"Human in the loop" is often said and rarely designed. In regulated industries the loop is the product: it decides what the agent is allowed to do, what the regulator sees, and whether your people trust the system after month three.
Three patterns
Approve. The agent prepares; a person decides. Every output waits for a human action before it has effect. Right for high-risk decisions under the AI Act, for anything irreversible, and for the first weeks of any deployment. Wrong as a permanent design for high-volume work, because approval fatigue turns the human into a rubber stamp within days.
Review. The agent acts; a person reviews a sample or a risk-weighted subset afterwards, with the power to reverse. Right for medium-risk, reversible actions at volume: routing, drafting, first-line resolution. The review rate is a dial. We start at 100%, drop it per category as the evaluation data justifies, and never below a floor agreed with compliance.
Observe. The agent acts; people watch dashboards and sampled traces. Right only for low-risk, fully reversible actions with strong evaluation coverage. Even here, every action is logged and attributable, and a kill switch is one click away.
How to choose
| Question | Approve | Review | Observe |
|---|---|---|---|
| Is the decision high-risk under Annex III? | Yes | No | No |
| Is the action reversible within a day? | Either | Yes | Yes |
| Evaluation accuracy on this category | Any | Above threshold | Well above, stable for weeks |
| Volume per person per day | Under 50 | Hundreds | Thousands |
Keeping the human real
- Show the reasoning, not just the answer. An approver who sees only "pay €1,240" will click yes. One who sees the coverage check, the fraud flags and the comparable claims makes a decision.
- Route by uncertainty. Send the agent's least confident cases to people first. Approvers who see hard cases stay engaged; approvers who see easy ones stop reading.
- Make override cheap and visible. One click to correct, and the correction feeds the golden set. People trust systems they can teach.
- Measure the human too. Approval time, override rate, agreement with later outcomes. If approvals take four seconds, you no longer have a loop.
The loop is not a safety feature bolted onto an agent. It is the operating model, and the agent is one participant in it.
What the regulator will ask
Who had oversight, were they competent and authorised, could they intervene, and did they? Article 26 of the AI Act says it plainly. The patterns above exist so that the answer is a record, not a recollection.