Blog / Personal data

Agentic AI: what France's data protection authority says about agents and personal data

No definition of agentic AI commands agreement, the CNIL notes. It nevertheless agrees with other authorities on one criterion, autonomy of action, and describes risks that stem less from the technology than from the delegation of power.

By Vincent Dorange, , 3 min read

5 degrees

of autonomy in the CIANum scale the CNIL refers to

On 2 September 2026 LINC, the digital innovation lab of France's data protection authority (CNIL), published "IA agentique : de quoi parle-t-on ?" (Agentic AI: what are we talking about?), a paper by Romain Darous, an engineer in the CNIL's artificial intelligence department 1. It presents itself as a survey of definitions. It is more useful than that, because it shows how the French data protection authority is starting to look at agents: through what they do with data and through the power that is delegated to them.

No shared definition, one shared criterion

The first finding is that no definition of agentic AI commands agreement. The paper recalls that the OECD listed about ten in a February 2026 report, then compares those of the European Commission, the UK ICO, Spain's AEPD, the European Data Protection Supervisor and France's CIANum. The CIANum's is the broadest: an AI agent is a program able to take decisions or carry out actions, including coordinating them, by relying on AI models.

Behind the variety of wording, one criterion keeps coming back: autonomy of action. A model that looks up information remains a tool; a system that acts on its environment, sending an email, booking an appointment or confirming a purchase, becomes an agent. The paper cites the CIANum's scale of 5 degrees of autonomy, from the mere application of predefined rules to fully autonomous flows with no human involvement in intermediate decisions. According to the CNIL, that last degree remains a horizon that raises fundamental legal questions.

The risks the CNIL describes

The paper's conclusion, which draws on a note written with the CIANum, sets out the risks to personal data 1.

  • Data circulating between many services, which users struggle to follow, with a real risk of losing control over their own data.
  • Persistent memory and interaction history, which multiply the data kept and encourage hyper-personalised profiles.
  • The sharing of responsibility: by automating decision processes that involve several parties, agents make it harder to identify what falls to whom.
  • Cybersecurity, whose risks extend to every service connected to the agent.

The CNIL does not call for new law. It considers that the GDPR and the European AI rules already apply, while finding that decision-making autonomy, persistent memory and the ability to act on the user's behalf require their implementation to adapt.

The phrase that matters: delegation of power

The most important passage comes down to one expression. The CNIL writes that agentic systems raise questions about how responsibility is shared because of the delegation of power they make possible. That is exactly what the Agentic Commerce Framework® is about: decisions are delegated to an agent, never the responsibility that goes with them.

For a company, this reading turns into very concrete GDPR obligations. The purpose limitation principle 2 applies to every use an agent makes of a piece of data, including when it draws on its memory for a new task. A data protection impact assessment is required before processing likely to result in a high risk 2. Article 22 gives everyone the right not to be subject to a decision based solely on automated processing that significantly affects them, subject to exceptions 2. Each of these rules assumes the company knows, decision by decision, what the agent does alone and what a human approves.

The ACF provides that breakdown. The agent's mandate sets what it may decide, the decision classification shows which decisions affect people, and a named owner answers for each of them. The CNIL describes the risk; an organisation that has written down its delegations can show how it keeps that risk under control.

Sources

Links to the official texts and original publications.

  1. 1IA agentique : de quoi parle-t-on ?, by Romain Darous (in French), CNIL, LINC, 2 September 2026.
  2. 2Regulation (EU) 2016/679 (GDPR), Articles 5, 22 and 35, Official Journal of the European Union, 27 April 2016.

Glossary terms used in this article

AI agent delegationAI agent mandateAI agent decision classificationAI agent owner

Vincent Dorange, creator of the Agentic Commerce Framework®

Also worth reading

95%

of organisations surveyed faced at least one agent-related incident in 12 months

Governance

95% of organisations have already had an AI agent incident, says Gartner

25

organisations interviewed; none lets its agents act with unrestricted autonomy

Governance

AI agents at work: what the OECD learned from 25 organisations

4 levers

transaction caps, budgets, approved counterparties and escalation rules, per Circle

Agentic economy

Agentic economy: agents that spend within delegated permissions

All articles