AI in medical practices: what Swiss data protection allows
AI in medical and therapy practices: what is allowed
AI in medical and therapy practices in Switzerland is allowed. The limits are the data and professional secrecy. Patient information is health data, which the Data Protection Act treats as sensitive personal data. Doctors, physiotherapists, psychologists and their assistants are bound by professional secrecy under the Criminal Code. Admin work without patient data can run in an approved tool. Patient data in a free consumer chatbot does not belong there. Everything in between needs a processor contract that also covers confidentiality. This article is not legal advice. It is the map a practice manager needs before the next referral letter lands in a chat window.
It is written for the manager or owner of a practice with a handful to about twenty people: practice software for appointments and records, an email inbox, perhaps Microsoft 365 or Google Workspace. Someone on the front desk has already asked whether ChatGPT could draft the letters.
Why the question comes up now
Since 1 January 2026, outpatient medical treatment in Switzerland is billed under a new tariff system. TARDOC and the outpatient flat rates replaced TARMED, which had been in force since 2004. TARDOC has around 1,400 tariff positions, compared with more than 4,600 under TARMED (Federal Office of Public Health, in German). For a medical practice team that means new positions to learn, software to adapt and questions from patients about their bills. Therapy practices are not billed under TARDOC, but their paperwork around prescriptions and insurers is just as real.
In a 2025 survey by gfs.bern for the FMH, doctors in outpatient practices reported 54 minutes a day of administrative work driven by authorities and insurers (report, in German). That is their figure, measured before TARDOC started, not our survey.
With that much admin, pasting a report into a chatbot at six in the evening is tempting. “Everyone uses it anyway” puts patient data into accounts the practice has no contract for. “AI is banned here” moves the same chat to a private phone. What works is a short rule about which tool, which data and who checks, the same approach an allow-list of AI tools describes for any office.
Three rules that apply to every practice
- Data classHealth data is sensitive
The Data Protection Act lists data relating to health as sensitive personal data (Art. 5 let. c).
- SecrecyArt. 321 Criminal Code
Listed professions and their assistants must keep patient secrets.
- CloudA processor, not a free pass
Outsourcing is allowed only if no secrecy obligation forbids it (Art. 9 FADP).
The Data Protection Act treats health data as sensitive personal data, so security and disclosure rules are stricter than for an ordinary customer email. Large-scale processing may require a data protection impact assessment (Art. 22). Your data-protection adviser decides whether that applies.
Art. 321 of the Criminal Code adds professional secrecy for listed professions and their assistants (Canton of St. Gallen, in German), including the front-desk assistant. Not every therapy profession is listed. Health data stays sensitive for all of them.
Does every AI tool need a processing agreement?
Yes, once a service processes patient data on a vendor’s servers. The vendor is then a processor. The FDPIC sets out the conditions: the processor may only process data the way you yourself may, it must keep the data secure, and outsourcing must not breach a duty of confidentiality such as professional secrecy. For patient data, the FDPIC also counts IT providers among the assistants bound by professional secrecy (FDPIC).
The FMH puts it in practical terms: as soon as a third party is commissioned to process data, the practice concludes both a processing agreement and a confidentiality agreement, and cloud services are explicitly among the cases (FMH guide, in German).
A free consumer chatbot fails because the contract is between the employee and the vendor, not the practice. Business plans in Microsoft 365, Google Workspace or ChatGPT come with processing terms. Your data-protection adviser checks whether those terms fit professional secrecy, storage location and sub-processors before patient data goes in.
The same applies to dictation apps, transcription tools and anything that records a consultation. They process health data by definition. For email, the FDPIC says patient data sent by email must be secured, for example by encryption, unless the patient has consented after being fully informed of the risks. Secured channels such as HIN Mail exist for this.
What is fine, and what is not
Fine in an approved tool
- A job advert for a medical practice assistant or a physiotherapist
- Templates for appointment confirmations and reminders, without names
- A general patient information sheet rewritten in plain language or translated
- A summary of a public tariff document or a newsletter from your association
- Internal checklists for opening, closing or the weekly rota
Not in a free consumer chatbot
- A report, referral or lab result with a patient’s name
- An insurer’s letter about a patient, pasted in to get a reply
- A consultation recorded with an app on a private phone
- A “de-identified” case where date of birth, a rare diagnosis and the village still point to one person
The left side has one thing in common: no patient appears in it. That is the test for the front desk. If a patient could recognise themselves in the text, it belongs in the practice software or in a tool your data-protection adviser has cleared.
The general traffic light for personal data in ChatGPT, including what to do when someone has already pasted a file, is in ChatGPT and customer data. A practice adds the secrecy layer on top.
What better looks like in a practice
Better is one page, not a ban. It names the approved tools, says that patient data does not leave the practice software unless a cleared tool with the right contracts is involved, and names the person who decides on new tools. Check the short version: does your allow-list name which tools may see patient data, is there a processing agreement for every cloud service that touches patient data, and who decides when a staff member proposes a new tool? Also ask your practice software vendor which AI features it offers and on what contract, since the data is already there.
Where automation pays off in a practice is usually the admin around the software: sorting the inbox, reminders, tracking cost-approval requests. Much of it runs without patient data, or keeps them inside the practice software.
Before you sign anything that touches patient data, talk to your data-protection adviser or your professional association: for doctors the FMH or the cantonal medical society, for physiotherapists physioswiss.
If you want to know whether one recurring admin task in your practice can be automated without patient data leaving the practice software, that is what the Process Check is for. You describe the task. No patient data is needed.
Written by
Aurum Avis Labs
Builds and ships at Aurum Avis Labs. Writes here about what we learn working with founders and SMEs in the DACH region.
Related Articles
You might also be interested in these articles
Automate practice administration: requests, reports, approvals
How a Swiss practice can automate appointment requests, reminders, incoming reports and cost approvals while patient data stays in the practice software.
ChatGPT and customer data: what Swiss data protection means for SMEs
May staff paste customer data into ChatGPT? Public/internal/personal traffic lights, Free vs Business, FDPIC. Not legal advice.
Approve AI tools at work: an allow-list in five steps
How to approve AI tools at work: take stock, move to business accounts, check the Microsoft 365, Google Workspace and ChatGPT terms, announce the list.