People decide

Throughout this course we keep returning to one principle: human in the loop. AI may read, summarise, draft, calculate and suggest. The decision stays with a person who can be held accountable. The question is, which decisions?

These decisions are always made by people:

• About people: accepting or rejecting applicants, assessing performance, imposing sanctions, deciding who gets help or a service.
• About money: approving payments, granting discounts, rejecting claims.
• About law and health: interpreting rules for a real case, deciding on medical action.
• That cannot be undone: public statements, official letters, deleting data.

In all of these, AI may help prepare material. But there is a subtle trap: once AI gives a neat, confident suggestion, we tend to approve it. This tendency is called automation bias, the habit of trusting a machine's suggestion over the evidence in front of us. It turns "human in the loop" into a human who only presses a button.

Real review has clear signs. You open the original data, not just the AI summary. You look for at least one reason against the suggestion. You can explain your decision to the person affected without the words "because the system said so". And you are genuinely ready to decide differently from the AI. If in a month you never once disagreed with the AI, that is not a sign the AI is perfect; it is a sign your review needs reviewing.

Decisions helped by AI need to be explainable, by you now and by a colleague six months from now. So record the AI's part briefly, in the same place as the decision record. Five lines are enough:

1. Decision: what was decided, by whom, when.
2. AI material: what the AI produced (summary, score, draft, suggestion) and which tool was used.
3. Source data: what data the AI read, and whether it was masked.
4. Human review: what you checked yourself and what you changed.
5. Reason: why the decision is what it is, in your own words.

This record is a simple audit log. It helps when there is a question, an objection, or an error to trace. People's right to object to automated decisions is covered in the next lesson.

Be careful with what the record itself holds. Do not copy passwords, tokens or unneeded personal data into a log; write a reference to where that data is kept. A log full of secrets only moves the problem somewhere new.

The same principle applies when the worker is not a chatbot but an AI agent that can use tools and run several steps by itself. An agent may gather material, draft, even prepare actions, but decisions about people, money, law or health still stop at a person's desk. Design the agent to wait for approval at that point, to show what it is about to do and on what data, and to log each step. If an agent cannot explain what it will do before doing it, it is not ready to be given authority over such decisions.