What does Human-in-the-Loop mean?

Human-in-the-Loop (HITL) describes AI systems in which a person is involved at defined points in the flow: they review, correct, or approve before the system continues. For AI agents this mainly concerns actions with external effect or high risk, such as postings, payments, or master-data changes. Exactly where the person steps in is a deliberate design choice between speed and control.

Why does this matter?

Autonomy without intervention points is neither permitted nor sensible in most business processes; constant check-backs, on the other hand, make an agent useless. Human-in-the-loop defines where responsibility stays with a person – and the EU AI Act explicitly requires effective human oversight for high-risk systems.

What does this look like in practice?

An agent prepares credit notes for complaints. Up to €200 it posts them itself; above that it puts a draft with reasoning and case history in front of the team lead. After three months the analysis shows that 96% of drafts were approved unchanged — reason to raise the limit to €500 and focus human review on the difficult cases.

Not to be confused with

Not to be confused with human-on-the-loop, where the system works on its own and the person supervises and steps in only when needed rather than approving each decision. Nor with manual rework, which fixes errors afterwards, whereas human-in-the-loop places the check before the effect.