Human in the loop

What is human in the loop?

Human in the loop, or mens in de lus in Dutch, means that a person has a fixed place in a process with AI. The AI makes a proposal and a person approves it, adjusts it or rejects it.

A European expert group names three forms of human oversight:

• Human in the loop: a person can intervene in every decision.
• Human on the loop: a person keeps an eye on how it works.
• Human in command: a person decides whether and how the system is used.

According to that group, a person in every decision is often not possible or desirable. Choose per task where the person sits.

How does it work in an SME?

The quote generator I made for De Rijplaten Specialist and De Slowfeeder Specialist, two online shops from Ede, works the same way. AI recognises a request in the email and proposes a price. Only after a click on approve does the quote go to the customer. See the case.

Is human oversight mandatory?

Not for every use of AI. It is in two cases:

• Decisions about people. Under the GDPR (Article 22), a decision that significantly affects someone may as a rule not be taken by a computer alone. Think of an automatic rejection for a loan.
• High-risk AI, such as AI that assesses job applicants. Under the EU AI Act (Articles 14 and 26), you then appoint people with enough knowledge and authority. They must be able to disregard the outcome and stop the system. For recruitment this applies from 2 December 2027.

Even without an obligation, human oversight is sensible for work that affects customers, money or agreements. AI can give a convincing answer that is wrong: a hallucination.

This is an explanation in plain language, not legal advice.

What works, and what doesn't?

Blindly clicking approve does not help. The EU AI Act calls that pitfall automation bias: relying too much on the system. This does work:

1. Choose the decision points. Sending, paying, deleting and decisions about people belong with a person.
2. Make checking easy. Show the proposal next to the source and make deviations stand out.
3. Record what was approved, adjusted or rejected. That way you see where AI is often wrong.
4. Teach your team to check. That is part of AI literacy and belongs in your AI policy.

This is how I build AI applications: AI does the groundwork, you decide. With an AI agent too, every step with consequences waits for your approval. Read more about AI in your business.

Frequently asked questions

Sources: GDPR (Article 22, in Dutch), EU AI Act (Articles 14 and 26, in Dutch), EDPB: automated decision-making (in Dutch), Ethics guidelines for trustworthy AI (EU expert group, 2019, p. 19, in Dutch) and European Commission (timeline). Consulted on 22 October 2026.