Every AI vendor says "human-in-the-loop." For compliance-heavy teams, the phrase is either an architecture or a fig leaf, and the difference is auditable.
A real loop has three properties
Confidence routing: the system knows what it doesn't know, and low-confidence cases go to a person by design, not by luck. Meaningful review: the human sees the evidence behind the output, not just an approve button. Rubber-stamp UIs recreate the error rate you were escaping. And a durable trail: every decision records what the model did, what the human saw, and who signed off.
Where the human sits is a design decision
Before the action (approval gates) when errors are irreversible; after it (sampling and audit) when speed matters and mistakes are correctable; at the exceptions when volume is high and rules are crisp. Most regulated workflows want a mix: intake sampled, filings gated.
Done right, HITL isn't a compliance tax on automation. It's the reason your compliance officer says yes.
