Private AI in healthcare: start with the administration, not the diagnosis

Healthcare organisations hold some of the most sensitive material any organisation holds, which makes the architecture question sharper here than almost anywhere else. It also makes overclaiming more dangerous, so this article will not suggest that a local deployment solves the regulatory problem.

The rule that shapes everything else

GDPR Article 9(1) is a prohibition with exceptions, not a permission with conditions:

Processing of personal data revealing racial or ethnic origin, political opinions, religious or philosophical beliefs, or trade union membership, and the processing of genetic data, biometric data for the purpose of uniquely identifying a natural person, data concerning health or data concerning a natural person's sex life or sexual orientation shall be prohibited.

Regulation (EU) 2016/679, Article 9(1)

Article 9(2) then lists the situations in which the prohibition does not apply, beginning with explicit consent. The practical consequence is that a healthcare AI use case needs its lawful basis identified before the architecture is chosen, not after.

The classification question, separately

The AI Act adds its own layer. Article 6(1) makes a system high-risk where it is a safety component of, or is itself, a product covered by the Union harmonisation legislation in Annex I and subject to third-party conformity assessment. Article 6(2) adds the Annex III use cases.

Article 6(3) can take a system back out, where it does not pose a significant risk and performs, for example, “a narrow procedural task” or is “intended to improve the result of a previously completed human activity”.

The Commission's current timeline puts the Annex III rules at 2 December 2027 and product-embedded high-risk at 2 August 2028. Neither date changes because the server is local.

What to deploy first

Administrative and knowledge work, not clinical decisions. Finding information in internal procedures. Summarising non-clinical documents. Drafting correspondence. Answering staff questions about process.

The risk profile of an internal procedures assistant and an AI system involved in a care decision are not adjacent. Treating them as one project is the most common way a healthcare AI programme stalls.

Data minimisation does not become optional indoors

• Does the model need this data at all, or only a redacted version of it?
• Can access be restricted to the people whose work requires it?
• How long should prompts and logs exist, and who deletes them?
• Who can retrieve which documents, enforced where the retrieval happens?
• Can the system operate without internet access, and does anything in the workflow require it?

Local processing narrows the set of parties involved. It does not supply a lawful basis, and it does not make a clinical use case appropriate.

The honest summary

A private AI environment can process information inside the organisation's own infrastructure, which reduces external processing dependencies and makes several GDPR questions easier to answer with evidence.

It is not a shortcut around Article 9, and it is not a defence if the use case was inappropriate. Start with low-risk internal document work, measure whether it helps, and expand only when governance and controls are ready for the next step.