Human-in-the-loop
Human-in-the-loop means a person sits at the exact point where the AI's output turns into a real action, and nothing proceeds until that person says yes. The AI does the work and proposes; the human approves, edits, or rejects. The hard part is placing the checkpoint where the stakes are real without making a human rubber-stamp everything, which just trains them to click yes.
Why it matters
Models are non-deterministic and confident even when wrong, so a fully automatic pipeline will eventually take an expensive action off a bad answer. A refund gets issued, an email goes to the wrong client, a database row gets deleted, and no one saw it coming. A human gate is your circuit breaker for the small slice of decisions where a mistake actually hurts. It also gives you a stream of corrections you can log and learn from.
How it works
You draw a line between what the AI can do on its own and what needs sign-off, usually by risk: low-stakes reads run free, anything that spends money or touches a customer waits for a click. The system pauses, shows the human the proposed action plus the reasoning and sources, and records the approve or reject decision. Watch the approval rate and how long reviewers actually spend, because a 99% approve rate at two seconds each means the human is asleep and you have no real gate.
A support bot drafts a $200 refund and, instead of issuing it, drops it into an agent's queue with the customer's history and the policy it relied on. The agent glances at it, sees the order was clearly outside the return window, and rejects it in one click. The bot did the tedious work; the human caught the call that would have cost real money.