Picture a fifteen-person trading company in France. Since September, an n8n workflow has been reading messages sent to the contact address. It sends them to Claude through the API to classify them and then prepares a draft reply. A second workflow collects supplier invoices as PDFs and asks a Mistral AI model to extract the amount, the due date and the bank details.
Every run sends the sender's name, email address and message, sometimes a phone number or an IBAN, from the company's server to a model provider's server. Nobody has reviewed these workflows from a personal data angle, because they were built as productivity tools.
GDPR compliance for an AI agent comes down to a handful of concrete checks. This article follows the data through this kind of workflow and turns it into a checklist for a company of 5 to 50 people. Legal decisions, such as choosing the legal basis, stay with the DPO or a lawyer. The sources are French, since the example company falls under the CNIL, the French data protection authority, but the questions apply to any company subject to the GDPR.
The personal data an AI agent actually touches
A July 2026 note from the Conseil de l'IA et du numérique, the French advisory council on AI and digital affairs, published on the CNIL website, points out that agentic AI remains fully subject to the GDPR. The note separates context (conversation history, instructions, interactions with other services), which is deleted at the end of the process, from memory, which keeps information from one process to the next.
In an n8n workflow connected to Claude or Mistral, personal data passes through four places:
- The prompt. The customer's message travels with the instructions: a signature, a phone number, a quoted thread, sometimes information about a third party.
- Attached documents. An invoice contains a company name, a contact's name, an address and an IBAN. A CV or a sick note that arrives by mistake follows the same path if nothing filters it out.
- Tool results. When the agent calls a tool, for example via MCP, the response enters the model's context: a CRM customer record, orders, outstanding balances.
- Traces on your side. The n8n execution history, error workflows that copy the message into an email or a Slack post, and the agent's memory if you configured one.
A simple test is to open a past execution in n8n. Everything you see there is stored somewhere.
Who is responsible: the company, the model provider, the host
The company decides to connect the agent, for what purpose and on which data. The note stresses that with an agentic system, a data controller must always be identifiable and able to meet its obligations. Here, that is the company.
Providers that process data on the company's behalf are processors. In its Q&A on using a generative AI system, published in July 2024, the CNIL states that relying on external infrastructure must be secured by a data processing agreement with the system's host and, where relevant, with the AI system provider. It adds that with an API, control of the system lies "almost exclusively in the hands of the provider" (my translation).
In the example, the processors are the model provider (Anthropic or Mistral AI), the host of the n8n server (or n8n itself on its cloud plan), the existing software vendors (email, CRM, accounting) and the contractor who maintains the workflows.
One case calls for legal advice. The same note observes that some providers may reuse their users' data to train or improve their models. If that applies to your plan, the provider is no longer processing only on your behalf, and the question of who plays which role goes to the DPO.
What to check with Anthropic, Mistral AI or any other provider
Provider terms change often and depend on the plan, so I won't summarize them here. Read the versions in force on the day you go live and archive them with their date. In the same Q&A, the CNIL states that it is up to the user or integrator to question the provider about its system's GDPR compliance.
Six points to check:
- The plan you actually use. Read the terms for the API key in the workflow, not those of the consumer app. An API key created on an employee's personal account is a warning sign.
- The data processing agreement (DPA). Does one exist for your plan? Do you need to sign it, and who accepted it on behalf of the company?
- Training. Is the data you send used to train or improve the models, by default or as an option? If the console has a setting for this, write down its state.
- Retention. How long are requests and responses kept, including for abuse detection?
- Processing location and transfers. Where is the data processed, and what mechanism covers a transfer outside the European Union? Mistral AI's French headquarters do not settle this on their own.
- Sub-processors. Is the list published, and how are you notified of changes?
The guide "Choisir une solution d'IA générative" (choosing a generative AI solution), produced by France Num with the CPME and the CNIL, helps frame this choice. If the answers are not acceptable for some of your data, an open model hosted on your own infrastructure remains an option (see open vs proprietary AI models).
Minimize before calling the model
The note restates the data minimization principle of Article 5(1)(c) of the GDPR (Regulation (EU) 2016/679, applicable since May 25, 2018): only data necessary for the purpose should be processed. In n8n, this happens in the steps placed before the node that calls the model.
- Send only the fields you need. To classify an email, the subject and body are often enough. A code step can strip signatures and the quoted thread. Headers and attachments don't need to be sent.
- Mask identifiers. Replace email addresses, phone numbers and IBANs with placeholders (
[EMAIL_1],[IBAN_1]) before the call. Put them back afterwards if the draft needs them. - Route without a model what must not leave. Simple rules on the sender or on keywords can send messages about health, HR or disputes straight to a human. Health data is one of the special categories of data under GDPR Article 9, which the note cites.
- Keep documents on your side. An invoice received in a structured electronic format can be read without a model (see processing incoming supplier e-invoices with n8n). For a PDF, send only the text you need.
- Restrict tools. An MCP server should expose only the tools and fields required, read-only where possible (see the n8n MCP server with controlled access).
Execution history, logs and memory: set retention periods
The storage limitation principle (GDPR Article 5(1)(e), cited in the note) means data is kept only as long as necessary. On your side, an agent leaves data in three places.
The n8n execution history. By default, n8n saves both successful and failed executions. Its documentation on execution data describes the settings for what gets saved and for automatic deletion. Check your instance's actual configuration, then set up pruning (see securing self-hosted n8n).
Copies. An error workflow that forwards the original message, a tracking spreadsheet or a Slack alert each creates a copy with its own lifespan, often unlimited. I recommend alerting with an execution ID rather than with the content.
The agent's memory. The note recommends isolating memory per agent and per processing purpose, limiting its size and making information expire automatically. In practice: one memory per workflow, one key per conversation, an expiry date.
I recommend setting and writing down retention periods workflow by workflow, for example a few days for successful executions and a few weeks for errors. The DPO signs off.
Record of processing, informing people and DPIA
The record. I recommend one entry per agent in the record of processing activities: purpose, categories of people and data, recipients (model provider, host, contractor), processing location, retention periods and measures taken.
Informing people. The note recalls the transparency principle (GDPR Article 5(1)(a)) and the rights in Articles 15 to 22, including access and erasure. Check that the information you give customers, suppliers or applicants covers this processing. For an erasure request, know where to look: the n8n server, the copies, the memory and the provider.
Automated decisions. GDPR Article 22, which the note cites, covers decisions based solely on automated processing that produce legal effects or significantly affect the person. An agent that sorts emails decides nothing about anyone. An agent that rejected a job application would.
Human approval only protects you if it is genuine. Citing the Court of Justice of the EU's SCHUFA ruling (C-634/21), the July 2026 note states that human intervention must be real, effective and able to influence the decision: a formal or automatic sign-off is not enough. The person who approves must see the file and be able to change the outcome.
Data protection impact assessment (DPIA). Whether a DPIA is mandatory is a question for the DPO or a lawyer. I recommend raising it as soon as an agent touches recruitment or employee evaluation, processes special categories of data, combines several sources about the same person or prepares decisions that concern that person.
What the CNIL plans on AI in 2026
The page IA, comment se mettre en conformité (AI: how to comply) gathers several of the resources cited here, including the generative AI Q&A and the guides for small businesses.
In its 2026 work program, published on April 7, 2026, the CNIL says it will finalize its work on the use of AI in the workplace and in healthcare. It also plans to clarify the responsibilities of the players along the AI value chain, from model developers to deployers. This is the question of roles discussed above.
On the AI Act itself, see what applies to companies under the EU AI Act.
GDPR AI agent checklist before going live
| Check | Who |
|---|---|
| List the data that leaves: prompt, attachments, tool results, memory | Technical lead |
| Locate the processing agreements: model provider, host, contractors | Company head |
| Archive the DPA, training rules, retention, processing location, sub-processors | Company head and technical lead |
| Tie the API key to the company account | Technical lead |
| Filter, mask and route before the call; restrict MCP tools | Technical lead |
| Configure retention: executions, copies, memory | Technical lead, approved by the DPO |
| Add the entry to the record, update the information given to people | Company head or DPO |
| Know where to look for an access or erasure request | Technical lead |
| Leave any decision that affects a person to someone who reviews the file and can change the outcome | Company head |
| Decide on the legal basis, ambiguous roles and the DPIA | DPO or legal counsel |
The last line cannot be delegated to a technical contractor. The note points out that all processing rests on one of the six GDPR legal bases, and that an agent's autonomy can weaken the link with the basis chosen at the start. The CNIL's 2026 work program describes the DPO's advisory role on AI use as essential. Without a DPO, a lawyer makes the call.
At Kirako, we provide audits and consulting on the technical side of these workflows: data flows, filters, pruning and tool permissions.
An agent connected to your email or accounting raises the same questions as any service provider: where the data goes, how long it stays and who can read it. It is better to answer them before going live than after a customer asks.
Sources
All CNIL and council documents are in French.
- CNIL: Q&A on using a generative AI system (July 2024), 2026 work program (April 7, 2026)
- Conseil de l'IA et du numérique: IA agentique et protection des données personnelles (agentic AI and personal data protection), note of July 2026
- European Union: Regulation (EU) 2016/679 (GDPR) of April 27, 2016
- n8n: execution data documentation
Also available: Read in French