What human oversight means for AI agents
Carlos Lebre Ribeiro ·
Human oversight of an AI agent means a competent person, with the authority to act, reviews the decisions that matter and can stop the ones that are wrong. It does not mean a person clicks yes on everything.
Why approving everything fails
When every action needs approval, approving becomes a reflex. Anthropic reports that Claude Code users approve 97% of permission prompts, and that people caught 13.6% of dangerous commands in a controlled test.
Oversight only works if the volume reaching a person is small enough for them to pay attention. So something has to decide what reaches them.
What the EU AI Act asks
Article 26 of the EU AI Act (Regulation (EU) 2024/1689) asks deployers of high-risk systems to give human oversight to people with the competence, training and authority to use it. It also asks them to monitor the system, report serious incidents and keep the logs it generates for at least six months.
Under the Digital Omnibus on AI, in force since 27 July 2026, these obligations apply to most high-risk systems (Annex III) from 2 December 2027, and to high-risk AI in regulated products (Annex I) from 2 August 2028.
What good oversight looks like in practice
- Safe actions clear on their own, against a written policy.
- Risky actions reach a named reviewer, with the context they need to decide.
- Every outcome is recorded: who or what decided, when, and on what basis.
- The share of decisions going to a person is measured, so you can see it drift.
This article is general information, not legal advice.