Definition
A design where a person reviews or approves what an AI system produces before it takes effect. The system prepares (drafts the message, assembles the recommendation, queues the change) and a human decides. It is the standard control wherever an error carries money, legal, clinical, or reputational consequences, which is most of what a small business actually cares about.
What it looks like in practice
A med spa's system drafts a reply to a client asking whether she can book a treatment four weeks after her last one. The draft is accurate on scheduling and correctly notes the required interval, but it stops short of answering whether her skin is ready. That is a clinical judgment. The message sits in the provider's queue with the client's history attached. The provider reads it in ninety seconds, adds one sentence, and approves. The client gets a fast, correct answer; nobody unqualified made a clinical call.
Why it matters
The useful version of this is specific rather than absolute. Reviewing everything defeats the point, if a person has to approve every appointment reminder, nothing has been taken off anyone's plate. Reviewing nothing is how a business ends up apologizing for something it never said. The workable line is drawn by consequence: routine, reversible, rule-covered actions run on their own, and anything touching price, scope, safety, or a sensitive conversation waits for a person. Where that line sits should be written down and yours to move.