Human-in-the-loop

Human-in-the-loop means a person decides, approves or corrects at a point in an AI system's workflow that was defined in advance. Not every output gets checked, only the ones where a mistake is expensive. What matters is that the checkpoint is deliberate rather than left to chance.

What does human-in-the-loop look like in practice?

Most systems use one of three patterns.

Approval before effect: the system prepares something, a person approves it, and only then does anything reach the outside world. This fits anywhere a mistake lands directly with a customer, such as quotes or credit notes.

Sampling after effect: the system acts on its own and a share of cases is reviewed afterwards. This fits high volume with limited damage per case, such as sorting incoming requests.

Escalation on uncertainty: the system acts while it is confident and hands over as soon as it is not. That assumes uncertainty is measured at all, which in practice is the demanding part.

Which pattern fits depends on the cost of a single error and on volume. An AI agent with no defined handover tends to collapse into the most expensive pattern of all: eventually somebody checks everything by hand, and the time saved disappears.

Why is human-in-the-loop more than a safety measure?

Three reasons that come up repeatedly with management, data protection officers and works councils.

Accountability: a machine cannot be accountable for a decision. When an invoice is approved in error, there has to be a person with whom the decision sat.

Law: the EU AI Act explicitly requires human oversight for high-risk systems. Even outside that category, data protection officers expect an answer to the question of who can overrule an automated decision.

Improvement: every correction is a data point. A system where people intervene at a defined step collects exactly the cases it gets wrong. Without that loop, it stays unclear why a result was poor.

Bitkom reports that 53% of German companies name legal uncertainty as a barrier to using AI. A control point that can be explained removes part of the ground that uncertainty stands on.

What human-in-the-loop is not

It does not mean somebody proofreads everything. Full review of every output usually costs more time than the automation saves, and it wears people down: after approving two hundred correct suggestions, nobody catches the mistake in the two hundred and first.

It is also not a temporary measure to be removed once a system is live. The checkpoint moves as the system matures, from approving every case towards sampling, but in serious applications it does not go away.

For how this plays out on the way into production, see Shipping AI that survives production.