Human-in-the-loop

A system design where a person must approve, edit, or reject an AI action before it takes effect. The action does not proceed without explicit human sign-off.

Human-in-the-loop (HITL) is a specific control pattern: a defined action waits for a person to approve, edit, or reject it. An insurance workflow might pause an unusual claim before payment, or a migration tool might hold a destructive schema change until an operator confirms the plan. The checkpoint should expose the evidence and decision the reviewer needs; a button with no useful context supplies delay without reliable control.

Builder example

HITL is useful when a person has information, authority, or judgment that the automated checks lack and the action can still be stopped. It is one option among deterministic validation, limited permissions, sandboxing, reversible execution, monitoring, and specialist review. Requiring approval for every action can create queues and routine clicking, so place the checkpoint around a named uncertainty or consequence rather than around AI use in general.

An agent drafts a reply to a client using meeting notes and CRM history.

When policy assigns sending authority to the account owner, place that person's approval immediately before the send operation.

Common confusion: The term describes a blocking approval, not a universal safety requirement. Reviewing logs after an action is human-on-the-loop or post-hoc oversight. Automated checks can be the stronger control when they directly establish the relevant fact, while some regulated or high-consequence decisions require a named person or role regardless of model quality.