Add an AI step.

So far your flow moves data from A to B. Useful, but a bit dumb. The thing that makes these automations genuinely clever is dropping an AI step into the middle, so the flow can read messy human text and do something smart with it. This is where the leverage really shows up, and it's also where a careless setup can embarrass you, so we'll do both halves: make it powerful, and keep it safe.

What an AI step is, in your flow

Make and Zapier both have built-in OpenAI and Claude steps. You drop one in like any other action, give it a connection (your OpenAI or Anthropic account and an API key), then map some text into it, usually the message field from the enquiry. It sends that text off to the model, gets an answer back, and hands that answer to the steps below it. From the flow's point of view it's just another step with an output you can map onward. The difference is what it's good at: turning unstructured, rambling, human writing into something tidy and usable.

The three jobs it does brilliantly here

• Summarise. Feed it a long, waffly enquiry and ask for a one-line summary. Now your CRM note and your phone screen show "Wants a quote for a bathroom reno in Geelong, budget around $20k" instead of six paragraphs.
• Classify. Ask it to read the message and reply with a single category: sales, support or other. That one word becomes the thing your flow routes on, which is the whole trick behind an auto-sorted inbox.
• Draft a reply. Ask it to write a friendly first-draft response in your tone. The flow doesn't send it; it parks the draft somewhere a person can glance at it, tweak it and hit go.

The prompt recipe that works

A vague prompt gives a vague, unreliable answer, and unreliable is no good when the next step depends on it. Three things make the output dependable enough to build on:

• Give it a role. Start with who it's being. "You sort incoming enquiries for a plumbing business."
• Give it the text. Map the enquiry message in clearly: "Here is the enquiry: [message]."
• Pin the format. Tell it exactly what to send back. "Reply with one word only: sales, support or other. Nothing else." When you need a category to route on, a fixed one-word output is what keeps the rest of the flow from breaking.

Run it on a few real-looking messages and check the answers are consistent. If it occasionally returns a sentence instead of the one word, tighten the instruction until it behaves. Five minutes of sharpening the prompt now saves a flow that misfires later.

The human check is not optional

Here's the line that keeps you out of the horror stories, and it's the same human-in-the-loop rule from the first course, now wired into the flow. AI is brilliant at the heavy lifting and it can still be confidently wrong. So decide, per step, whether the output is safe to use automatically or needs a person first:

• Safe to automate: a summary written into an internal CRM note, or a category used to route a message. If it's slightly off, the cost is small and a human sees it next anyway.
• Needs a human first: anything that reaches a customer, states a fact, quotes a price, or carries legal or money weight. The flow drafts it; a person approves it before it goes.

The neat part is you can build that check right into the automation. Have the flow draft the reply and drop it into a Slack message or an email to you with an "approve to send" step, rather than firing it off on its own. You keep the speed of AI doing the work and the safety of a person signing off the things that matter.

One more habit worth carrying over: keep genuinely sensitive data out of it where you can, and use your proper business AI account rather than a free consumer login. With the AI step understood and a human on the risky bits, you've got everything you need to build the real thing. Next lesson, we put the whole lead follow-up flow together, start to finish.