AI phone assistant in a medical practice
An AI phone assistant answers calls, books appointments and notes prescription requests. In a medical practice, four levels apply: professional secrecy, data protection, AI disclosure and recording. What you need to check before you start, and why a data processing agreement alone is not enough.
In August 2025, consumer advocates in Germany called for callers to medical practices to be told in advance when they are talking to an AI assistant. Since 2 August 2026, the AI Act has required exactly that. It is not the only rule for AI on the practice phone. Providers promise a relieved practice team, and some advertise their product as “100% GDPR compliant”. But the practice remains responsible, and it is subject to a stricter framework than a trade business using the same software: even the fact that someone is a patient is covered by professional secrecy.
Key points at a glance
- An AI phone assistant is allowed in a medical practice. According to the German Medical Association and the National Association of Statutory Health Insurance Physicians, professional secrecy and data protection apply in full.
- The provider and all of its subcontractors must be bound to confidentiality. Otherwise you may yourself commit a criminal offence under Section 203 of the German Criminal Code (StGB) if a patient secret gets out there. A data processing agreement alone is not enough.
- You do not need callers’ consent for appointments and prescription requests. You do need it for recordings, appointment reminders by text message and any use of the data for the provider’s own purposes.
- Under the German supervisory authorities’ list, a data protection impact assessment is the rule for AI on the phone. Under current German law, this also triggers the obligation to appoint a data protection officer, regardless of the size of your team.
- The assistant must identify itself as AI at the start, on the phone with a spoken sentence. A beep is not enough.
- Appointments yes, medical assessment no: if the assistant assesses symptoms and urgency, it can become a medical device.
Can a medical practice let AI answer the phone?
Yes. No law prohibits practices from having calls answered by AI. Typical tasks are booking or rescheduling appointments, taking prescription and referral requests, answering questions about opening hours and passing on requests to the team.
In their joint guidance on professional secrecy and data protection in medical practices (October 2025), the German Medical Association (Bundesärztekammer) and the National Association of Statutory Health Insurance Physicians (KBV) make one thing clear, however: both apply in full when artificial intelligence is used. For the phone assistant, this means checking four levels, ideally in this order, and respecting one limit.
Professional secrecy: it starts with the first sentence
According to the German Medical Association and the KBV, medical confidentiality also covers the fact that a person is receiving treatment. Anyone who calls and says “I’m a patient of yours and need a repeat prescription” is therefore already disclosing a patient secret. A doctor or member of another healthcare profession who discloses such a secret without authorisation may commit a criminal offence (Section 203(1) no. 1 StGB).
Disclosure is permitted to service providers who contribute to the professional activity, to the extent necessary for their work (Section 203(3) sentence 2 StGB). The German Medical Association and the KBV expressly name telecommunications providers and providers of practice management systems as examples. An AI phone assistant belongs to this group. Two conditions come with it:
- Ensure confidentiality: you must make sure the provider is bound to secrecy. If you fail to do so and someone at the provider discloses a secret without authorisation, you may yourself commit a criminal offence (Section 203(4) sentence 2 no. 1 StGB). According to the German Medical Association and the KBV, you either bind the provider’s staff yourself or you contractually oblige the provider to do so for you. With a cloud service, only the second route is realistic, and it should be in writing.
- The whole chain: an AI phone assistant usually consists of several services: telephony, speech recognition, language model, speech output and storage. If your provider uses other companies for this, it must bind them to secrecy too, otherwise it may commit a criminal offence itself (Section 203(4) sentence 2 no. 2 StGB). Ask for the list of subcontractors and for confirmation that all of them are bound.
Is a data processing agreement enough?
No. The data processing agreement under Article 28 GDPR governs data protection. The German Medical Association and the KBV expressly state that it does not constitute authorisation within the meaning of Section 203 StGB. So you need both, ideally in one contract. If your provider’s contract says nothing about Section 203 StGB, the confidentiality obligation is missing. What else a data processing agreement must contain is explained by Dieter Datenschutz in the article on the data processing agreement.
Data protection: what works without consent and what doesn’t
Calls almost always involve health data: the reason for the appointment, the medication for the repeat prescription, the symptoms. These are special categories of personal data (Article 9 GDPR). The good news: for what is part of the treatment, you do not need consent. Article 9(2)(h) GDPR and Section 22(1) no. 1(b) of the German Federal Data Protection Act (BDSG) allow processing for treatment and for the management of healthcare services, provided the data is processed by or under the responsibility of persons subject to professional secrecy.
In June 2025, the German Data Protection Conference (DSK), the joint body of the German supervisory authorities, set out what this means for external appointment management and online booking. An AI phone assistant essentially does the same, just on the phone. In our view, its positions can therefore be applied:
| What the assistant does | Allowed? | What matters |
|---|---|---|
| Book an appointment: name, date of birth, treating doctor, type of appointment, callback number | yes, without consent | only the data needed for the specific appointment |
| Outsource appointment management to a service provider | yes, as processing on your behalf without consent | inform patients and name the service provider |
| Send all patient master data to the provider in advance | no | not necessary, the provider only gets what it needs for the appointment |
| Appointment reminder by text message or email | only with explicit consent | the practice must be able to prove the consent |
| Use call data for the provider’s own purposes, such as training | no | the provider acts only on your instructions |
| Keep appointment data after the appointment | no | delete shortly after the appointment, the calendar is not treatment documentation |
Three further points:
- Where the data is processed: “servers in Germany” does not answer the question if speech recognition or the language model run elsewhere. The DSK requires you to clarify before commissioning the service whether individual steps such as support, maintenance or administration take place in a country outside the EU. If so, you need a legal basis for this under Article 44 et seq. GDPR.
- Inform: the assistant belongs in the practice’s privacy notice, with the provider listed as a recipient, and in the record of processing activities. On the phone, layered information is sufficient: the European data protection authorities’ guidelines on transparency allow you to give the key information during the call and to refer to the privacy notice for the rest.
- Keep a human within reach: according to the DSK, there should always be another way to get an appointment besides digital booking. On the phone, this means forwarding to the team or a callback.
Impact assessment: the rule for AI on the phone
The German supervisory authorities have published a list of processing operations that always require a data protection impact assessment (DPIA) (Article 35(4) GDPR). It includes the use of artificial intelligence that controls interaction with data subjects. As a typical field of application, the list names customer support using AI, with the example of a system that interacts with customers in conversation and processes personal data using AI. In our view, a phone assistant that talks to patients and processes health data fits this description exactly.
So plan the impact assessment from the outset and ask the provider for supporting documents, for example on data flows, subcontractors and safeguards.
The AI disclosure on the phone
Since 2 August 2026, Article 50(1) of the AI Act has applied: AI systems that interact directly with people must be designed so that those people are informed that they are dealing with an AI. This does not apply where it is obvious to a reasonably well-informed person. The information must be provided clearly and unambiguously at the latest at the time of the first interaction and must be accessible (Article 50(5)).
In its guidelines of July 2026, the European Commission is specific about phone and voice assistants:
- A spoken sentence at the start, such as “This is an AI-powered assistant”, with reminders during longer calls where appropriate.
- A beep alone is not enough, nor is a vague reference to an “assistant” or a sentence that only appears in the terms and conditions or on the website.
- “Obvious” hardly ever applies on the practice phone. If a system is accessible to the general public, which includes older people, the provider cannot rely on the exception according to the guidelines. And a human-sounding voice argues against it being obvious.
Formally, the obligation lies with the provider, so with a ready-made product, with the company that makes it; the practice is the deployer. But since you usually configure the greeting yourself, in practice it is up to you to make sure the sentence stays in. More on Article 50 can be found in the article on AI labelling requirements.
As a deployer, the practice must also take measures to support the AI literacy of its team (Article 4 of the AI Act, as amended with effect from 27 July 2026). You do not have to guarantee a specific level, and you do not need a dedicated AI officer for this. For the practice team, it mainly means knowing what the assistant can and cannot do, reading transcripts correctly and spotting errors. What lean training looks like is explained in the article on AI literacy training.
Recordings and transcripts
Anyone who records the privately spoken word of another person without authorisation commits a criminal offence (Section 201(1) no. 1 StGB). If the phone assistant stores recordings of calls, you need the callers’ consent. Because the call contains health data, the consent must be explicit (Article 9(2)(a) GDPR). Active agreement is the safe option, such as a “yes” to the question of whether the call may be recorded. Under the GDPR, silence does not constitute consent (Recital 32).
Whether the brief technical processing of speech for speech recognition already counts as a recording within the meaning of Section 201 StGB has not been settled. The safe route: recordings off by default, and only a transcript or a summary for the practice team. These are health data too. Decide who can see them and when they are deleted.
Appointments yes, diagnosis no
Some providers advertise that their assistant recognises emergencies or urgency. According to the guidance of the European Medical Device Coordination Group (MDCG 2019-11, June 2025 version), software for booking appointments is not a medical device in itself. Additional modules, however, can be medical devices in their own right. If an assistant assesses symptoms to determine urgency, such a module may be involved. What counts is the intended purpose set by the manufacturer. And the AI Act classifies AI systems for triaging patients in emergency care as high-risk systems (Annex III, point 5(d)). Their obligations apply from 2 December 2027.
The simplest line for the practice: the assistant books appointments and records requests, it does not assess symptoms. A fixed safety rule belongs in it all the same. If someone mentions an emergency keyword such as chest pain or shortness of breath, the assistant immediately refers to the emergency number 112 and puts the caller through to the team. Outside opening hours, it gives the number of the German out-of-hours medical service, 116117. If a provider advertises triage, ask about the intended purpose and whether the product is labelled as a medical device.
What belongs in the greeting
The greeting combines the AI disclosure, the first layer of data protection information and, if calls are recorded, the request for consent. Everything else is in the privacy notice on the website.
Checking the provider: ten questions
Instead of a product recommendation: every provider should be able to answer these questions in writing before you sign.
| Question for the provider | Why |
|---|---|
| Do you and all your subcontractors commit in writing to confidentiality under Section 203 StGB? | otherwise criminal liability risk for the practice |
| Is there a data processing agreement with a complete list of subcontractors? | Article 28 GDPR |
| Where do telephony, speech recognition, language model and storage run, including support and maintenance? | transfers to third countries, Article 44 et seq. GDPR |
| Do you use call data for training or other purposes of your own? | must be excluded |
| Are recordings off by default, and when are transcripts and appointment data deleted? | Section 201 StGB, storage limitation |
| Is the AI disclosure built into the greeting? | Article 50 AI Act |
| What data do you receive from the practice management system? | only what is necessary, no master data kept in reserve |
| Does the system assess symptoms or urgency, and with what intended purpose? | boundary with medical devices |
| Can every caller be forwarded to the team? | alternative route, emergencies |
| What documents do you provide for the impact assessment? | Article 35 GDPR |
Six steps to an AI phone assistant
- Define the purpose: what should the assistant do (appointments, prescription requests, opening hours) and what should it explicitly not do (medical assessment, information on test results)?
- Check the provider: with the ten questions above, in writing.
- Sign the contracts: a data processing agreement and a confidentiality obligation under Section 203 StGB, including subcontractors.
- Document data protection: impact assessment, record of processing activities, privacy notice with the provider as a recipient. Involve your data protection officer early.
- Set up the greeting and emergency rules: AI disclosure, data protection notice, forwarding to the team, 112 and 116117. Check everything with test calls.
- Train the team and adjust: the team knows how to check transcripts and what to do when something goes wrong. Record in your AI policy what the assistant is used for.
The most common mistakes
- Taking “GDPR compliant” on the provider’s website as sufficient. The practice is responsible, and the GDPR does not govern professional secrecy.
- Signing only the data processing agreement. Without a confidentiality obligation under Section 203 StGB, the criminal law safeguard is missing.
- Choosing a vague AI disclosure. “Your digital assistant” does not make clear that an AI is speaking.
- Leaving recordings on by default. Without active consent, this is a risk under Section 201 StGB and the GDPR.
- Letting the assistant assess symptoms. That is no longer appointment booking.
- Offering no route to a human. The DSK expects another way to get an appointment, and not everyone can cope with AI on the phone.
How Klaus helps medical practices
In the free AI check-up, you add your practice’s AI tools to the AI inventory, Klaus classifies their risk and creates your AI policy. The free plan covers one person and one AI system, and the training is then for you yourself.
From the Basic plan onwards, Klaus checks the phone assistant and your other AI systems in the 6-phase check, trains the practice team and keeps you informed about changes in the law. For the data protection side, such as appointing a data protection officer, there is Klaus’s big brother Dieter Datenschutz.
Conclusion
An AI phone assistant can noticeably relieve the practice team, and it is allowed. The work comes before the launch: confidentiality under Section 203 StGB for the provider and the whole chain, a data processing agreement, an impact assessment and a proper privacy notice. On the phone itself, the rule is: the assistant says at the start that it is an AI, does not record without asking, does not assess symptoms and lets everyone through to a human. Then patients know where they stand. What the same questions look like in a law firm is covered in the first part of the series: AI in law firms.
Frequently asked questions
Is an AI phone assistant allowed in a medical practice?
+
Do I need patients’ consent?
+
Does the AI phone assistant have to say that it is an AI?
+
Is a data processing agreement with the provider enough?
+
May the AI phone assistant record calls?
+
Do I need a data protection impact assessment for the AI phone assistant?
+
May the assistant decide how urgent a request is?
+
Sources and further reading
- Gesetze im Internet: Section 203 StGB, violation of private secrets (in German) ↗
- Gesetze im Internet: Section 201 StGB, violation of the confidentiality of the spoken word (in German) ↗
- Gesetze im Internet: Section 22 BDSG, processing of special categories of personal data (in German) ↗
- Gesetze im Internet: Section 38 BDSG, data protection officers of private bodies (in German) ↗
- General Data Protection Regulation (EU) 2016/679, full text (gdpr-info.eu) ↗
- EUR-Lex: Regulation (EU) 2024/1689 (AI Act), consolidated version of 27 July 2026 ↗
- European Commission: Guidelines on the transparency obligations under Article 50 of the AI Act (July 2026) ↗
- German Medical Association and KBV: guidance on medical confidentiality, data protection and data processing in medical practices, October 2025 (PDF, in German) ↗
- German Data Protection Conference (DSK): data protection in appointment management by healthcare practices, position paper of 16 June 2025 (PDF, in German) ↗
- German Data Protection Conference (DSK): list of processing operations that require a DPIA (PDF, in German) ↗
- Article 29 Working Party: Guidelines on transparency (WP260 rev.01), endorsed by the European Data Protection Board (PDF) ↗
- European Commission: MDCG 2019-11 rev.1, qualification and classification of software as a medical device, June 2025 (PDF) ↗
- Deutsches Ärzteblatt: debate about AI assistants in medical practices, 28 August 2025 (in German) ↗
Ready for your AI check-up?
In 20 minutes, Klaus shows you which EU AI Act obligations apply to you. Honest, concrete, no panic. Start for free, no credit card required.

