What it means
Human-in-the-loop places a person at specific decision points rather than reviewing everything or nothing. The useful version is targeted: the AI does the work, and a human approves the step that actually carries consequence — the payment, the email to a customer, the production change.
The design question is which points. Reviewing everything destroys the efficiency the automation was for; reviewing nothing means a wrong action ships. The workable rule is to gate on reversibility and blast radius: a draft nobody sends needs no gate, an irreversible external action always does.
The well-documented failure mode is rubber-stamping. When a system is right ninety-something percent of the time, reviewers stop reviewing and start clicking approve — and approval volume looks identical whether or not anyone read anything. Oversight that exists on the org chart but not in practice is worse than none, because it produces false assurance.
Why it matters
This is the main practical control on agentic AI, and it is increasingly a compliance requirement — the EU AI Act mandates meaningful human oversight for high-risk systems, with emphasis on *meaningful*. A rubber-stamp does not satisfy it.
In practice
Gate on irreversibility, not on confidence scores. Give reviewers the information needed to actually judge — sources, the reasoning, what changes if they approve — and sample your approvals to check whether review is really happening.