EU AI Act for hospitals
For hospitals, the AI Act applies depending on the use. AI for doctor's letters, summaries or questions on guidelines mainly brings the duty to promote staff AI literacy (Art. 4). High-risk AI is AI assessed by a notified body as a medical device (from 2 August 2028), AI for triage in the emergency department and AI in personnel decisions (from 2 December 2027). Deployer obligations then apply, often a fundamental rights impact assessment too.
Which AI in a hospital is high-risk?
| Use | Classification | From |
|---|---|---|
| Drafts of doctor’s letters and findings, summaries, questions on guidelines, translations | not high-risk, as long as there is no medical intended purpose | Art. 4 already applies |
| AI as a medical device assessed by a notified body (MDR class IIa and above), such as image analysis or diagnostic suggestions | high-risk under Art. 6(1) and Annex I | 2 August 2028 |
| Triage of patients in the emergency department, classification of emergency calls | high-risk under Annex III point 5(d) | 2 December 2027 |
| Screening job applications, allocating shifts or tasks, monitoring performance | high-risk under Annex III point 4 | 2 December 2027 |
| Biometric identification, emotion recognition | high-risk under Annex III point 1; emotion recognition of staff is prohibited, except for medical or safety reasons (Art. 5) | prohibition since 2 February 2025, otherwise 2 December 2027 |
The deadlines were postponed by Regulation (EU) 2026/1744 (Digital Omnibus, in force since 27 July 2026). Whether software is a medical device depends on its intended purpose under the MDR. See the guide Is AI a medical device?.
What must hospitals already do today?
- AI literacy (Art. 4): Train medical staff, nursing, administration and IT who use AI, and document it. Templates in the guide AI literacy under Art. 4.
- Transparency (Art. 50, since 2 August 2026): A chatbot for patients, relatives or referring doctors must be recognisable as AI. Published AI texts on matters of public interest without editorial review are labelled. See Transparency obligations under Art. 50.
- Create an AI inventory: Which systems are running, for what, with which data, and how are they classified?
In addition, the GDPR, § 203 StGB (German criminal law on professional secrecy), § 393 SGB V for cloud services with health data, and co-determination by the works council or the employee representation body (MAV, in church-run organisations) apply.
What applies when you deploy high-risk AI?
As a deployer, under Art. 26 you must, among other things, use the system in line with the instructions for use, assign human oversight to competent people, check input data for suitability, monitor operation and report serious incidents, keep logs for at least six months, and inform staff and their representatives in advance. Publicly owned hospitals and private hospitals with a public healthcare mandate usually need a fundamental rights impact assessment for Annex III systems.
If your hospital builds its own AI application with a medical intended purpose, it is a manufacturer under the MDR and therefore also a provider under the AI Act. Special rules apply to in-house manufacturing in health institutions (Art. 5(5) MDR).
What changes for university hospitals?
AI developed and put into service solely for research and development does not fall under the regulation (Art. 2(6)). The boundary is use in patient care: if a research model is used for treatment decisions, the MDR and the AI Act apply.
How does this work with basebox?
basebox is a platform, not a high-risk system, and not a medical device: it serves general tasks such as drafting letters or searching guidelines. For tasks involving patient data, basebox delivers drafts and summaries with sources, no diagnosis and no therapy suggestion; every result is checked by the doctor or nurse. If your hospital builds its own medical device and uses models from basebox through it, your hospital is the manufacturer; for the technical documentation, basebox provides the information of a supplier: model register with version and licence, software bill of materials and ISO 27001 certificate. For the EU AI Act, you receive training, logs and the model register. More under AI for hospitals and Security.
Not legal advice: As of 11 October 2026. The boundary between the MDR and the AI Act is being clarified further by guidelines from the Commission and the Medical Device Coordination Group. Check uses with a clinical connection in advance with your legal department and your data protection officer.
Frequently asked questions
Is AI for doctor's letters high-risk AI?
As a rule, no. A draft that the doctor checks, changes and signs has no medical intended purpose within the meaning of the MDR and does not fall under Annex III. It may be different if the software itself suggests diagnoses or therapies.
Does the EU AI Act apply to research projects at a university hospital?
AI systems specifically developed and put into service solely for scientific research and development are exempt (Art. 2(6)). As soon as a system is used in patient care, the regulation applies.
Do church-run and private hospitals have to carry out a fundamental rights impact assessment?
If they deploy high-risk AI under Annex III, probably yes: Art. 27 also covers private bodies providing public services, and the recitals name healthcare as an example. Art. 27 does not apply to AI in medical devices (Annex I).
Can hospitals be fined?
Yes, private and non-profit operators under Art. 99 of the AI Act. Whether a publicly owned hospital falls under the German exemption for public bodies in § 17(2) KI-MIG (the German act on AI market surveillance) depends on its legal form and on state law.
Sources
- Regulation (EU) 2024/1689 (AI Act), Art. 2, 6, 26, 27, Annex I and III, EUR-Lex ↗
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), new Art. 113, EUR-Lex ↗
- Regulation (EU) 2017/745 (Medical Device Regulation, MDR), EUR-Lex ↗
- European Commission, AI Act Service Desk: Annex III ↗
- KI-MIG, § 17 (gesetze-im-internet.de, German) ↗
- Bundesnetzagentur: press release on the KI-MIG (29 July 2026, German) ↗
This article was created with the help of AI, based on the sources listed. Responsible for the content: basebox GmbH. Not legal advice, no liability. Research as of 15 October 2026.