Healthcare organisations get more AI proposals than almost any other sector, and approve fewer of them. The reason is usually correct caution applied at the wrong point: the legal question is asked about the model, when it should have been asked about the data and about what the software is for.
Under Swiss law, health data is sensitive personal data, and clinical staff are bound by professional secrecy. That does not prohibit AI. It decides which projects are straightforward and which need months of preparation, and the difference has little to do with how clever the model is. It holds for a 12-person group practice, a 60-person home care organisation and a 200-person private clinic alike; only the paperwork scales.
What does Swiss law actually require for AI on health data?
Health data is sensitive personal data under the revised Federal Act on Data Protection. That does not ban AI, but it raises the bar: a clear purpose, no more data than needed, transparency towards patients, and a data protection impact assessment before any processing likely to be high risk, which large-scale processing of sensitive data typically is. Professional secrecy under the Criminal Code applies on top.
Article 5 of the Federal Act on Data Protection lists data relating to health among sensitive personal data, and Article 22 names large-scale processing of sensitive data as a case where high risk arises and an impact assessment is required beforehand. The Federal Data Protection and Information Commissioner has stated that the Act applies directly to AI-supported processing, that providers and users must make the purpose, functionality and data sources of AI-based processing transparent, and that high-risk processing is permitted in principle when appropriate safeguards are in place.
Professional secrecy is the part technology projects underestimate. Article 321 of the Swiss Criminal Code makes it an offence for doctors, dentists, pharmacists, psychologists, nurses, physiotherapists and several other professions, and their assistants, to disclose confidential information learned in their work, unless the patient consents or the supervisory authority authorises it in writing. The data protection Act adds that processing may be handed to a provider only if no statutory or contractual duty of confidentiality prohibits it. How a particular cloud contract sits with professional secrecy is a question for a lawyer who knows healthcare. The practical point is that the answer must exist in writing before the first record is sent, not be reconstructed afterwards.
A general AI law is not imminent. The Federal Council has chosen to implement the Council of Europe's AI Convention with amendments that are sector-specific as far as possible, and according to the Federal Office of Communications a consultation draft is due by the end of 2026. Until then, data protection, professional secrecy and medical device law are the rules that apply.
Which AI projects clear the bar quickly?
Projects that avoid identifiable patient data, or keep it inside systems the organisation already controls. Rostering, capacity planning, supply forecasting and drafting administrative correspondence clear it quickly. Coding and billing support on pseudonymised data is feasible after an assessment. Triage and diagnostic support are slow because medical device rules may apply, and anything that sends patient records to an external model needs advice first.
| Project | Data involved | Usual verdict |
|---|---|---|
| Roster and capacity planning | Staff and operational data | Straightforward |
| Supply and consumable forecasting | Purchasing history | Straightforward |
| Appointment reminders and no-show prediction | Scheduling data, minimal patient detail | Straightforward with controls |
| Document and correspondence drafting | Administrative text, no patient identifiers | Straightforward with controls |
| Coding and billing support | Pseudonymised episode data | Feasible, needs assessment |
| Dictation and clinical note drafting | Identifiable clinical conversations | Feasible, needs assessment and patient information |
| Imaging or triage decision support | Identifiable clinical data | Slow: device rules may apply |
| Anything sending patient records to an external model | Identifiable clinical data | Stop and get advice first |
The pattern is consistent. Administrative and operational work carries most of the near-term value and almost none of the legal weight; clinical decision support carries the reverse. Organisations that start at the bottom of the table spend a year in review and deliver nothing. Organisations that start at the top fund the harder work with results.
For a 12-person group practice, the top of the table usually means appointment reminders and drafting routine letters. For a 200-person private clinic it means rostering across wards and forecasting consumables, where the savings are measurable in hours and stock. A home care organisation sits in between, with route and shift planning as the obvious first candidate.
When is AI software a medical device?
When its manufacturer intends it for a medical purpose for an individual patient, such as diagnosis, prevention, monitoring, prediction or treatment, and it does more than store, archive, simply search or transmit data. Software for scheduling, resource planning and billing is not a medical device. A triage or diagnostic recommendation for a specific patient usually is, and the Medical Devices Ordinance then applies.
Swissmedic's information sheet on medical device software sets out the test. Software is explicitly included in the definition of a medical device in the Medical Devices Ordinance. The manufacturer's intended purpose is decisive, regardless of whether the software runs in the cloud, on a computer or on a phone. Software for hospital resource planning, reimbursement and managing doctors' visits, and electronic records that simply replace paper, are not medical devices. Devices are classified from class I to class III, and the classification rule for software puts strict limits on what can stay in class I.
Two practical consequences follow for a practice or clinic. When buying a tool that makes patient-specific suggestions, ask the vendor for its intended purpose and its conformity marking, because medical devices placed on the Swiss market must carry one. And when staff use a general-purpose chatbot to suggest a diagnosis, nobody has assessed it for that purpose. Software a healthcare institution builds and uses only in-house is not exempt either: Swissmedic states that it must still meet the general safety and performance requirements.
For a healthtech startup selling into the EU, the EU AI Act adds a layer. Under Article 6, AI that is itself a medical device or a safety component of one, and that needs a third-party conformity assessment, is treated as high risk. The European Commission's AI Act page currently gives 2 August 2028 for the rules on AI in such regulated products. Separately, Annex III lists emergency healthcare patient triage systems among high-risk uses. Both timelines have already moved once, so check them again before planning around them.
Can we use a cloud AI service with patient data?
Sometimes, with conditions documented before the first record is sent: a contract that makes the provider a processor bound by confidentiality, a known processing location with adequate protection or safeguards, training on your data switched off, access logging, and an assessment of whether the processing is high risk. Because of professional secrecy, the answer for free consumer tools is no.
In practice the checklist has six items. A data processing agreement that covers sub-processors and confidentiality. The countries where data is stored and processed, including support access, since disclosure abroad needs a country with adequate protection or appropriate safeguards under the Act. Training on your data switched off in the contract, not only in a settings screen. Logs showing who accessed what. An impact assessment where the processing is large-scale or otherwise high risk. And a privacy notice that tells patients what is processed and where.
Dictation and note-drafting assistants are where many practices meet this first. They are genuinely useful and they process the most sensitive material a practice holds: the conversation itself. Treat them as an assessed project, inform patients, and keep the clinician responsible for the final note.
Reusing records to build or validate a model is a different question again. If the project generates knowledge about diseases from health-related personal data, it may fall under the Human Research Act, which does not apply to anonymised data but requires prior information and no objection for further use of coded data, and consent for uncoded data. Ask the competent ethics committee early; it is faster than asking late.
Why do pilots stall even when they are permitted?
Because the pilot was scoped as a demonstration rather than as a change to how work is done. Clinical and administrative staff have no spare capacity to run a parallel process next to the real one, so a pilot that adds a step is abandoned regardless of its accuracy, while a pilot that removes a step survives even when it is imperfect.
This is not specific to healthcare. It is the pattern described in why AI pilots stall in Swiss companies, amplified by workload. The question to ask of any proposal is whether it adds or removes a step, and the answer is usually visible in whether anyone has been asked to stop doing something.
The second stall is ownership. A model that drafts correspondence needs someone accountable for what it drafts, a review path when it is wrong, and a record of both. Without that, the first bad output ends the programme, and rightly so. Drafting tools that keep a person in charge of every output are compared in generative AI use cases for small businesses.
A realistic first year
Start with one administrative process that is measurably slow, where the data never leaves your control, and where a named person can say whether the output was good. Run it in parallel, keep the record, and decide at the end of a quarter with evidence rather than enthusiasm.
In parallel, not afterwards, document the data protection position for the class of projects you intend to pursue, not just the one in front of you. The assessment is reusable; doing it once per project is what makes each project feel prohibitively slow. The records, identifiers and access rights it depends on are covered in is your data ready for AI?, and the security baseline, which matters more once an AI tool can read across systems, in cybersecurity priorities for Swiss SMEs.
In the engagements we have run, the first administrative project takes a small practice weeks of part-time effort and a clinic of 100 to 300 people one to three months, most of it spent on the data position rather than the tool. These are typical observations, not a quote. The wider obligations under the revised FADP are in what Swiss companies actually have to do, and for practices whose next question is patient-facing rather than internal, how a new medical practice gets found in Switzerland covers that side.
Our AI and machine learning practice works with healthcare organisations on exactly this ordering: the data position first, one operational process second, clinical questions only once both are settled. It is slower to start and considerably faster to finish.
Sources
- Fedlex — Federal Act on Data Protection (FADP), SR 235.1, Articles 5, 9, 16 and 22
- Federal Data Protection and Information Commissioner — AI and data protection
- Fedlex — Swiss Criminal Code, SR 311.0, Article 321 on professional confidentiality
- Federal Office of Communications (OFCOM) — Artificial intelligence, the Federal Council's regulatory approach
- Swissmedic — Information sheet: Medical Device Software, BW630_30_007, version 3.0, April 2026
- Fedlex — Medical Devices Ordinance (MedDO), SR 812.213
- European Commission — AI Act: regulatory framework on AI, application timeline
- AI Act Service Desk, European Commission — Article 6: Classification rules for high-risk AI systems
- AI Act Service Desk, European Commission — Annex III: High-risk AI systems
- Fedlex — Human Research Act (HRA), SR 810.30, scope and further use of health-related data
Related reading
- The revised FADP: what Swiss companies actually have to do
- Why AI pilots stall in Swiss companies, and what gets past that
- How does a new medical practice get found in Switzerland?
- AI and machine learning consulting
- Healthcare consulting
Working on this right now?
Tell us where you are in two questions. A consulting partner reviews every enquiry within 2 business days.