Also known as: HITL
In plain English
Human-in-the-loop means a person stays involved. The AI might draft a reply or suggest a decision, but a human checks it before it goes out.
In practice
Put humans where errors are costly or irreversible: payments, customer commitments, legal or medical content, and access changes. Use the model's confidence to route only uncertain cases to people, so review stays manageable and doesn't become rubber-stamping.
Under the hood
HITL patterns include pre-action approval, after-the-fact review, exception handling for low-confidence cases, and human feedback used to improve the model. Effective designs track override rates and review times, and guard against automation bias, where reviewers over-trust the AI.
Example
"Refunds over $500 need human-in-the-loop approval."