Human-in-the-loop: approvals and control

When you finish, you can design a workflow where a person approves the steps that matter.

Why keep a human in the loop

An agent can be wrong and still sound sure. If it also has the power to act, a mistake can spread before anyone sees it. Human-in-the-loop means a person reviews or approves certain steps before they happen. It is the main safety tool for beginners.

It also keeps responsibility clear. You are accountable for what your agent does in your name. If it emails a client, signs up for a service or edits a record, those actions belong to you. Approvals keep you in charge.

Which steps need approval

Ask three questions of each step. Can it be undone? Does it reach other people? Does it cost money or touch private data? If the answer to any is yes, require approval. Reading public information might run automatically. Sending a message, deleting, paying or sharing a file should wait for you.

A weak setup is “The agent will email suppliers.” A strong setup is “The agent will draft emails to suppliers and place them in a folder. I will read and send each one.” You still save the writing time, and you keep the final act.

Make approvals meaningful

Approvals only work if you actually look. If a tool asks you to click approve fifty times in a row, you will start clicking without reading. Design for this: group similar items into one review, show the agent’s reasoning and sources, and keep the number of approvals small by limiting the scope.

Ask the agent to show what it plans to do, in plain words, before it acts. “List the three actions you will take and why. Wait for my yes.” Also ask for a short report after: what it did, what it skipped and what it was unsure of.

Worked example: scheduling interviews

A hiring coordinator in Bangalore uses an agent to schedule interviews. The agent reads availability, proposes time slots and drafts invitations. The coordinator reviews each batch and sends them. The agent does not decide who is invited; that decision stays with people.

Once, the agent proposed a slot during a public holiday. The coordinator caught it at the approval step. Without that review, a candidate would have received a confusing invitation. The approval cost a minute and saved an awkward correction.

Practice now

Take a workflow you might automate. For each step, answer: can it be undone, does it reach others, does it cost money or touch private data? Mark approval points.
Write an instruction telling an agent to list its planned actions and wait for your yes before acting, and a request for an end-of-run report.

Use any AI assistant you have. Do not paste private, client or patient information into it.

Check yourself

1. Which step most needs approval?

2. Why can too many approval requests weaken safety?

3. Who is accountable for actions an agent takes in your name?

Ask the AI tutor

Ask anything about this lesson, in any language. The tutor is an AI: it can be wrong, and it never needs your personal details.

Next lesson: Where agents fail

AI-drafted. Checked by an automated validator and a separate AI reviewer; a person has not read every lesson. Translations and tutor answers are AI-made. Found a mistake? Tell us. AI can be wrong. General information only, not professional advice. Not affiliated with any AI company named. Current University is not accredited.