AI inside automations: classify, extract, draft

Where AI actually fits in a workflow

Strip away the hype and an AI step inside an automation does one of three jobs: classify (which bucket does this email belong in?), extract (pull the name, date, and amount out of this messy text), or draft (write a first-pass reply a human will approve). That's it. Every impressive "AI agent" a client has seen on TikTok decomposes into those three verbs plus the plumbing you learned in lesson 3.

In Make, n8n, and Zapier this is literally one module: an OpenAI or Anthropic (Claude) step that receives text from the previous module, runs your prompt, and hands its answer to the next module. The plumbing is easy. The craft is the prompt.

Chat prompting vs production prompting

When you chat with ChatGPT, a rambling answer is fine — you're the reader. Inside an automation, the next module is the reader, and it needs the same shape of answer every single time. A reply that starts "Sure! Here's the classification you asked for:" breaks your filter that expects exactly the word refund.

A real classification prompt

Here's a prompt that triages a client's support inbox — the kind of build that gets quoted at real money. It shows every production habit at once: role, fixed labels, JSON-only output, an escape hatch, and one example.

Notice what makes it deterministic: a closed list of labels (the model can't invent "misc-question"), a forced JSON shape the next module can parse, a numeric rule instead of "use your judgment," and a defined behavior for uncertainty. Set temperature to 0 (or as low as the module allows) so the model stops being creative — creativity is a bug here.

Guardrails: never trust, always verify

Even a good prompt fails sometimes — models occasionally return broken JSON or an off-list label. Production builds assume that and catch it.

Human-in-the-loop is a feature, not a weakness

The pattern clients trust most: AI drafts, human approves. The classifier files the email AND a draft reply lands in a "pending" folder; a human clicks send. For anything customer-facing, money-touching, or reputation-risking, sell this pattern deliberately.

What the AI step costs

API calls are metered per token (roughly, per word in and out). A short classification like the one above costs a fraction of a US cent per email; even a few thousand emails a month is usually a rounding error next to the hours saved. Two habits keep it that way: keep prompts tight (don't paste a whole thread when the last message will do), and put a cheap non-AI filter first — no point paying the model to classify obvious newsletter spam a keyword filter could catch.

Do this now

Twenty minutes. Copy the triage prompt above into ChatGPT or Claude and adapt it to a business you know — change the categories and the urgency rule. Then test it against five realistic emails you write yourself: one obvious booking, one angry complaint, one ambiguous message, one spam, one written in Taglish.

Log results in a doc titled Prompt Test — email triage: the prompt, each input, each output, pass or fail, and one sentence on what you tightened after failures. That doc is a portfolio piece — it proves you test AI like an engineer instead of trusting it like a fan.