Human-in-the-loop

Human-in-the-loop means the AI proposes and a person decides. The model drafts a reply, suggests a category, prepares a refund, or extracts invoice fields, and nothing happens until someone accepts, edits, or rejects it. It's the default safety mechanism for AI features that touch customers, money, or records.

It comes in degrees. For low-risk output, review can be optional: a tag anyone can change with one click. For higher-stakes actions, writes become proposals that need explicit confirmation, often shown as a summary or diff of what will change. Many teams send only uncertain cases to review, using signals such as failed validation, missing citations, or a risk label, and let confident ones through. Automation platforms such as Zapier and n8n include approval steps for the same pattern.

The risk is rubber-stamping. When reviewers see hundreds of mostly correct suggestions a day, they stop reading, and the loop gives false comfort. Keep the review surface small and specific, highlight what actually needs checking, log every approval and edit, and track how often people change the output. Those edits are also your best source of eval and training data, so keep them.