Human-in-the-loop in production
Human-in-the-loop (HITL) is not a checkbox. It is a production design pattern for routing uncertain, risky, or high-impact model behavior to people who can decide, correct, approve, or escalate.
When to insert humans
- High-impact actions: money movement, legal advice, account changes, deletion.
- Low confidence or conflicting evidence.
- Safety/policy uncertainty.
- New intent clusters not covered by evals.
- User disputes, corrections, or appeals.
The review threshold should match blast radius. Low-stakes outputs can be logged; irreversible actions need approval.
Design the review packet
Show the human:
- User request and relevant context.
- Model answer or proposed action.
- Evidence, retrieved chunks, or tool results.
- Confidence/eval signals and policy flags.
- Clear approve/edit/reject/escalate controls.
Bad HITL UX creates rubber-stamping. The human needs enough signal to decide quickly.
Feedback capture
Every review decision should become data: label, correction, reason, segment, and version. That feeds feedback loops and future evals.
Pitfall
Putting a human at the end does not make the system safe if the human lacks context, time, authority, or a usable interface.
Connects to: agent HITL · AI product UX · feedback loops