Human in the loop (HITL, "the person in the loop") is a way of designing an AI system where a person participates in the decisions that matter: they review, correct or approve before an action takes effect. The term comes from control engineering, where a human operator is part of a machine's feedback loop, and transitioned to AI to describe systems that do not act alone in irreversible matters.
It is not the same as retrospective monitoring. In a HITL system, approval is within the flow: the agent prepares the payment, the dispatch or the data change, and pauses until someone says yes.
What usually goes through a person
- Irreversible actions or those difficult to undo: deleting data, sending an email or a message to a client, processing a payment, deploying to production.
- Decisions with legal or financial consequences: accepting a quote, modifying an invoice, changing terms and conditions.
- Low-confidence cases: when the model is uncertain, when data does not match or when the action is out of the ordinary.
Why it matters
An agent amplifies both the judgement and the errors of whoever directs it. Human approval is not a brake on speed: it is what allows for significant delegation without delegating responsibility. In the EU AI Act, this idea is referred to as human oversight.
In NAiOS, this is a product principle: agents and connectors propose the action and execute it only with human approval, which can be given from the chat, from Talk or from the module itself.