worked example: classifying a lead enquiry

Say an interiors firm in Punjabi Bagh gets this message on its website form: 'Hi, we just bought a 3BHK in Dwarka, need modular kitchen and maybe wardrobes later, budget around 4-5 lakhs, call after 6.'

A loose prompt ('Read this lead and tell me what they want') returns something like: 'This customer is interested in a modular kitchen for their new 3BHK flat in Dwarka and may want wardrobes in the future. Their budget is ₹4–5 lakh.' Perfectly readable, and useless to the next step. A filter cannot reliably pick 'kitchen' out of that sentence, and next time the wording will differ.

The pattern above returns this instead:

  • {"category": "kitchen", "city": "Dwarka", "budget_inr": 450000, "summary": "3BHK, modular kitchen now, wardrobes later. Call after 6 PM."}

Critique: the category is one allowed label, so a router can send it to the kitchen designer. The budget is a single number (the model took the midpoint), which you can sort and filter in the sheet. One judgement call to check: you may prefer the model to return budget_min and budget_max rather than a midpoint. Decide the rule and write it into the prompt.

The 'call after 6' detail survived in the summary because a human needs it. That split, structured fields for the automation and a short summary for the person, is the habit that makes AI steps dependable.

Setting it up in Zapier and Make

  1. Build and test the automation without AI first: trigger, then sheet row.
  2. In Zapier, add an AI by Zapier or ChatGPT (OpenAI) action between them. In Make, add an AI tools module or the OpenAI module after the trigger.
  3. Paste your prompt and map the enquiry text into the input slot.
  4. If using your own key, choose a small, low-cost model first and switch on JSON or structured output if the module offers it.
  5. In Make, add a Parse JSON module after the AI step; in Zapier, use the step's structured output fields or a Formatter step.
  6. Map each field (category, city, budget) into its own sheet column.
  7. Run ten real past enquiries through it and check every output by hand before going live.

What does an AI step cost per 1,000 runs?

For short classification tasks on a small model, the API cost is usually tiny: roughly ₹10–50 per 1,000 runs as of November 2026. Larger, top-tier models can cost ten times that or more, and you also pay your Zapier or Make usage for the step.

The arithmetic: a prompt like the one above plus a short enquiry is roughly 400–600 tokens in and 80–120 tokens out (a token is a word fragment, about three-quarters of an English word). For 1,000 runs that is around 5 lakh input tokens and 1 lakh output tokens. Multiply by the per-million-token prices on the provider's pricing page and convert at the day's rate. Prices drop often, so recheck each quarter.

In practice the platform cost often outweighs the AI cost. An extra Zapier task per lead at volume can cost more than the model itself.

Keeping AI steps reliable after launch

Before launch, build a small test set: twenty real past inputs, including the awkward ones (a message in Hinglish, a one-word enquiry, a spam message, an enquiry that mentions two products). Write down the answer you expect for each, run them all, and count how many match. If fewer than about eighteen match, fix the prompt before you fix anything else. Keep the same twenty inputs and rerun them whenever you edit the prompt or switch models, so you can see whether a change helped or quietly broke something.

A common beginner mistake is to keep adding instructions to the prompt every time one output looks odd. Past a point, long prompts become contradictory. Prefer fewer, clearer rules plus the unknown fallback.

  • Add a filter that stops the run, or sends it to a 'check manually' sheet, when the category is unknown or the JSON fails to parse.
  • Log the raw AI output in a hidden column so you can see what went wrong later.
  • Do not let AI output go straight to a customer without either a human check or a fixed template around it.
  • Re-test when you change the model; a newer model can format answers differently.
  • Never pass sensitive personal data into the prompt that the task does not need.

Once your workflow has several AI decisions chained together, you are edging towards what people call an agent. Our explainer on AI agents vs automation workflows helps you decide whether that is what you need, and the no-code AI automation guide shows where AI steps fit in the bigger picture.

Prompting for structured output is covered with live exercises in Spettro School's Applied AI Professional course.

Frequently asked questions

Do I need a ChatGPT Plus subscription to use ChatGPT in Zapier?

No. The ChatGPT (OpenAI) integration uses the OpenAI API, which is billed separately from Plus. Built-in options such as AI by Zapier need no OpenAI account at all.

Why does my AI step give different answers each time?

Language models vary their wording unless you constrain them. Fix the output format, list the allowed answers, add an unknown fallback and use a low temperature setting if the module offers it.

Can I use Claude or Gemini instead of ChatGPT in Make?

Yes, Make and Zapier offer integrations for several AI providers. The same prompt pattern works; test a few real inputs because formatting habits differ between models.

How do I make ChatGPT return JSON in Make?

Ask for JSON only in the prompt, switch on the module's JSON or structured response option if available, and add a Parse JSON module after it. Route bad outputs to a manual check.

Is it safe to send customer data to ChatGPT through Zapier?

Send only what the task needs, use business accounts, and check each provider's data-use terms. Avoid passing ID numbers, bank details or health information into AI steps.

Want to learn this with a trainer?

Leave your details and our counsellor will call you within one working day with batch dates, fees and the right course for you.

  • Free counselling call, no obligation to join
  • Weekday and weekend batches, online or offline
  • Small batches of 15 to 20 students

Or chat on WhatsApp

Enter a 10-digit mobile number.

We use your details only to contact you about Spettro School courses.

Chat on WhatsApp