Human-in-the-Loop (HITL)
An AI architecture that keeps a person embedded at defined decision points to review, approve, or correct the model's output before it affects the customer.
Human-in-the-loop (HITL) is an architecture, not a metric. It keeps a person embedded at defined points in an AI workflow, with the authority to review, approve, correct, or reverse what the model does. It spans a spectrum: human-in-command, where nothing reaches the customer until a person approves it; human-on-the-loop, where the AI acts but a person monitors and can intervene; and the feedback loop, where every human correction becomes training signal that sharpens the next answer.
In support the pattern is usually risk-gated. The AI handles the routine, and anything expensive or irreversible — a large refund, a VIP account, an action the model itself is unsure about — is escalated to a human before it lands. Regulators increasingly treat some version of this as mandatory: the EU AI Act requires meaningful human oversight of high-risk AI systems, so for a growing set of use cases HITL is a compliance requirement, not just a quality one.
What it hides: a human in the loop is not the same as oversight. Pile too many approvals on too few reviewers and they rubber-stamp to clear the queue; hand people fluent, confident AI drafts and automation bias nudges them toward approving without really checking. Oversight is only real when the reviewer has the time, the context, and the standing to say no — and when someone actually measures how often they do. A loop that never rejects anything is decoration.