Can AI Process Patient Data in the Netherlands? The Rules

Can AI process patient data in the Netherlands?

AI may process patient data in the Netherlands under the same rules that govern any processing of health data, plus two of its own.

Yes, when the AI works inside the treatment relationship, on a documented legal basis, after a DPIA, under a processor agreement, in a system that meets NEN 7510, and with a clinician making the decision. No, when staff paste patient records into a public chatbot. This guide sets out which rule says what, in the order a healthcare provider meets them.

By Tom Joseph · Last updated: 14 September 2026

Which rule says what

Rule	What it requires for AI on patient data
GDPR (AVG) article 9	Health data is a special category: processing is prohibited unless an exception applies, and every processing needs a legal basis, a purpose and a retention period.
UAVG article 30	The Dutch exception: care providers, and insurers for specific tasks, may process health data as far as needed for treatment or care. This is the basis most clinical AI runs on; separate consent is usually not the basis.
WGBO, article 7:457 Civil Code	Medical confidentiality. Data leaves the treatment team only with the patient's consent or a legal duty; an AI supplier is inside the circle only under a processor agreement and only for the treatment purpose.
DPIA (AVG article 35)	Mandatory before AI on health data goes live; the Dutch DPA treats AI and large-scale health data as high risk, with hospitals always counted as large-scale and practices from about 10,000 patients.
NEN 7510, 7512, 7513	The Dutch information-security norms for healthcare: access control, secure exchange and logging of who saw which record. An AI system logs like any other system in the EPD.
EU AI Act	AI that is a medical device under the MDR, or that triages emergency care, is high-risk: documentation, human oversight, accuracy monitoring and a conformity route. The Article 4 AI-literacy duty already applies to every provider whose staff use AI.
EHDS, Regulation (EU) 2025/327	In force since 26 March 2025, applying from 26 March 2027 with phases to 2035: interoperable electronic health records and a regulated route for secondary use of health data, including for AI development.

Four common cases, judged

A clinician pastes a discharge letter into ChatGPT to summarise it. Not allowed. The provider of a public chatbot becomes a recipient without a basis or an agreement; the DPA counts this as a data breach, and it has to be reported.

Decision support inside the EPD that proposes a classification the clinician accepts, adjusts or overrides. Allowed, with a DPIA, a processor agreement, MDR and AI Act conformity, and logging. This is how Crux Digits built the AMAN AI platform, which has supported about 100,000 assessments since February 2025 with the clinician deciding every case.

Drafting discharge documentation from the record, for the doctor to sign. Allowed on the same conditions; the model runs in the provider's own tenancy, is not trained on the data, and every sentence traces to the source record.

Using patient records to train a model for a new product. Not covered by the treatment basis. It needs anonymisation that actually holds, or consent, or the EHDS secondary-use route once it applies; pseudonymisation alone keeps the data personal.

The six design decisions that make it lawful

Where the model runs. In the provider's own Azure or AWS tenancy in the EU, or on-premise. No US public API on identifiable data.

What the model keeps. Nothing. No training on patient data, no retention of prompts, a written statement from the supplier to that effect.

Who sees the output. The treating clinician, inside the treatment relationship; the system proposes, the person decides, and the log records both.

Minimisation. The model sees the fields the task needs and no more; pseudonymise where the task allows it, and say in the DPIA why it does not where it does not.

Explainability. Every output traces to the source data that produced it, so a clinician, an auditor or the patient can follow the reasoning.

Paper before code. Purpose, basis, DPIA, processor agreement and the MDR and AI Act classification are written before the first line is built, and updated when the model is retrained.

What compliant clinical AI costs to build

The compliance work is not a separate project; it is the shape of the build. Crux Digits scopes clinical AI in three fixed steps: a €2,500 audit that produces the intended use, the risk classification under the MDR and the AI Act, the data map and the DPIA outline; a €20,000 Production-ready MVP that runs on de-identified or in-tenancy data against an agreed acceptance criterion; and production from €50,000, integrated with the EPD, logged to NEN 7513 and monitored in service. Excl. VAT, all code and documentation handed over, so the provider can present the file to an inspector without us in the room. See AI for healthcare for the use cases and the AI Act checklist for the deadlines.

Frequently asked questions

Usually not as the legal basis. For treatment, care providers rely on UAVG article 30 and the treatment agreement under the WGBO; consent is the basis only when the processing goes beyond care, such as research or product development. Patients must still be informed, and a clinician must remain able to decide without the system.

Is a DPIA mandatory for AI on patient data?

Yes. The Dutch Data Protection Authority lists AI and large-scale processing of health data among the cases that require one, hospitals are always large-scale and practices from about 10,000 patients. The DPIA is written before go-live and revisited when the model or its use changes.

May we use ChatGPT, Claude or Gemini on patient records?

Not the public versions. Entering identifiable patient data into a public chatbot is a data breach: the provider becomes a recipient without a basis or an agreement. Enterprise deployments in your own EU tenancy, with a processor agreement and no training on your data, are a different product and can be used under the same rules as any other system.

Is clinical AI a medical device?

If it supports diagnosis, prognosis, treatment or monitoring, it is software as a medical device under the MDR, and as such a high-risk AI system under the AI Act. That means a conformity route, technical documentation, human oversight and monitoring in service. Administrative AI, such as summarising a letter for the doctor to sign, is generally not a device but still needs the DPIA and the processor agreement.

Can patient data be used to train an AI model?

Not on the treatment basis. Training needs anonymised data that cannot be re-identified, explicit consent, or the secondary-use route under the European Health Data Space once it applies from 2027. Pseudonymised data remains personal data under the GDPR.

What does Crux Digits do differently for healthcare?

The clinician-led maker-checker model: the system proposes with its reasoning, the clinician accepts, adjusts or overrides, and the log records both. AMAN AI has run this way at a national provider since February 2025, about 100,000 assessments in, with every output explicitly indicative. The DPIA, the classification and the processor agreement are deliverables of the audit, not afterthoughts.

Can AI process patient data in the Netherlands?

Which rule says what

Four common cases, judged

The six design decisions that make it lawful

What compliant clinical AI costs to build

Compare further

Frequently asked questions

Is a DPIA mandatory for AI on patient data?

May we use ChatGPT, Claude or Gemini on patient records?

Is clinical AI a medical device?

Can patient data be used to train an AI model?

What does Crux Digits do differently for healthcare?

Building AI on patient data and want it to pass the DPIA?

Can AI process patient data in the Netherlands?

Which rule says what

Four common cases, judged

The six design decisions that make it lawful

What compliant clinical AI costs to build

Compare further

Frequently asked questions

Is patient consent required before AI processes the data?

Is a DPIA mandatory for AI on patient data?

May we use ChatGPT, Claude or Gemini on patient records?

Is clinical AI a medical device?

Can patient data be used to train an AI model?

What does Crux Digits do differently for healthcare?

Building AI on patient data and want it to pass the DPIA?