Why this matters
The AI step is the part of an automation nobody tests properly. It reads an email and decides where it goes, but if it answers in a sentence instead of the exact fields the next step expects, or makes up a policy number, the workflow breaks or quietly does the wrong thing.
ChatGPT can write that prompt for you if you give it the rules: the one job, the allowed answers, the exact JSON fields, what to do in a tie, and what it must never invent. Then you make it run the prompt on real examples and look for the weak spots before any customer email goes through it.
Practice
Do every step yourself in a practice copy of the real tool. It checks your work, and “Do it for me” finishes any step you get stuck on.
What you’ll build
You work at Calloway Insurance Agency (made up), where one person reads about 90 emails a day and forwards each to Claims, Quotes, Certificates, Billing or Other. You attach the desks and six sample emails and get a complete, paste-ready prompt that returns five JSON fields.
Then ChatGPT runs it on all six emails, shows the JSON for each, says which ones a person should still read, and rewrites two rule blocks where the test showed a weakness.
What you need
- Nothing to start the practiceThe lesson is a practice copy of ChatGPT. The agency, its clients and policy numbers are made up.
- A ChatGPT account, to write the prompt for realAny plan. The answers shown came from a Plus account.
- Five to ten real examplesCopies of the kind of emails, forms or messages the AI step will see, with names removed if your rules require it.
- Your automation tooln8n, Make or Zapier, with an AI step where the prompt goes. Test it there before going live.
Concepts
What a good AI-step prompt contains
One job. The exact list of allowed answers (here, five desk names). The output format, field by field, with types: urgent is true or false, policy_number is text or null. A rule for ties (claims first, mention the other issue in the summary). And what it must never do: invent a policy number, add Markdown, add extra fields.
The real prompt also ends with the email itself in three placeholders. Each tool maps its own fields into those.
Test it on examples, then read the misses
All six sample emails got a valid desk and no invented policy number. The interesting result is email 6: two inches of water in a basement since Sunday, a ruined carpet, and a billing question. It went to Claims with confidence "medium" and urgent false.
Would your agency want water damage flagged urgent? That is your call, not the model's. It is also why ChatGPT suggested sending anything below "high" confidence to a person.
Simulated is not tested
ChatGPT said it plainly: "These are simulated AI-step outputs, not results from a live automation." The model inside your tool, its settings and the way the email arrives can all change the answers.
Paste the prompt into your AI step, run the same examples through it there, and keep a branch that sends low-confidence results to a person.
Do it for real
Now repeat it in your own account. Tick each step as you go.
Downloads
- sample-inbox-emails.txtThe desks and six sample emails used in the lesson. Made up.
Quick check
3 questions to check you’ve got it.
1. Why ask for JSON with fixed fields instead of a sentence?
- JSON sounds more professional
- The next step in the automation can read fixed fields reliably
- ChatGPT answers faster in JSON
Show the answer
The next step in the automation can read fixed fields reliably. A router or a spreadsheet step needs the desk in the same field every time. A sentence has to be guessed at.
2. An email has no policy number. What should the prompt make the AI return?
- A likely number based on the sender
- null
- The word "unknown" in the desk field
Show the answer
null. Never invent a policy number. null tells the next step there was none.
3. ChatGPT ran your prompt on six emails and all six look right. Is it ready?
- Yes, it was tested
- Not yet: run the same emails through the AI step in your own tool first
- Only if it said confidence high for all six
Show the answer
Not yet: run the same emails through the AI step in your own tool first. ChatGPT simulated the step. Your tool's model and settings may answer differently.
ChatGPT: Read an Automation Error and Fix a Workflow You Didn't Build
The prompt is ready. Next, the day an automation stops working and you did not build it.
Start module 3 →