The CNIL on agentic AI: 8 pointers for protecting data when you use agents
After Spain's AEPD in February, the French data protection authority has published its own note on AI agents. It creates no new obligations, but it shows what European supervisors will look at.
What this note is
The CNIL published it on 20 July 2026 with the French AI and digital council. It is an exploratory note: it raises pointers and questions, not requirements. It is still useful, because it shows what supervisors expect to find when they review data processing carried out with agents.
The 8 pointers
- Traceability: be able to reconstruct the whole path of a decision: which personal data were used, which agents took part, which external services were called and in what order.
- User control: let each person decide which data the agent can access, especially sensitive data.
- Memory with limits: partition memory by agent and by process, with a maximum size and automatic expiry.
- Filters at every step: check what goes in and out not only at the user's first request, but every time an agent calls a model.
- Risk-based oversight: classify actions by risk and require human approval for the higher-risk ones.
- Containment: run agents in sandboxes and give users a kill switch.
- Evaluation: include data protection and security in independent testing of agents.
- Automated decisions: clarify, sector by sector, when an agent triggers Article 22 of the GDPR.
How it fits with Spain's AEPD guidance
Spain's AEPD published its guidance on agentic AI on 18 February 2026, an 81-page document. They agree on the essentials: control memory, keep human oversight and design on the assumption that errors will happen. The AEPD considers dedicated governance of agentic systems, across their whole life cycle, the most important measure, and asks for the data protection officer to be involved.
If you operate in Spain, the AEPD guidance is your main reference. The CNIL note works as an extra checklist.
What comes next
The European Data Protection Board and the Commission are preparing joint guidelines on how the GDPR and the AI Act fit together, expected by the end of 2026. The CNIL suggests they include specific recommendations for those who build agents and those who use them.
What to do now
- Check whether you can reconstruct what each agent did in a specific operation. If you can't, start with logging.
- Review what each agent remembers and for how long.
- Flag the actions that need human approval: payments, deletions, anything sent outside the company.
- Make sure someone can stop an agent within minutes, and knows how.
- If the agent makes decisions about people, check with your data protection officer whether Article 22 of the GDPR applies.
Frequently asked questions
Is the CNIL note binding outside France?
No. It is an exploratory note from the French authority. Each country's data protection authority is the reference for processing there, although the ideas are very similar.
What is a kill switch?
A simple, fast way to stop an agent when it does something it shouldn't, without waiting for the technical team.
For information only, not legal advice.
What's your Data Governance maturity?
Free assessment with your priority gaps and the templates on the Data Governance path.