Human-in-the-Loop (HITL)
A system design where a person reviews, approves, or corrects an AI system's decisions at defined points, combining model speed with human judgement for high-stakes actions.
Human-in-the-loop (HITL) places a person inside an AI workflow rather than outside it. The model drafts, classifies, or recommends; a human confirms before the result takes effect. It is the standard pattern wherever errors are expensive or irreversible: medical triage, financial decisions, content moderation, autonomous agents that spend money or send messages.
Common patterns:
- Approval gates: the agent pauses for confirmation before sensitive tool calls.
- Review queues: low-confidence predictions are escalated to a person.
- Active learning: human corrections are fed back to improve the model.
- Escalation paths: the system hands off to a human when inputs fall outside its competence.
HITL is also how models learn human preferences: RLHF and related methods use human rankings as training signal. The design tension is throughput — more human review means fewer errors but higher cost and latency. Mature systems tune the threshold so that humans concentrate on genuinely ambiguous cases, and so that no single AI mistake can cause lasting harm.