The Dutch Data Protection Authority Says It Is Difficult to Comply With Data Minimization if a Generative AI System Has Access to All of an Organization's Data

On July 13, 2026, the Netherlands' data protection authority, the Autoriteit Persoonsgegevens or AP, published a practical tool for the deployment of generative AI and the GDPR. Its page for the tool says it is "primarily intended for organisations that wish to grant generative AI applications access to existing files or data within your organisation." The tool consists of three chapters of steps, and each step is accompanied by statements. Parts of it follow, not the whole.

What the Tool Covers, and What It Leaves Out

The authority recommends that organizations "explore whether it is possible to deploy generative AI without personal data." Where personal data is processed, the General Data Protection Regulation (GDPR) applies, and the tool sets out the steps needed before an organization can start using generative AI.

The tool "deals exclusively with the deployment of generative AI" and, it says, is not suitable for organizations that develop generative AI models themselves. It "does not address the requirements of other relevant legislation, such as the AI Act." It is "not exhaustive and is not a substitute for a Data Protection Impact Assessment (DPIA)," and simply completing it "is therefore not sufficient to comply with the GDPR." The tool is based on an opinion the AP gave after a prior consultation. Because that consultation does not address the legal basis for processing personal data, the tool does not address it either.

An Explanation for Every Statement

For each statement, the tool asks an organization to check whether the step has been completed and to provide an explanation. "Providing an explanation is necessary to fulfil your accountability obligations," the tool says. It asks that privacy professionals, IT specialists and representatives of the units that want to use generative AI complete it, and that a Data Protection Officer, where an organization has one, be involved in the entire process at an early stage.

The box is the easy part. The explanation beside it is what a later reviewer could test against what the system was actually allowed to see.

Access, Storage and the Supplier

Before deployment, the tool asks whether an organization has an authorization policy and puts it into practice. The reason it gives is data minimization: "not every employee should have access to all (personal) data within your organisation, including through the generative AI system." It calls that step particularly important where a generative AI system is granted access to the same files as the user.

On choosing a system, the tool says some integrated generative AI systems have access to all files that the user can also access. "If a generative AI system has access to all data within your organisation, it is difficult to comply with the principle of data minimisation," it says. Data minimization, in its words, "can be ensured by determining, prior to processing, to which (personal) data the generative AI system has access," for example by blocking access to files or applications such as email or chat. An organization can also specify which categories of personal data users may enter into the system.

The tool also asks for agreements with the supplier and other relevant parties, among them on whether those parties may use the data an organization shares to train their own model. Where a supplier gives no clarity about how long data is kept, or no way to adapt it, the organization "remains obliged to comply with the GDPR."

Answers, Decisions and Human Intervention

On accuracy, the tool says a generative AI system "does not always provide the right answer," because it "does not 'reason' with knowledge of the world, but predicts which words fit best in a given context." An incorrect answer "may therefore appear logical or convincing at first glance." It asks organizations to make employees aware of this so that they do not blindly rely on the output, and says it is important that users have sufficient knowledge and time to assess the output critically.

Where generative AI is used in processes that decide about individual people, the tool says this can lead to automated decision-making. Automated decision-making without meaningful human intervention, it says, "is in principle prohibited, unless there is an exception." It asks that users be authorized to deviate from the output, and adds a sentence about time: "The less time an employee has, the less meaningful the human intervention will be."

That sentence ties oversight to the clock, not only to a name. A review squeezed into a few minutes can still leave a signature on the file.