Putting AI inside the workflow

Automations that think in the middle

Everything so far moves data without judging it. The Level 6 move is inserting one thinking step into the pipeline: trigger → AI step → action. The form entry doesn't just land in a sheet — it arrives classified. The week's numbers don't just sit there — they arrive summarized. The lead doesn't get a template — she gets a personalized draft waiting for your approval.

Both Zapier and Make offer AI steps as ordinary modules — add a step, choose the AI app (ChatGPT, Claude, or the platform's built-in AI), write the instructions, map the incoming fields into your prompt. If you can map a field, you can add a brain. Model choice: keep it boring. The default model each platform offers is fine for classify-summarize-draft work; you're not writing a novel in step 3 of a Zap.

Prompting a robot inside a robot

Here's what nobody tells you: prompting inside an automation is the opposite of chatting. In chat, you want conversation — follow-ups, nuance, personality. Inside an automation, the AI's output feeds a machine step that cannot ask clarifying questions. So the prompt must force a predictable shape every single run:

• One job per step. Don't ask one AI step to classify AND summarize AND draft. Three jobs, three steps — each testable alone.
• Lock the output format. "Return ONLY one word from this list," "Return exactly 3 bullets," "Return subject line, then body, nothing else." Say only and mean it — one stray "Sure, here's your summary!" breaks a downstream mapping.
• Give it the edge cases. Tell it what to do when the input is empty, spam, or gibberish ("if the message is empty or spam, return: SKIP").

Three battle-tested shapes — copy, then adapt the brackets:

You are a classifier. Read the form message below and return ONLY one word from this list, nothing else: SALES, SUPPORT, PARTNERSHIP, SPAM.
If the message is empty or nonsense, return: SPAM.

Message: {mapped form message field}

Summarize the text below in exactly 3 bullet points, each under 20 words, plain language, no introduction and no sign-off. Focus on: what happened, what changed, what needs a decision.

Text: {mapped content field}

Write a short, warm reply email (under 120 words) to the person below. Reference one specific detail from their message so it feels personal. Sound like a busy, kind founder — no corporate phrases, no exclamation marks in a row. Return ONLY: a subject line on line 1, then the body.

Their name: {name field}
Their message: {message field}
Context about my business: {one sentence about what you sell}

The guardrails (non-negotiable)

1. Draft, don't send. Anything outward-facing — replies, posts, DMs — the AI step writes it and the automation saves it as a draft or sends it to you. A human clicks send. You'll be tempted to skip this after week one goes perfectly. Don't. The one bizarre AI reply that reaches a real customer costs more than every minute this checkpoint ever takes.
2. Keep private data out of the pipe. Don't route health details, finances, or anyone's personal secrets through an AI step. Same rule you learned in Level 1, now with the extra weight that nobody's watching each run.
3. The intern test. Before switching any AI step to unsupervised: "Would I let a brand-new intern do this exact task with nobody checking?" Filing, labeling, summarizing for your own eyes — yes. Talking to customers, spending money, deleting things — no.