Human-in-the-Loop
Human-in-the-loop (HITL) is a design approach in which people take part in an AI or automated system's process by reviewing, correcting, or approving its outputs at defined points.
The term has two common uses. In machine learning, it describes people labeling data, correcting predictions, or rating outputs so a model improves over time. In automation and AI agents, it describes checkpoints where a person must review or approve an action before it takes effect, especially when the action is high risk or hard to undo. The approach is common in fields such as healthcare, finance, and content moderation, where mistakes can cause real harm.
Human-in-the-loop is often contrasted with human-on-the-loop, where people monitor a system and can step in but don't approve each action, and human-out-of-the-loop, where the system acts fully on its own. The right level depends on the risk of errors, not on how capable the technology seems. A common misconception is that adding a reviewer solves the problem. Oversight only works when the person has enough context, time, and authority to catch mistakes.
Examples
A manager approves a payment an agent has prepared before it is sent.
A reviewer checks low-confidence predictions before they reach a customer.
Annotators label data used to train or evaluate a model.
An agent hands a case to a person when a request falls outside its limits.
