AI agents, capable of taking initiatives within corporate information systems, are disrupting existing governance models. Identity, access control, traceability, and human supervision become essential elements to reconcile autonomy and security, as the AI Act gradually comes into effect.
- AI agents act autonomously, which challenges the traditional identity-based governance model.
- Each agent needs an identifiable owner, a limited access perimeter, and regularly reviewed permissions.
- Sensitive actions require human validation, and agents must be interruptible when they exceed their scope.
- The AI Act’s high-risk rules apply from December 2, 2027, but governance must start as soon as agents are granted rights.
The traditional governance model is based on a simple principle: one person corresponds to one identity, specific rights, and a clearly identified manager. With AI agents, this logic is challenged. These programs can be created quickly, evolve over time, and see their access rights expand according to their missions. Deployed outside of usual processes, they can cause a company to lose track of the number of active agents, their managers, and what they are actually authorized to do.
According to the Okta Global CISO Insights 2026 report, 81% of surveyed security leaders express concern about excessive access granted to AI, while only 47% claim they can identify all AI agents present in their environment. The AI Act does not regulate ‘AI agents’ as a category. However, for certain systems and certain risk levels, it imposes an increased ability to identify, frame, monitor, and control deployed systems. The question is therefore no longer just whether the company uses AI, but what is authorized to act on its behalf.
For high-risk systems, the regulation imposes requirements for risk management, traceability, monitoring, and human control. Before allowing an agent to intervene in a critical process, it is necessary to know what data it can access, which applications it can use, what actions it can perform, who is responsible for it, and how to suspend its access if necessary. The text is also part of a broader movement, including notably DORA and NIS2, which converge towards the same requirement: identify resources, understand access, and be able to demonstrate what happened.
Identity, the foundation of the chain of responsibility
Each agent must be identifiable, linked to a team or a manager, have a defined access perimeter, and have its permissions regularly reassessed. An agent is not an employee, and it should not be managed like one. The challenge is to be able to reconstruct the chain of responsibility when it acts, meaning linking an automated action to an authorization and a responsible party. Some tasks can be fully automated, others must remain subject to human validation or be interruptible when they present a particular risk.
Four steps to regain control
Map the existing landscape. The first step is to inventory the tools and agents used, their data access, and the people who control them. This inventory must also cover agents created in development environments and automated workflows that escape traditional IT processes.
Assign an identity and a perimeter. An agent must have a designated owner and permissions corresponding to its mission, following the principles of least privilege, time-limited access, and separation of duties. Its machine nature should not lead to granting it more trust than a human user.
Keep the human in the loop. For sensitive actions, the company must define when human intervention is necessary, whether it’s approval before execution or the ability to quickly interrupt an agent that goes beyond its perimeter.
Verify business value. Autonomous projects launched without a clear objective or supervision can multiply risks without demonstrating their usefulness. It is therefore advisable to define the problem to be solved, the expected value, and how to measure results, which also helps determine the necessary level of autonomy.
For high-risk AI systems falling under Annex III, the rules will come into application from December 2, 2027. The source, however, reminds us that the governance of an agent is built when the company grants it the right to act, not at the regulatory deadline. By treating these agents as an identity issue, and not just a compliance or technology one, organizations can move from a ‘shadow AI’ logic to a governed AI.
