As soon as an AI project touches customer, employee or prospect data, the GDPR enters the room. And with it, often, a kind of paralysis: some teams give up on useful projects for fear of getting it wrong, others move ahead without asking the question and discover the problem during an audit or a complaint. Both attitudes are costly.
The good news is that the GDPR does not forbid AI. It requires asking the right questions at the right time. This article presents the checklist we go through with our clients before sending any personal data to a model. It does not replace advice from your DPO or a lawyer, but it lets you come to them with a solid file.
1. Is there really personal data?
The first question seems obvious, yet it is often poorly answered. Personal data is any information that can identify a person, directly or indirectly: a name, of course, but also a work email, a case number, a “job title + department + city” combination that points to a single person, or the free text of a support ticket.
In many projects, the honest answer is “yes, a little, without meaning to”. That is exactly where things are decided: a base of internal procedures that contains, here and there, employee names and examples drawn from real cases is not an anonymous base.
2. For what purpose, and on what legal basis?
Every processing operation needs a clear purpose and a legal basis. For a business AI project, this is most often performance of a contract or legitimate interest, more rarely consent. Legitimate interest is not a master key: it requires checking that the use stays proportionate and in line with what people can reasonably expect.
The classic pitfall is reuse. Data collected to manage customer contracts cannot automatically be used to train a sales scoring model. A new purpose must be compatible with the original one, or it needs its own justification.
The GDPR does not ask “are you allowed to use AI?”, but “are you allowed to use this data, for this?”.
3. Are we sending only what is necessary?
Data minimization is probably the most useful principle in practice, because it reduces both legal and technical risk. A model that classifies a support ticket does not need the customer’s name or address. An assistant that summarizes a contract does not need the bank details in the appendix.
- Exclude, from indexing onwards, the fields and documents the use case does not need.
- Pseudonymize before sending to the model when identity is not needed for the reasoning.
- Never send sensitive data (health, opinions, biometric data…) without a specific prior analysis.
4. Who processes the data, and under which contract?
As soon as a model provider processes personal data on your behalf, it acts as a processor under the GDPR. This requires a contract that sets out what it may do: the data processing agreement, or DPA. Major providers offer one for their business plans and APIs, generally not for consumer plans.
You also need to look at the whole chain: the tool vendor, the host, the model provider, the monitoring tool. Each is a processor or sub-processor and must appear in the documentation. The same reasoning applies to a service provider like us: a team working on your data must be covered by a contract, with clear commitments on confidentiality, security, and the return or deletion of data at the end of the engagement.
5. Does the data leave the European Union?
Processing outside the European Economic Area is not forbidden, but it must be framed. For the United States, the EU-US Data Privacy Framework allows transfers to certified companies; elsewhere, the European Commission’s standard contractual clauses are most often used. In every case, these transfers must be identified and documented.
In practice, the question applies to every link: where the model runs, where logs are stored, where the team that can access the data is located. Choosing a European hosting region simplifies things a lot, without removing the need to check the rest of the chain.
6. Is an impact assessment required?
A data protection impact assessment (DPIA) is mandatory when processing is likely to result in a high risk to individuals. This is frequently the case with AI: large-scale processing, evaluating or scoring people, combining sources, sensitive data, innovative use. The French data protection authority (CNIL) and other European regulators have published lists of such processing and recommendations specific to AI systems, which are a good starting point.
Even when it is not mandatory, a light risk analysis forces the right questions and leaves a written record of the choices made, which is valuable in case of an inspection.
7. Do individuals’ rights still apply?
Data subjects keep their rights: information, access, rectification, erasure, objection. You must therefore be able to answer a simple question: if a customer asks for their data to be deleted, do you know where it is in the AI system? In a RAG index, the answer is usually yes, provided you planned for deleting a document and its chunks. In a model fine-tuned on that data, it is much harder, which is one more reason to favor RAG for personal data.
Finally, when a decision with significant effects on a person (a credit refusal, a rejected application) relies solely on automated processing, the GDPR requires specific safeguards, including the right to human intervention. A human in the loop is then not a nice-to-have.
What about the AI Act?
The European AI regulation adds a layer of obligations that depends on the system’s risk level: bans for certain uses, stricter requirements for high-risk systems such as recruitment or credit scoring, transparency obligations for others. It does not replace the GDPR: both apply in parallel. We cover it in a dedicated article.
Our recommendation
Build these questions into the project scoping, not the go-live. Half a day of work with your DPO upfront avoids weeks of rework later, and turns the GDPR into an argument rather than a brake: an AI system for which you can explain exactly what data it uses, why, and where it goes is a system your customers and teams can trust.