AI for medical practices: dictation, correspondence, triage – what is legally allowed and what is not

What the industry does – and where AI fits

About 17,500 Swiss physicians run their own practice or shared office (FMH physician statistics 2024). General practices, specialist offices, group practices and hospital-affiliated outpatient setups all share the same backbone: anamnesis, examination, diagnosis, treatment, documentation, billing. Documentation and billing consume between 30 and 50 percent of physician time depending on the study – time not spent with patients.

Practice AI does not mean AI-diagnosis. That is a high-risk area under EU AI Act Art. 6 and MDR. Practice AI as of May 2026 is sober and concrete: speech recognition for consultation notes (Whisper, OpenAI gpt-4o-transcribe, local models), automated letters to health insurers, FAQ bots for phone triage (administrative, not diagnostic), template generation for reports, OCR for incoming results. These are administrative writing aids that do not replace a trained physician – they free time for the patient.

Swiss practice-software is fragmented: tomedo, Vitomed, Achilles, Triamed, open-source Elexis, Medi-PC. Most vendors have shipped a first Whisper integration in 2025/2026. Practices unwilling to change their KIS can layer AI on top as a separate component: a revDSG-compliant dictation frontend that passes the finished text via copy-paste or API into the existing system.

Why it matters now

Swiss physician shortage is real: the OBSAN 2025 study estimates an additional 5,000 to 6,000 GPs needed by 2030, especially outside urban centres. Those who stay in practice will fight for every hour not spent at the desk. Two observations from 2024/2025 matter.

First: dictation AI has become productive in the German-speaking market. Players like Nuance/Microsoft DAX Copilot, Doc2Data, Dictat.ai and open-source Whisper forks (whisper.cpp, Faster-Whisper) achieve sub-5-percent word error rates on medical German – comparable to human typing. A 12-minute consultation now produces a complete record in 4 to 6 minutes of post-editing, down from the usual 15 minutes of typing.

Second: patient data is highly protected personal data under Art. 5 lit. c of the revised DPA. Consequence: a data-processing agreement is mandatory, cross-border data transfer requires a TIA, and the EU AI Act classification must be documented. Cloud dictation with US servers is not lawful without TIA and SCCs. The pragmatic path: EU hosting (Frankfurt, Zurich) or on-prem Whisper.cpp on a local GPU box.

FMH has stated several times in 2024 and 2025 (Schweizerische Ärztezeitung) that physician ultimate responsibility for AI-supported acts remains with the physician – medically and under FMH professional code Art. 3. That applies in particular to AI-generated letters and insurer correspondence: the physician must read, correct and approve before sending.

How AI actually fits a practice

Three workflows cover 80 percent of practice value. All three are administrative, not diagnostic – and therefore outside the EU AI Act high-risk class, as long as no triage or treatment advice is given to the patient.

Consultation dictation to structured record. The physician dictates either live (microphone at the exam table) or right after the consultation. Whisper transcribes. A downstream language model (Claude Sonnet, Mistral Large, local Llama 3.3) shapes the free text into the KIS structure: anamnesis, findings, diagnosis, procedere. The physician reviews, then the workflow passes the text into tomedo/Vitomed/Triamed. See n8n-workflow-automation for the glue logic.

Insurer correspondence. A health insurer query (cost coverage, medical certificate, clarification of a TARMED/TARDOC position) is received by email or letter. The AI reads the query, searches the patient file (RAG via local vector database – see retrieval-augmented-generation) for the relevant entries and drafts the reply. The physician edits, signs and sends. Time saved: 60 to 80 percent per reply.

Phone triage (NOT diagnostic!). A voice bot answers calls outside office hours, handles administrative questions (opening hours, emergency number, appointment booking) and prioritises incoming callback requests. For medical complaints the bot gives NO recommendation – it only records name, phone and the words "urgent / not urgent" and triggers the callback workflow. See voice-agent-telefon.

There are use-cases with clearer high-risk classification too: AI-supported image analysis (X-ray, skin lesions, ECG) is a medical device, falls under MDR and EU AI Act Art. 6 Annex III. That is the territory of CE-certified software products, not of a practice integration project. We explicitly leave that class outside our scope.

6 steps to productive AI dictation in a practice

1. 01Practice audit: inventory the KIS (tomedo/Vitomed/Triamed/Achilles), map interfaces, document data flows, capture revDSG register and EU AI Act classification.
2. 02Lock data residency: EU/CH-only via Mistral Frankfurt, OpenAI EU region or on-prem Whisper. Sign the data-processing agreement, archive the transfer impact analysis.
3. 03Define pilot scope: 2 physicians, 4 weeks, one documentation type (e.g. GP consultation). Success metric: minutes per record before vs. after, plus correction rate per record.
4. 04Roll out dictation hardware: lavalier mic or headset per pilot physician, web UI or mobile app, a clearly documented off switch for sensitive conversations.
5. 05Enforce the release workflow technically: every record and generated letter enters the KIS in "draft" – release only with physician signature (eID, BAG HIN login, or click with authentication).
6. 06Activate the audit trail: every AI action (dictation capture, model call, release click) is logged with timestamp, user, input-text hash and output-text hash. Retention 10 years, aligned with the patient record under KVG.

When AI makes sense in a practice

Pragmatic entry path: audit first, then pilot, then managed. The audit (ai-readiness-audit) maps practice software, workflow and revDSG context in 2 to 3 days. Output: three prioritised use-cases with effort and risk estimates.

A pilot usually runs on one of the three workflows above – most often dictation, because time saving is immediate and the legal case is clearest. Pilot duration: 4 to 8 weeks, one to two physicians, one documentation set, a clear metric (minutes per record, correction rate). If the pilot succeeds, operations move to the managed phase (managed-service-monitoring): server monitoring, model updates, audit trail, revDSG processing register.

AI fits when (a) the practice has at least 3 physicians or 1,500 quarterly consultations – below that the setup does not pay back; (b) documentation time is a measurable burden; (c) the practice software has an interface (API, HL7, copy-paste). Solo practices on paper need digitisation first, AI later.

Where AI must NOT be used in a practice

There are four red lines to communicate in every practice project.

First: AI does not replace a medical decision. Diagnosis, therapy and patient triage remain physician tasks (FMH professional code Art. 3, MedBG Art. 40). A practice AI setup must not give the patient a medical recommendation – not even "likely harmless" or "please go to the ER". Those sentences are reserved for the physician.

Second: patient data does not go into opaque cloud models without a data-processing agreement and without EU/CH region. Concretely: free ChatGPT in a browser is unlawful for patient content – no DPA, no deletion right, US servers. Lawful: OpenAI Enterprise with EU region and signed DPA, Anthropic via EU endpoint, Mistral La Plateforme in France, or on-prem Llama 3.3 / Mistral.

Third: AI image analysis as a practice integration. Running skin lesions, X-rays or fundus images through your own AI setup operates an uncertified medical device (MDR breach) and sits in the high-risk band (EU AI Act Art. 6). These products must be obtained as CE-certified third-party software – not as a homemade pipeline.

Fourth: fully automated insurer correspondence without physician release. Even if the AI drafts 95 percent correctly, every letter remains a medical statement with liability consequences. The release step must be technically enforced in the workflow – no "send" without a physician click.

Trade-offs

FAQ

Related topics

Sources

1. FMH Ärztestatistik 2024 – Demographie und Praxisverteilung in der Schweiz · 2024-12
2. Schweizerische Ärztezeitung – KI in der Arztpraxis: Verantwortung bleibt ärztlich · 2025-09
3. OBSAN Bulletin 2025 – Versorgungsbedarf Hausärzte Schweiz 2030 · 2025-04
4. EDÖB – Leitfaden: Bearbeitung besonders schützenswerter Personendaten unter dem revDSG · 2025-11
5. Europäische Kommission – EU AI Act, Hochrisiko-Systeme Annex III (Medizin) · 2024-07