Why this matters
Once people see AI write an email, they want it in every step. That is how a cheap, reliable automation turns into one that costs money on every run and sometimes makes things up. The opposite mistake is just as common: a team keeps sorting client emails by hand because a keyword rule keeps missing the ones that are worded oddly.
The skill is knowing which steps need AI and which do not. Plain rules handle anything that is the same every time, for free. AI earns its place only where a step has to read, write or judge messy words. And some decisions, like refunds, stay with a person.
This module gives you a quick test for any step, and the safe shape of an AI automation you can build in n8n, Make or Zapier.
What you’ll build
You sort six jobs for Harbor Lane Bookkeeping (made up) into plain rules, AI steps and human decisions. Then you walk through one AI automation from top to bottom: a trigger, one small AI step that picks a label and drafts a reply, a rule that acts on the label, a person who approves, and an action that saves a draft.
Last, a cheat sheet shows how n8n, Make and Zapier each add the AI step, and what the second bill is.
What you need
- NothingEvery screen is a plain diagram, not a real app. No account, nothing to install, nothing is sent.
- Your list of repeat jobs, if you have oneAt the end you mark each job as a rule, an AI step or a person’s call.
Concepts
3 ideas you’ll use in the practice.
Rule, AI step or a person?
Ask one question about each step: is it the same every time? If yes, use a plain rule. If it has to read, write or judge words that change every time, like sorting an email by what it says, an AI step helps. If it involves money, a promise to a client or a judgment call you would not hand to a new hire, a person decides.
The safe shape: trigger, AI, rule, person, action
Give the AI step one job and a fixed list of answers, such as "new client, billing or documents", so the next step can rely on what comes back. Let a rule (an IF node, a Make router or a Zapier filter) act on that answer. Keep a human in the loop for anything a client will read, and save drafts instead of sending until you trust the results.
Remember the second bill
The AI model is paid for separately from your automation tool. With the OpenAI app in Make or Zapier, or an OpenAI node in n8n, you paste an API key and OpenAI bills per use, separately from any ChatGPT plan. Small, focused AI steps keep that bill small. For a side-by-side on cost and AI features, see n8n vs Make vs Zapier for AI automation.
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.
Do it for real
Now repeat it in your own account. Tick each step as you go.
Quick check
3 questions to check you’ve got it.
1. Which job should stay a plain rule, with no AI?
- Sorting emails by what they say
- Sending a reminder 3 days before a deadline
- Drafting a personal reply
Show the answer
Sending a reminder 3 days before a deadline. A date and a template are the same every time. AI would only add cost and risk.
2. Why give the AI step a fixed list of labels to choose from?
- So the rule after it can rely on the answer
- Because AI cannot write sentences
- To make it run faster
Show the answer
So the rule after it can rely on the answer. A rule can route "billing" reliably. It cannot route a free-form paragraph.
3. What is the safest last action while you are still testing an AI reply?
- Send it straight to the client
- Save it as a draft for a person to send
- Post it in a public channel
Show the answer
Save it as a draft for a person to send. A draft lets a person read every AI reply before it goes out.
Zapier: ChatGPT Drafts Every Reply, Signatoro Signs It
You know where AI belongs. Last, build it: a Zap where ChatGPT drafts every reply, Signatoro signs it and a person clicks Send.
Start module 7 →