A register of AI systems describes, for each system, a purpose, a provider, data and a classification. An AI agent adds what none of these fields covers: a technical account, rights in applications, actions that it chooses and carries out itself. Agentic AI governance deals with this addition.
It is obtained by extending the existing set-up (register, classification, roles, approval process) with six additions, in this order: the inventory of agents and their rights, an autonomy level per action with its human approval, a designated owner, traces, a periodic review and a shutdown procedure. Each step delivers a document that the next one uses. “Agentic AI” and “autonomy level” are terms of usage, and this method is a proposal, which we do not draw from any framework. It assumes governance put in place as described in our article on implementing AI governance.
The inventory lists the agents with their rights
The existing register provides a starting point. An agent may appear in it as a simple assistant, or be absent from it if it was connected to internal tools by an employee, a form noted in our entry on shadow AI. The inventory therefore cross-checks other sources: technical accounts and access keys created for AI applications, connected tool servers, for example through the MCP protocol, and automation platforms.
The criterion for inclusion is the one in the glossary entry: the model chooses the next action and carries it out through a tool. Each agent’s record lists its tools, the account used, the rights opened in each application, the sources it reads and the model on which it is based.
OWASP, in the “Excessive Agency” entry (LLM06:2025) of its Top 10 for LLM applications (genai.owasp.org), read on 12 November 2026, generally attributes an agent’s excess of power to one or more of three causes: too much functionality, too many permissions, too much autonomy. The inventory records the first two; the next step deals with the third. The security review of these rights follows our article on AI agent security.
The result is an “agents” section in the register, with, for each one, the matrix of its tools and rights.
One autonomy level per action, and the associated human approval
The level is set for each action: the same agent can read, propose and execute. We propose a four-level convention. Other breakdowns are in circulation, and we attribute none of them to any body.
| Level | The agent | Human approval | Who authorises this level |
|---|---|---|---|
| Reading | Consults and reports, without writing anything | Review by the recipient of the result | The agent’s owner |
| Proposing | Prepares an action that a person carries out | Agreement given before each action | The agent’s owner |
| Reversible execution | Writes to an internal application, within caps | Check on a sample, human takeover outside the scope | The agent’s owner and the application owner |
| Binding execution | Carries out an irreversible action or one directed at a third party | Prior agreement for each action, unless a written exemption exists | The AI governance body |
These levels are based on the classes of action in our article on AI agent use cases in the enterprise; the convention adds the person who authorises each one. The highest-level action sets the process that the agent’s file follows.
Moving an action up a level is a dated decision, recorded in the register and taken on the basis of records: proportion of proposals approved as they stand, reasons for human takeovers. The technical safeguards that enforce these levels are designed as described on our AI agents page, which also refers to Regulation (EU) 2024/1689 for human oversight of an agent classified as high-risk.
This step delivers the grid of levels, carried over into the inventory matrix.
A designated owner, and traces that reach them
The register already names a business owner per system. An agent crosses several applications, and the designation is supplemented: its owner decides on the task entrusted and proposes the autonomy level, each application owner grants the rights that concern them, and one person manages the technical account. When several agents cooperate, we propose that a single owner covers the entire chain, whatever mode is chosen among those in our article on AI agent orchestration. The designation is written, accepted, and provides for a deputy.
Traces are this owner’s material. Governance sets what they contain (what triggered the execution, each tool call, each approval requested and the person who gave it), who reads them and how often. The measures drawn from them are described in our article on generative AI observability. As they may include personal data, their retention period is agreed with the data protection officer.
Two documents join the file: the act of designation and the logging rule.
The periodic review of agentic AI governance
The review compares, for each agent, what was authorised and what happened. It examines the rights granted and those the agent actually used, in order to withdraw rights that have remained unused. It rereads the level of each action in the light of the records. It lists the changes (model version, tool added, instruction modified), incidents and human takeovers.
Some events trigger it without waiting for the due date: a tool added, a change of model, the departure of the owner, an incident. Its frequency is set according to the agent’s highest level.
Frameworks can structure the exercise. The NIST AI Risk Management Framework, voluntary in use, is under revision according to the page the institute devotes to it, consulted on 12 November 2026. The standard presented in our ISO 42001 entry concerns the organisation’s management system. We attribute none of the steps proposed here to them.
The review concludes with minutes and an updated register.
Shutdown is prepared as a decision, with its holders
The technical cut-off (deactivating the account, revoking tokens, closing access to tools) is described in our article on agent security. Governance settles what surrounds it: the people authorised to suspend an agent, including outside working hours, the findings that lead to it, what happens to tasks in progress, the people informed, starting with the owners of the applications affected. It also sets the return: who authorises putting the agent back into service, and after which review. Final withdrawal is added to this: accounts and keys deleted, traces archived, register entry closed.
The register and approval process that receive these additions are described on our AI governance page. Once the six steps have been completed, an agent’s record there shows:
- its tools, its account and its rights in each application;
- the level of each action and the associated approval;
- its owner, the deputy and the application owners who granted the rights;
- the logging rule and the retention period for traces;
- the date of the last review and of the next one;
- the people authorised to shut it down, and the date of the last shutdown test.