Human-in-the-loop (HITL) is a design in which certain decisions made by an AI system are routed to a person, who approves, edits or rejects them before they take effect. It is used for high-impact agent actions, uncertain classifications and evaluation grades where automated judges are unreliable.
Where it applies
- Action approval: an agent must get a confirmation before sending money, deleting data or emailing outside the organisation.
- Escalation from guardrails: borderline inputs or outputs go to a reviewer instead of being silently blocked.
- Evaluation: humans label a sample of outputs to calibrate an LLM judge and audit its errors.
- Training: people provide the preference data behind RLHF.
Worked example: sizing a review queue
Illustrative numbers: a support agent handles 50,000 conversations a day and proposes refunds in 4% of them. Policy sends refunds above a threshold to a person; that is 25% of refund proposals, or 500 reviews a day. At two minutes each, that is about 17 hours of reviewer time daily. If reviewers also check a 1% random sample of all conversations for quality, add another 500 items. Sizing the queue before launch shows whether the threshold is affordable or needs a cheaper intermediate check.
Failure modes
- Automation bias: reviewers approve almost everything because the system is usually right. Track approval rates and seed known-bad items to measure detection.
- Context starvation: the reviewer sees only "approve transfer?" and not why the agent proposed it. Show the relevant part of the trajectory.
- Fatigue: too many prompts train people to click through. Reserve approval for actions that are irreversible or high value.
Why it matters
Human review is the main control for decisions too risky to automate, and many regulations require meaningful human oversight for high-risk AI. It only works if it is designed: the right decisions routed, enough context shown, and reviewer accuracy itself measured. Done well, it also produces labelled data that improves the automated layers over time.