Skip to content
All articles
Industry and business

Private AI in healthcare: start with administration, not diagnosis

GDPR Article 9(1) prohibits the processing of data concerning health unless an exception applies. That sentence is why a healthcare AI project should begin with the documents nobody's treatment depends on.

BastionUpdated 3 min read

Healthcare organisations hold some of the most sensitive data there is, so where an AI system runs matters more here than in most sectors.

What GDPR Article 9 says about health data

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. So a healthcare AI use case needs its lawful basis identified before the architecture is chosen.

When the AI Act makes a healthcare system high-risk

The AI Act applies separately from GDPR. Two routes lead into its high-risk category, each with its own date.

RuleWhat it coversApplies fromSource
Article 6(1)A system that 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 assessment2 August 2028Regulation (EU) 2026/1744, European Commission
Article 6(2)The Annex III use cases2 December 2027Regulation (EU) 2026/1744, European Commission
The two routes into high-risk under the AI Act, with the Commission's current timeline

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”. Neither date changes because the server is local.

What to deploy first: administrative work

Deploy administrative and knowledge work first, not clinical decisions: finding information in internal procedures, summarising non-clinical documents, drafting correspondence and answering staff questions about process.

The risk profiles of an internal procedures assistant and of an AI system involved in a care decision are far apart. Treat them as one project and the procedures assistant waits for the checks that the care system needs.

Data minimisation still applies on your own server

Five questions still need an answer when the system runs inside the organisation:

  • 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?

What local processing changes, and what it does not

Local processing keeps information inside the organisation's own infrastructure, which reduces external processing dependencies and makes several GDPR questions easier to answer with evidence. It supplies no lawful basis under Article 9, and it does not make a clinical use case appropriate. Measure whether the first workload helps, and expand only when governance and controls are ready for the next step.

With Bastion

What Bastion changes

Bastion is a private AI system delivered as one sealed box that runs inside your building. One monthly fee covers the hardware, the model, the hardened operating system and support, and nothing your team types leaves the building.

Bastion keeps inference and retrieval inside one box, at the organisation's site or in a data centre. No prompt and no document goes to an outside AI service, so the list of processing parties to document gets shorter.

The lawful basis, the minimisation decisions and the clinical appropriateness remain the organisation's.

Questions this article answers

Does GDPR allow AI to process health data?
Article 9(1) prohibits the processing of data concerning health, and Article 9(2) lists the situations in which the prohibition does not apply, beginning with explicit consent. The lawful basis has to be identified before the architecture is chosen.
Which healthcare AI workloads are realistic first?
Administrative and knowledge work: finding information in internal procedures, summarising non-clinical documents, drafting correspondence and answering staff questions about process. Their risk profile is far from that of a care decision.
Does an on-premise system make healthcare AI compliant?
No. It keeps processing inside the organisation's own infrastructure, which reduces external processing dependencies and makes several questions easier to evidence. It supplies no lawful basis and does not make a clinical use case appropriate.

On the record

  1. 1
  2. 2
  3. 3
  4. 4

    European Commission

    AI Act. Regulatory framework for AI

    read 2026-09-22