‘Persistent memory’: the CNIL and CIANum name the risk of agentic AI, and it is exactly what a domiciliation branch reception has to explain
In their exploratory note of 20 July 2026, the CNIL and France's Council for AI and Digital Affairs write that agentic AI ‘marks a change of scale that renews and amplifies the risks to personal data’: data flowing ‘between many services’, retention ‘of the history of interactions with the user’, ‘hyper-personalised’ profiles, responsibilities harder to pin down. A conversation that follows a visitor from one branch to the next is a persistent memory. Better to say so plainly.
The note does not target a product: it describes a mechanism. An agent acting on the user's behalf has to remember, pass things on and chain actions. The CNIL and CIANum draw four risks from that: personal-data flows between services multiply the leak points; the memory of interactions increases the volume of data retained; those memories favour hyper-personalised profiles and a ‘real risk of losing control over one's personal data’; finally, decision-making autonomy ‘complicates the identification of each party's responsibilities’, and cyber risks extend ‘to all the services connected’ to the system. The legal framework remains the GDPR, but the way it is implemented has to be adapted, in the note's own words.
For a domiciliation network, this note is an honest test to apply to AppH's Pack Accueil. Its promise, ‘one conversation that follows the visitor from the screen to the phone’, is a persistent memory in the CNIL's sense. So here is what it holds and what it does not. It keeps the context useful to the domiciliation file: branch visited, request in progress (mail collection, certificate, missing documents), agreed times. It builds no profile of the visitor beyond that file, flows to no service other than those configured by the provider, and its retention period is a setting the customer fixes, not an AppH default. No act is triggered by the assistant alone: every hand-over, every document, every change goes through a branch staff member who approves it. Hosting in France, a data-processing agreement under article 28, and a written answer to ‘who is responsible for what’ before going live, not after.
For AppH
- The note describes agents chaining actions across many services. The Pack Accueil is deliberately narrow: one trade, one network of branches, one file, the simplest scope to explain to a visitor and to a prefecture.
- ‘A human approves every act’ answers the note's point on responsibilities directly: the staff member remains the author of the decision, the provider remains the data controller, AppH remains the processor.
Against / the honest limit
- It is an exploratory note, not a recommendation or a ruling: it opens questions (memory duration, reuse of exchanges, security of connected services) without answering them. A network waiting for an official checklist does not have one yet.
- A conversation memory, however narrow, is still one more personal-data store to record in the register, document and purge. AppH supplies the settings; the provider keeps the burden of deciding and justifying them.
We could have read this note as a threat to our sales pitch, since continuity of the conversation is what we sell. We read it as a list of questions to put to any vendor, ourselves included: what does the assistant retain, for how long, where does it flow, and who decides when the conversation leads to an act. A domiciliation network under a prefectoral licence is already used to answering for its premises and its registers; it can demand the same precision about the memory of its reception. If a vendor cannot answer on one page, the July note explains clearly enough why that should worry you.
Reviewed by a human at AppH