FREN

Let’s keep in touch

Social

Regulation· 12 min read

Generative AI and GDPR: what to check before deploying an internal assistant

Generative AI and GDPR: contract, data, transfers, DPIA, works council and training. What to check before deploying an internal AI assistant.

An internal AI assistant, meaning a generative AI tool made available to employees to write, summarise or query the company’s documents, almost always processes personal data: what employees enter into it, what is in the documents it reads, and the conversation history. The GDPR therefore applies as soon as it is used. Before deploying one, five topics need checking: the contract with the provider, the data the assistant can access, the country the data goes to, the impact assessment and, in French companies with at least 50 employees, consultation of the works council (CSE). This guide covers them one by one, based on the CNIL (the French data protection authority), the European Data Protection Board and recent court decisions.

It is aimed at executives, CIOs and compliance managers who want to equip their teams with an off-the-shelf assistant such as ChatGPT or Copilot, or to build one connected to their own documents, in-house or with an artificial intelligence provider.

Which deployment mode should you choose for an AI assistant?

The deployment mode determines who controls the data: the more the assistant is hosted by the company, the more control it keeps; the more it is self-service, the more that control shifts to the provider. The CNIL distinguishes four cases in its questions and answers on using a generative AI system.

Deployment modeWho controls the systemWhat the CNIL saysFor which uses
Consumer serviceThe providerOnly for non-confidential uses, with a work email address, no personal account, data reuse switched offRewording a public text, brainstorming
Professional cloud offeringShared, depending on the contractA processing agreement that sets out responsibilities and accessWriting assistant for teams
Model called through an API from an applicationThe provider, “almost exclusively”Avoid sending personal data through itFeatures without personal data
On-premises deploymentThe company“Generally more appropriate and more secure”Internal documents, confidential or sensitive data

Table based on the CNIL’s questions and answers of 18 July 2024. The “For which uses” column is our summary.

The CNIL recommends starting from concrete needs rather than from the tool. The risk depends on the use: an assistant that helps write an email does not expose people in the way a tool that helps make a decision about a customer or a candidate does. The practical sheets for small businesses, published with France Num and the CPME, follow the same logic: the need first, the solution second.

Who is responsible: the company or the provider?

The company that deploys an AI assistant is the controller for the data its employees enter into it; the provider that processes that data on its behalf is, in principle, its processor, and the contract must say so. The CNIL asks for these roles to be defined before deployment: processing on behalf of the company, or joint controllership if the provider pursues its own purposes. The processing agreement contains the clauses required by Article 28 of the GDPR.

Six points deserve a line in the contract or in the settings of the offering.

  • The provider does not reuse prompts, documents or answers to train or improve its models.
  • Retention of the conversation history: how long, and whether it can be switched off.
  • The list of sub-processors and the country where each one processes the data.
  • The hosting location and, if there is a transfer outside the EU, the tool that governs it.
  • Deletion or return of the data at the end of the contract.
  • Security measures and the provider’s assistance in the event of a data breach.

The processing is entered in the record of processing activities (Article 30). The CNIL makes a point that is often overlooked: a company that fine-tunes the model with its own data, or connects it to a database of personal data, becomes the controller for that processing. The model itself is not neutral either: according to Opinion 28/2024 of the European Data Protection Board (EDPB), a model developed with unlawfully processed data could affect the lawfulness of its deployment, unless it has been duly anonymised. Asking the provider how its model was trained is part of the choice.

What data can the assistant see?

An AI assistant should only see the data needed for its intended use: this is the data minimisation principle in Article 5 of the GDPR, and it is where projects most often go wrong. The CNIL reminds users that they should only submit information they are authorised to share. For an assistant connected to company documents, the rule translates into five measures.

  • Connect only the useful sources. The folders for the use case, not the whole mailbox or the whole drive.
  • Carry over access rights. The assistant only answers from documents that the person querying it can already open.
  • Exclude sensitive data. Health, opinions, trade union membership and the other categories in Article 9 are prohibited in principle, with limited exceptions.
  • Set retention periods. Conversations, indexes and document copies have a retention period, and a purge.
  • Mask the logs. The tool’s technical traces contain the questions asked: they are protected like everything else.

How an assistant connected to internal documents works is explained in our article on RAG in the enterprise. For development and acceptance testing, the same principles apply: see test data and GDPR.

Where does the data go: hosting and transfers outside the EU

If the provider, or one of its sub-processors, processes the data outside the European Economic Area (EEA), that is a transfer, which must be governed by one of the tools in Chapter V of the GDPR: an adequacy decision by the Commission or, failing that, appropriate safeguards such as standard contractual clauses. The CNIL points this out for generative AI as for any processing.

For the United States, the framework in force is the Data Privacy Framework: adequacy decision (EU) 2023/1795 of 10 July 2023 covers certified US companies. On 3 September 2025 the EU General Court dismissed an action for annulment of this decision, and an appeal was lodged with the Court of Justice on 31 October 2025. Before relying on this framework alone, check that the provider is on the list of certified companies and where the proceedings stand.

Hosting in the EU does not settle everything. According to the EDPB’s Recommendations 01/2020, remote access from a third country to data located in the EEA is also a transfer. You therefore need to know the country of the technical support and of the sub-processors, not just that of the servers.

Do you need an impact assessment (DPIA)?

A data protection impact assessment (DPIA) is mandatory when the processing is likely to result in a high risk to individuals: in practice, when it is on the list published by the CNIL, or when it meets at least two of the nine criteria set by the European authorities (Article 35 of the GDPR). The nine criteria, according to the CNIL:

  • evaluation or scoring of individuals, including profiling;
  • automated decision-making with legal or similar effect;
  • systematic monitoring;
  • sensitive or highly personal data;
  • data processed on a large scale;
  • matching or combining datasets;
  • vulnerable data subjects;
  • innovative use or application of new technological solutions;
  • processing that prevents individuals from exercising a right or using a service or contract.

An AI assistant can count as an innovative use: a second criterion is then enough for a DPIA to be required. An assistant that helps write emails does not have the same profile as a tool that summarises annual reviews or sorts job applications: the latter evaluates people and affects their employment. These employment-related uses are also classified as high-risk by the AI Act, with obligations postponed to 2 December 2027 (see our update on the AI Act). In all cases, the CNIL recommends involving the data protection officer (DPO) from the start.

Do you need to consult the works council?

In a French company with at least 50 employees, introducing an AI assistant may require consulting the works council (comité social et économique, CSE): Article L2312-8 of the French Labour Code provides for it for “the introduction of new technologies” and for any significant change affecting health and safety or working conditions.

Judges do not treat it as an automatic rule, as two recent decisions show. On 14 February 2025, the Nanterre court, in summary proceedings, suspended the deployment of AI tools until the works council consultation was complete: the “pilot phase” put forward by the employer was analysed as a first implementation. On 10 February 2026, the Paris court, in summary proceedings, ruled the other way: a four-month voluntary test of Copilot 365, with no precisely identified impact on work, did not require consultation.

The common criterion is the foreseeable effect on work: what the tool changes for employees, not the name given to the project. Informing the works council from the pilot onwards, describing its scope, duration and the data processed, then consulting it before the general roll-out, avoids seeing a project suspended along the way.

What rules should users be given?

Users need written rules and training: the CNIL recommends an internal policy setting out authorised and prohibited uses, and the AI Act requires AI literacy measures for staff who use these tools. A useful policy fits in five points.

  • What may be entered, and what never is: only information one is authorised to share.
  • Template prompts for common uses, which avoid pasting in unnecessary data.
  • Checking outputs: the CNIL asks that the answers of these systems never be reproduced as they are, and that their accuracy be checked.
  • Which account to use: a work email address, never a personal account.
  • Who to contact when in doubt: the DPO and the chief information security officer (CISO).

Article 4 of the AI Act covers providers and deployers of AI systems. According to the European Commission’s questions and answers, a company whose employees use ChatGPT is concerned: at a minimum it must inform them of the risks, such as “hallucinations”. No certificate is required; an internal record of training is enough. The obligation has applied since 2 February 2025, and its supervision by the authorities since 2 August 2026.

The checklist before deploying

  • The use case is written down, with the data it requires.
  • The deployment mode is chosen according to that data.
  • The contract rules out reuse of the data for training and sets the retention of the history.
  • The sub-processors and the countries of processing are known, and transfers are governed.
  • The assistant only accesses the documents it needs, with the user’s rights.
  • The processing is entered in the record, and the need for a DPIA is examined with the DPO.
  • The works council is informed, and consulted if the project changes working conditions.
  • The policy is circulated, users are trained, and the training is recorded.

Frequently asked questions

Can ChatGPT be used at work without breaching the GDPR?
Yes, provided the offering matches the data. The CNIL limits consumer services to non-confidential uses, with a work email address and data reuse switched off. As soon as personal or confidential data goes into the tool, you need a professional offering with a processing agreement, or an on-premises solution.
Does an internal AI assistant need a DPIA?
Not always. An impact assessment is mandatory if the processing is on the CNIL’s list or meets at least two of the nine European criteria, such as innovative use, evaluation of individuals or sensitive data. An assistant that touches human resources or decisions about people almost always requires one.
Do you need to consult the works council before deploying Copilot or ChatGPT?
In a French company with at least 50 employees, yes if the tool is a new technology that changes working conditions. Judges look at the foreseeable effects: a limited, voluntary pilot was held not to require consultation in Paris in February 2026, while a pilot treated as a deployment was suspended in Nanterre in February 2025.
Can the provider train its model on our data?
Only if the contract or the settings of the offering allow it. The CNIL recommends objecting to such reuse and to the recording of the history when it is not needed. Check the terms of the offering and write it into the contract.
Can you choose a US provider?
Yes, if the transfer is governed: by the Data Privacy Framework for a certified company (adequacy decision of 10 July 2023, upheld by the EU General Court on 3 September 2025 and subject to an appeal lodged on 31 October 2025), or by standard contractual clauses.
Do employees who use AI need training?
Yes. Article 4 of the AI Act requires companies that deploy AI systems to take measures to ensure the AI literacy of their staff. No certificate is required, but training is recorded in an internal register.

An AI assistant is deployed as a project, not as a subscription

Opening accounts is not enough: an internal AI assistant needs a use case, a contract, a data scope, a risk analysis and rules for users. Handled that way, it is deployed without unpleasant surprises; handled as a mere subscription, it exposes the company to data processing it does not control, or even to suspension by a court.

iones’ nearshore team designs and integrates assistants connected to company documents (LLM and RAG integration) and helps teams get to grips with them (strategy and AI adoption). You can also write to us for an initial conversation.

Sources

  1. Les questions-réponses de la CNIL sur l’utilisation d’un système d’IA générative – CNIL, 18 July 2024
  2. Utiliser l’IA générative dans les TPE et PME : fiches pratiques – CNIL, France Num and CPME, March 2025
  3. Ce qu’il faut savoir sur l’analyse d’impact relative à la protection des données (AIPD) – CNIL
  4. GDPR, Articles 5 and 9: principles and special categories of data – Text of the regulation, CNIL website
  5. GDPR, Articles 28, 30 and 35: processor, record, impact assessment – Text of the regulation, CNIL website
  6. GDPR, Chapter V: transfers of data outside the Union – Text of the regulation, CNIL website
  7. Opinion 28/2024 on AI models and GDPR principles – EDPB, 18 December 2024
  8. Recommendations 01/2020 on measures that supplement transfer tools – EDPB, final version of 18 June 2021
  9. Implementing Decision (EU) 2023/1795 on the adequacy of the EU-US Data Privacy Framework – European Commission, 10 July 2023
  10. Data Privacy Framework: General Court ruling of 3 September 2025 and appeal of 31 October 2025 – WilmerHale, 1 December 2025
  11. AI literacy: questions and answers on Article 4 of the AI Act – European Commission, updated 20 October 2026
  12. Article L2312-8: works council powers in companies with at least 50 employees – French Labour Code
  13. Quand le CSE stoppe le déploiement de l’IA (Nanterre court, 14 February 2025, no. 24/01457) – CMS Francis Lefebvre, 14 April 2025
  14. Copilot 365 n’est pas un projet important rendant nécessaire la consultation du CSE (Paris court, 10 February 2026, no. 25/57412) – CMS Francis Lefebvre, 12 March 2026

Start a project

Ready to take action?

Tell us what you need – we will get back to you quickly with a technical assessment and an action plan.

  • Reply within 48–72 hours
  • No-obligation conversation
  • NDA confidentiality