What is human-in-the-loop?

Human-in-the-loop is a system design in which a person must take part in each decision before it takes effect: approving, editing or rejecting what the system proposes. The system prepares; the person decides. It is the strongest form of human oversight, and the slowest, so it belongs on decisions where a person's judgement and accountability are worth that cost.

Why does human-in-the-loop matter once agents act?

When an agent can act on its own, human-in-the-loop is the main way to keep a person in charge of decisions that matter. Payments above a limit, closures of accounts, decisions affecting someone's credit or benefits: these are the actions where a business will want to say a person decided.

But a person in the loop is only a control if they can and do say no. A reviewer approving hundreds of agent drafts an hour, without the evidence to judge them or the authority to reject them, is in the loop in name only. That is the argument of human in the loop is not a control unless the person can say no. The wider design is set out in agent governance is a bound, not a model.

How does human-in-the-loop work in practice?

For agents, human-in-the-loop is put into practice through a human gate: the agent drafts an action, the verdict is held, and a named person approves, edits or rejects it. For it to work:

• The trigger is set by rule, by capability and by the customer's limits, not left to the agent.
• The approver is named and has authority for the decision.
• The evidence is shown: sources, the rule applied and the agent's draft.
• Rejecting is as easy as approving.
• The decision is recorded with the person, the time and any edits.

Whether every agent needs this is a separate question, discussed in should every AI agent have a human in the loop?

What does that look like in an example?

A local authority uses an agent to process applications for council tax reductions. The agent gathers the applicant's evidence, checks it against the scheme's criteria and drafts a decision. Every decision is held for a benefits officer, because each one affects a resident's bill and may be appealed. The officer sees the evidence, the criteria and the agent's reasoning, and approves, changes or rejects the draft. Over time, the authority can see which decisions officers change most, and improve the agent's brief there.

How is human-in-the-loop different from human-on-the-loop?

In human-in-the-loop, a person takes part in every decision before it takes effect. In human-on-the-loop, the system acts on its own, and a person supervises: watching, sampling and able to step in or stop it, but not approving each action. For agents, held actions are in-the-loop; allowed actions within bounds are on-the-loop. Most agents use both, depending on the capability.

Human-in-the-loop is also narrower than human oversight in law. Under the EU AI Act, Article 14 requires that high-risk systems be designed so that people can oversee them effectively, which can include, but is not limited to, approving each decision. The difference is explored in human-in-the-loop vs human oversight under the EU AI Act.