EU AI Act Article 14: human oversight for AI agents
Article 14 of the EU AI Act says high-risk AI systems must be designed so that people can oversee them effectively while they're in use. For a chatbot that means reviewing answers. For an AI agent that takes actions, oversight has to happen before the action runs, or it's just reading the incident report.
This is a practical reading for engineering and security teams, not legal advice. Whether your system is high-risk, and which deadlines apply, is a question for your counsel; the timeline for high-risk obligations has been the subject of EU proposals to delay parts of it, so check the current dates.
What Article 14 asks for
Paraphrasing the article, the people assigned to oversee a high-risk system must be able to:
- understand what the system can and can't do, and monitor how it operates, so problems are spotted;
- stay aware of the tendency to over-trust automated output ("automation bias");
- interpret its output correctly;
- decide not to use it in a given situation, or disregard, override or reverse its output;
- intervene, or interrupt it through a stop button or similar procedure that brings it to a safe halt.
Oversight measures must match the system's risk, autonomy and context of use. Article 12 adds automatic record-keeping: logs that let you trace what the system did.
Who this applies to
Article 14 applies to high-risk systems, mainly those listed in Annex III: for example AI used in hiring and managing workers, creditworthiness, access to essential services, education, and critical infrastructure. A coding agent on a developer's laptop usually isn't high-risk by itself. An agent that screens candidates, adjusts credit limits or changes systems that run critical infrastructure may be. Providers design the oversight in; deployers must assign people who can actually use it.
What it means for agents that act
- Monitor: every tool call the agent makes is visible, with its input, result and the reason it was allowed.
- Override before, not after: actions that can't be taken back wait for a person who can reject them. Reversing a sent payment or a deleted record is often impossible.
- Counter automation bias: ask rarely and clearly. A person asked 100 times a day approves by reflex. Show the effect in words and the reason it was held.
- Stop: a person can halt one action, one session, or every agent at once.
- Records: who approved what and when, in logs that can't be quietly edited, kept with the version of the rules that made each decision.
How Squidbrake maps to it
Squidbrake sits between agents and the tools they call. A person can see, hold, reject and stop an agent's actions: one action, one session, or every agent. Every decision, approval and settings change goes into a hash-chained audit trail with signed checkpoints, exportable and verifiable offline. That produces evidence for Article 14 and Article 12; it doesn't make a system compliant on its own, and your auditor or counsel decides what it satisfies. The full mapping, including SOC 2 CC8.1 and ISO/IEC 42001, is on our security page.
Try it on your own agents. pipx install squidbrake, then
squidbrake connect all. Start in shadow mode: nothing is blocked in week one, you see what would have been.
Quickstart · Benchmark · Book a demo