GDPR-compliant with AI: Privacy checklist for Dutch organizations

The integration of AI (such as Large Language Models) offers unprecedented opportunities for Dutch companies, from automating customer service to rapid data analysis. But feeding these models with company data immediately raises an important question: How do you remain compliant with the General Data Protection Regulation (GDPR)?

In this practical checklist, we walk through the key pillars for a responsible, privacy-friendly AI strategy.

1. What data are you allowed to process? (Data minimization)

The GDPR prescribes that you must have a legal basis for processing personal data, such as consent, performance of a contract, or legitimate interest. With AI, data minimization is crucial.

• Do not throw everything into the prompt: Train employees or systems to filter out personal data (names, citizen service numbers [BSNs], financial data) before text is sent to an AI model.
• Anonymize or pseudonymize: Use scripts (such as Named Entity Recognition) to replace sensitive entities with placeholders (e.g., [CUSTOMER_NAME]) before the model processes the data.
• Purpose limitation: Collected data may only be used for the purpose for which it was originally obtained. Reusing data to train your own AI is often not permitted without further justification.

2. The Data Processing Agreement (DPA)

Do you use APIs from external vendors (such as OpenAI, Anthropic, or Google)? Then they act as processors of your (customer) data.

• Exclude training data: Ensure that the contract explicitly states that your inputted data (prompts and context) is not used to train the provider's models. (This is now standard for many Enterprise APIs, but not for consumer versions).
• Zero-retention: Where possible, demand a zero-retention policy, where your data is deleted from the provider's servers immediately after inference (generating the output).
• Data location: Check where the servers are located. Within the EEA (European Economic Area) is preferred. If data goes to the US, check whether the provider falls under the EU-US Data Privacy Framework.

3. Local vs. Cloud Models (Privacy by Design)

From the design principle of Privacy by Design, the architecture of your AI solution determines your risk profile.

Cloud APIs

Easily scalable and very powerful, but your data leaves your network. This requires watertight DPAs and strict access controls.

Local (On-premise) Models

By running open-weight models (such as Llama 3 or Mistral) on your own hardware, you maintain full control. Data does not leave your servers. For organizations working with medical records, legal documents, or highly secured intellectual property, local AI hosting (or bare-metal hosting in a Dutch data center) is often the safest route to guarantee GDPR compliance.

4. DPIA (Data Protection Impact Assessment) Outline

A DPIA is mandatory when data processing entails a high privacy risk, which is often the case with the systematic deployment of AI. What should you include in this?

1. Description of the processing: How exactly does the AI tool work? What data goes in, what comes out?
2. Necessity and proportionality: Is AI really the only or best way to achieve this goal?
3. Risk assessment: What happens in the event of a data breach? What if the model hallucinates about a person and this leads to incorrect decision-making?
4. Mitigating measures: What technical (e.g., local hosting) and organizational (e.g., 'Human-in-the-loop' control) measures are you taking to cover these risks?