GLOSSARY
Human-in-the-Loop
An AI system design where a human reviews, approves, or overrides agent decisions at defined points. The default control mechanism for high-stakes trade decisions where full autonomy isn't yet defensible.
Human-in-the-loop (HITL) describes any AI workflow that routes decisions to a human reviewer before they take effect. The interaction can be confirmatory (the agent recommends, the operator approves), corrective (the operator labels the agent's mistake to retrain), or escalation-based (the agent only stops to ask when its own confidence drops below a threshold).
Why it matters
For high-stakes trade tasks — clearing a sanctions match, signing off on an HS classification on dual-use goods, releasing a wire above an AML threshold — HITL is what makes the agent legally defensible under EU AI Act Article 14, OCC SR 11-7 model-risk expectations, and equivalent supervisory regimes. The right question isn't whether to keep a human in the loop, but which decisions require it and what context you give the reviewer.
Related terms
- Human Oversight
- Explainability
- Model Risk Management
- Escalation