Human-in-the-loop — a system design in which a person reviews, corrects or approves what an AI produces before it takes effect.
Where to put the person
Not everywhere — that defeats the point. The useful rule is to put a person in front of anything that is hard to undo or that speaks for the business in public: a refund, a message to a client, a published review response, a price change. Drafting, sorting and summarizing can usually run unattended and be spot-checked.
Why it matters more than model choice
Every model is wrong some of the time, and the rate is not zero for any of them. A system with a review step turns a wrong output into a correction. A system without one turns it into an incident. For health, legal and financial businesses, the review step is often also what makes a use of AI defensible to a client or a regulator.
A review step people click through without reading is not a review step. Track how often reviewers actually change the output; if the answer is never, either the system is excellent or nobody is looking.
Related: AI agent, AI governance, hallucination, prompt injection.