The example business is Harbor Lane Bookkeeping, a made-up firm.
What are the most common mistakes when adding AI to an automation?
| # | Mistake | Fix |
|---|---|---|
| 1 | AI in every step | Use plain rules for anything that is the same every time |
| 2 | Keyword rules for messy words | Use one AI step to read the meaning and pick a label |
| 3 | A free-form AI answer | Give the AI a fixed list of labels to choose from |
| 4 | One AI step doing everything | Give the AI step one clear job |
| 5 | Sending AI replies straight to clients | Save a draft; a person clicks Send |
| 6 | Letting AI make money decisions | AI can sum up the history; a person makes the call |
| 7 | Forgetting the second bill | The AI model is billed separately from your automation tool |
Where does AI not belong?
1. AI in every step. Copying every website inquiry into the client tracker is the same every time, with no judgment. A plain automation does it and never gets it wrong. So does sending a reminder 3 days before each quarterly tax deadline: that is a date and a template, and a schedule trigger handles it. Adding AI to either one only adds cost and risk.
6. Letting AI decide refunds. In the lesson, "decide whether a client gets a refund" is the one job marked A person decides. It touches money and a client relationship. AI can summarize the client's history for whoever decides, but it shouldn't make the call.
Where does AI help, and how does it go wrong?
2. Keyword rules for messy words. Sorting incoming emails into new client, billing question or documents sounds like a rule. But people word things a hundred ways, so a keyword rule misses some. This is where an AI step earns its place: it reads the meaning and picks one label.
3. A free-form answer. If the AI step returns a paragraph, the next step can't act on it. In the lesson, the AI step must return one label from a fixed list (new client, billing or documents), so a rule can rely on it: IF the label is billing, go one way; otherwise go the other. In n8n that rule is an IF node, in Make a router, in Zapier a filter.
4. One AI step doing everything. The lesson's AI step has one job: read the email, return a label and a short draft reply. Small, focused AI steps are easier to check, and they keep the AI bill small.
Try it free, step by step
When to Add AI to an Automation (n8n, Make or Zapier). Sort six office jobs into plain rules, AI steps and human decisions, then see the safe shape of an AI automation and how n8n, Make and Zapier each add the AI step. You do every step yourself in a practice copy of Automation basics, it checks your work as you go, and “Do it for me” finishes any step you get stuck on. Free, and nothing touches your real accounts.
Start the free lessonWhat is the safe way to end an AI automation?
5. Sending straight to clients. While you are still testing an AI reply, the safest last action is to save it as a draft for a person to send. The lesson's flow ends with Priya reading and approving the draft, then the reply is saved as a Gmail draft in the client thread. Anything a client will see gets a human check until you trust it. That is the human in the loop pattern. The Guardian approval gate mistakes guide shows one way to add an approval step in n8n.
The full safe shape is: trigger, one small AI step, a rule that acts on its answer, a person who checks, then the action.
What do people forget about cost?
7. 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 that key per use, separately from any ChatGPT plan. Every unneeded AI step adds to it. Check the provider's own pricing before you switch an automation on.
Key takeaways
- Most of an automation should be plain rules.
- Use AI only for steps that read, write or judge messy words.
- Give the AI step one job and a fixed list of answers.
- Keep a person on money, promises and anything a client reads.
- Save drafts, not sends, until you trust the results.
More answers are in the when to add AI to an automation FAQ. For a broader checklist on what to automate at all, see when to automate a task with AI.
