The dilemma: confidentiality meets the strictest data category in the GDPR
No industry faces a sharper legal backdrop on the subject of AI than medicine. Two sets of rules apply at the same time — and both in their strictest form:
- §203 StGB — the medical duty of confidentiality: A doctor or practice staff member who discloses someone else's secret without authorization is liable to prosecution. Entering patient data into an AI service that is not contractually bound to secrecy as an „other participant“ under §203 para. 3 StGB can be exactly such a disclosure.
- Art. 9 GDPR — health data: Data concerning health belong to the special categories of personal data — the most strongly protected category in the entire regulation. Their processing is prohibited in principle and permitted only under narrow exceptions. A standard data processing agreement (AVV) is not enough here; the bar lies considerably higher than for „ordinary“ personal data.
For tax advisors or lawyers, the AI question can be solved with a §203 agreement and a data processing agreement (AVV) under Art. 28 GDPR. For medical practices, the Art. 9 hurdle comes on top — and with today's cloud AI technology, it cannot be cleared in a serious way. Anyone who tells you otherwise is selling you a risk.
The red line: what does not belong in any cloud AI today
Before we talk about use cases, the boundary — clear and with no fine print:
Identifiable patient data does not belong in any cloud AI in 2026. That includes findings with names, doctor's letters with patient references, diagnoses with dates of birth, lab values with insurance numbers — anything that makes a person identifiable and concerns their health. This applies to ChatGPT, to Microsoft Copilot, to every cloud provider. And it applies to ClapNClaw too: in the current phase we deliberately do not process health data under Art. 9 GDPR. This honesty is uncomfortable — but it is the state of things.
Why this severity? Because with Art. 9 data, the combination of criminal-liability risk (§203 StGB), fine risk (Art. 83 GDPR: up to €20 million or 4 % of turnover) and damage to patient trust is out of all proportion to the benefit. The good news: a large share of the practice work that eats up time has no patient reference at all. That is exactly where safe AI use begins.
The 5 safe use cases: AI practice administration without patient data
1. Appointment management: templates for recall, cancellations and reminders
The AI creates and maintains the text modules of your appointment organization: recall letters for preventive check-ups, friendly cancellation and rescheduling texts, reminder templates for SMS and email, waiting-list notifications. All of them generic templates with placeholders — the names are only inserted in your practice software, not in the AI.
Practical example: Your medical assistant asks the AI for a recall template set for skin cancer screening: a letter, an email, a short SMS version — each in a friendly tone, with placeholders for name and preferred appointment. Two minutes later, three coordinated templates are ready that would previously have cost an hour of wording work.
2. Patient communication: practice FAQs, information texts, website
What does a private medical service (IGeL) cost? How do I prepare for the colonoscopy? What do I need to bring to my first appointment? The AI formulates generally understandable answers to recurring patient questions — for notices, the practice website, info sheets or the answering-machine text. Not a single patient record is needed for this; the medical sign-off naturally remains with the doctor.
Practical example: The practice is switching to online appointment booking. The AI drafts the patient information for it: a step-by-step guide in plain language, an FAQ list with eight typical questions and a short notice for the waiting room — on request additionally in English and Turkish.
3. Referral drafts: structuring them anonymized
This is where it gets trickier — and therefore precise: the AI can help with structuring and wording referral and inquiry texts when the input is fully anonymized. That means: no names, no dates of birth, no insurance numbers, no combinations of attributes that make a person identifiable. The abstract case („patient, male, mid-50s, with the following symptoms“) becomes a cleanly structured draft — you add the identifying details only in the practice system. Important: anonymization is more demanding than it sounds. With rare clinical pictures or small towns, even a text without names can be re-identifiable. When in doubt: do not enter it.
Practical example: A referral to a cardiologist. The doctor dictates the anonymized facts to the AI in keywords; it structures them into a clear draft with the clinical question, the diagnostics so far and the medication as a placeholder list. Name, date of birth and insurance details are only added in the practice software — the AI never saw them.
4. Document classification: sorting internal, non-identifiable records
Every practice collects documents with no patient reference: supplier invoices, maintenance contracts for medical technology, quality-management documents, hygiene plans, circulars from the Association of Statutory Health Insurance Physicians (KV), training materials. The AI classifies these holdings, summarizes long circulars and answers questions about them („When does the maintenance contract for the ultrasound device expire?“). Findings and patient files stay out of it — only what makes no person identifiable is classified.
Practical example: 40 pages of KV circular at the quarter change. The AI delivers a half-page summary with the three points that actually affect your practice — including references for follow-up reading. Reading time: two minutes instead of forty.
5. Team organization: emails, duty rosters, minutes
The underrated classic: the administrative work behind the practice. Draft replies to supplier and authority emails, the draft of a job ad for the new medical assistant, the minutes of the team meeting from bullet points, the checklist for onboarding. Here AI can be used immediately and without any gray area — these are the same office tasks as in any other business.
Practical example: After the team meeting, the practice manager types seven bullet points into the chat. The AI turns them into structured minutes with responsibilities and deadlines — and drafts the two emails that follow from the decisions right away.
Checklist: introducing AI in the practice — without data protection risk
- Define the red line in writing: No identifiable patient data into the AI — as a short, clear work instruction for the entire team, not as an assumption in individual heads.
- Vet the provider: Is a data processing agreement (AVV) under Art. 28 GDPR in place? Processing exclusively in the EU? No use of your inputs for AI training? A confidentiality obligation under §203 StGB? Without these four points: hands off.
- Start with use case 5: Team organization and emails are risk-free and show the benefit fastest. After that, templates (1) and patient information (2).
- Train the team: A short briefing on what is allowed and what is not — including the note that even „just quickly summarizing the findings“ is off-limits.
- Involve the data protection officer: Update the record of processing activities, and if needed assess whether a data protection impact assessment is required. That costs one appointment and creates legal certainty.
- Readjust regularly: Review quarterly which tasks can be added — and whether all uses still stay on this side of the red line.
Outlook: when is AI coming for patient data?
The honest answer: the industry is working on it, but today it is not there yet. The technically most plausible direction is locally operated language models — AI that runs entirely on hardware in the practice or in your own data center, so that health data never leaves the building. Smaller open models grow more capable every year, and for narrowly defined tasks such as documentation support, this path is foreseeable. Until then the rule stands: anyone selling you cloud AI today for findings and patient files is not solving the Art. 9 problem — they are pushing it onto your desk.
You will find the full legal framework — GDPR, §203 StGB, the data processing agreement (AVV) and the requirements per profession — in our complete GDPR AI guide 2026.
Frequently asked questions about AI in the medical practice
Is ChatGPT allowed for doctors?
For tasks without any patient reference — such as templates, practice FAQs or team organization — the use of AI is generally possible, provided GDPR basics like a data processing agreement (AVV) are met. Identifiable patient data, however, may not be entered into any standard cloud tool such as ChatGPT: doing so violates the medical duty of confidentiality under §203 StGB and processes health data under Art. 9 GDPR without a viable legal basis.
May AI process patient data?
Health data are, under Art. 9 GDPR, special categories of personal data and are subject to the strictest requirements in the entire regulation. On top of that comes criminal liability under §203 StGB for disclosure without authorization. In practice this means: identifiable patient data does not belong in any cloud AI today — regardless of the provider. Safe AI use in the practice therefore starts with tasks that have no patient reference.
Does ClapNClaw process patient data?
No. In its current phase, ClapNClaw deliberately does not process health data under Art. 9 GDPR. Its use in the medical practice focuses on practice administration, communication templates, anonymized drafts and team organization — all without identifiable patient data. We communicate this boundary openly, because it is the current state of safe AI use in healthcare.
Which AI tasks are safe in the medical practice today?
Safe are tasks without identifiable patient data: templates for appointment management (recall, cancellations, reminders), generic patient communication such as practice FAQs and information texts, anonymized referral drafts without names and dates of birth, the classification of internal non-identifiable documents, as well as emails and organization within the practice team. The prerequisite remains a GDPR-compliant provider with a data processing agreement (AVV) under Art. 28.
AI for your practice administration — with clear boundaries
Claude on your own server in Frankfurt. A data processing agreement (AVV) under Art. 28 GDPR and a §203 confidentiality agreement from day 1 — for practice work without patient data. Setup in 3 minutes, 14 days free to test.
Learn more for medical practices →Note: This article serves general information purposes and does not constitute legal advice or medical advice. The legal assessment of AI use in your practice — in particular with regard to §203 StGB, Art. 9 GDPR, the model professional code (Musterberufsordnung) and state-law requirements — depends on the individual case. Before introducing AI tools, consult your data protection officer and, if needed, a law firm specialized in medical law. As of: June 2026.