AI agent governance: what to control when AI acts on its own

Why an agent is different

An assistant that drafts text is like a writer: someone reviews what it writes. An agent is like an employee with access to your systems: if it errs, it acts. So it needs employee-grade controls: permissions, limits and supervision.

Six basic controls

1. Inventory: which agents exist, who owns them and what they do. You cannot control what you do not know (watch for agents teams create on their own).
2. Least privilege: only the permissions needed, only on the systems needed.
3. Human approval for sensitive or irreversible actions (payments, deletions, external communications).
4. Logs: what it did, when and why, so it can be audited.
5. Clear limits: what it is forbidden to do and the maximum amount or volume it can handle.
6. Escalation and stop: who is alerted if something fails and how the agent is stopped quickly.

How it fits with the AI Act

The AI Act does not define an agent category: use is what matters. If an agent operates in a high-risk area, the obligations of that area apply, including effective human oversight under Article 14. If it interacts with people, the Article 50 transparency obligations may apply.

Common mistakes

• Broad access “so it works”.
• Rubber-stamp approval: the person approves without understanding what they approve.
• No logs: what happened cannot be reconstructed.

Start here

1. Inventory the agents and AI tools in use.
2. Assign an owner per agent.
3. Review permissions and which actions need approval.
4. Turn on logging and define how to stop them.

Frequently asked questions

What is an AI agent?

A system that plans and carries out actions with some autonomy.

What minimum controls do I need?

Inventory, limited permissions, human approval, logs and a stop procedure.

Informational content, not legal advice.