In short
- Good candidates are administrative steps: appointment and recall requests, sorting incoming documents, drafting referral and clinical letters from dictation, preparing forms and invoices for insurers.
- Diagnosis, triage decisions, prescribing and every clinical judgement stay with clinicians. We do not build software for medical purposes, which counts as a medical device.
- Health data is special category data (Art. 9 GDPR): processing in the EU or UK, a data processing agreement, a data protection impact assessment where required and clear access rights.
- Patients ask AI assistants whom to see. Practices with clear service pages, consistent profiles and register entries are easier to name, and the advertising rules for medicine still apply.
Typical workflows in a medical practice
These are typical processes, not client projects. In each, the AI prepares and a member of the team checks. Clinical decisions are never part of the automation.
1. Appointment requests and recalls
Requests arrive by telephone, email, web form and messaging. Reception reads each one, finds the right appointment type and checks the calendar.
- Input: a request by email, web form or messaging service.
- AI prepares: reads the reason in the patient’s own words, matches it to the appointment types the practice has defined, proposes slots and drafts the confirmation or the question about missing details. Messages that contain red-flag words the practice has defined go straight to a person. The AI never decides on urgency itself.
- Staff check: every flagged message and every request outside the rules. The practice decides which simple requests may be booked without a check.
- Output: a booking proposal or a draft reply in the practice’s wording.
2. Incoming documents and results
Hospital letters, laboratory results and reports arrive by post, email and national messaging systems. Someone sorts them, assigns them to the patient and passes them on for review.
- Input: scanned or electronic documents.
- AI prepares: identifies patient, document type, sender and date, matches the document to the record and suggests the routing with a one-line summary.
- Clinician checks: every result and letter is reviewed and actioned by a clinician. The AI never marks a result as normal or files it as “no action”.
- Output: the document in the right record, routed to the right person.
3. Letters from dictation
Referral letters, reports for insurers and discharge summaries are written after the consultation, often late in the day.
- Input: a dictated note or, where patients have been informed, a recorded consultation, plus the existing record.
- AI prepares: a structured draft in the practice’s template, with medication and dates taken from the record and uncertain passages marked.
- Clinician checks: the whole text, corrects it and signs.
- Output: the letter in the record and ready to send.
4. Forms and billing preparation
Insurer forms, private invoices and requests to payers repeat the same data from the record.
- Input: the record entries the clinician has released for billing.
- AI prepares: fills the form, checks it for completeness against the payer’s requirements and lists what is missing.
- Clinician and staff check: diagnoses and codes are chosen by the clinician, and an authorised member of staff submits.
- Output: a completed form or invoice as a draft.
What stays with clinicians, and what we do not do
- Diagnosis, triage and treatment: the AI makes no clinical assessment and gives patients no medical advice.
- The medical device line: software intended for diagnosis, treatment decisions or the triage of patients is regulated as a medical device. We build administrative tools, not software for medical purposes.
- Results and prescriptions: reviewed and signed by clinicians, always.
- Patient-facing chat: answers organisational questions such as opening hours or how to request a repeat prescription, says clearly that it is an AI and points to the practice’s phone line or the emergency number for anything that sounds clinical.
- Decisions about people: we do not build scoring or selection of patients or staff. Such systems are high-risk under Annex III of the AI Act, see risk categories.
Rules for health data
Confidentiality about patient information is the core duty everywhere. The table lists the main sources by country as an orientation.
| Country | Basis | What applies in essence |
|---|---|---|
| England | GMC, Confidentiality, NHS England, guidance on AI‑enabled ambient scribing and MHRA, software and AI as a medical device | Patient information is confidential. For tools that record or summarise consultations, the NHS guidance expects a clinical safety case, a data protection impact assessment and a check of whether the product counts as a medical device |
| Ireland | Medical Council, position statement on AI | Doctors remain accountable for clinical decisions. Where AI is used to record consultations, doctors must explain how the tool works and what happens to the patient’s data |
| Germany | § 203 StGB and § 630f BGB | Doctors and their assistants must keep patient secrets, and service providers must be bound to secrecy. The treatment must be documented in a patient record |
| Austria | § 54 ÄrzteG | Doctors and their auxiliary staff must keep confidential everything entrusted to them or learned in the course of their work |
Across the EU, health data is special category data under Art. 9 GDPR, and the UK GDPR is similar. New AI tools that process it usually trigger a data protection impact assessment (Art. 35), and Art. 28 requires a data processing agreement with every provider. Art. 22 limits decisions based solely on automated processing. Software with a medical purpose is a medical device under Regulation (EU) 2017/745, and AI‑based devices that need an assessment by a notified body are high-risk under the AI Act. Chatbots must identify themselves as AI (Art. 50). This overview is general information, not legal advice. Check the current texts with your professional body and your data protection officer.
Systems you probably use
UK practices typically work with EMIS Web, SystmOne or Vision, and use Accurx or Docman for messaging and documents. German practices use practice management systems such as CGM M1 PRO, medatixx or Medistar and the telematics infrastructure with KIM. In Austria, ELGA is the national health record. We connect through the interfaces these systems offer and work alongside them where none exists. The names are for orientation only and do not imply any partnership.
How AI assistants answer when patients search
Patients ask ChatGPT, Gemini or Google’s AI Overviews for “a GP accepting new patients near me” or “a dentist for anxious patients”. The assistant draws on the practice’s own pages, the Google Business Profile, directories, register entries, regulator reports and reviews. Health content is held to a higher standard of expertise and trust, so named clinicians, qualifications and sources count.
- A page for each service: who it is for, how it works, who provides it and what to bring.
- Named clinicians with their registration: only titles and qualifications that are actually held.
- The same details everywhere: name, address, opening hours and services identical on the website, the Google Business Profile and directories.
- Reviews answered without detail: replies never confirm that someone is a patient.
Our article SEO for dentists and doctors covers the details, including advertising rules.
Is it worth automating? Eight questions
- Does reception spend a large part of the day on appointment requests that follow the same pattern?
- Is incoming mail sorted and assigned to patients by hand?
- Do clinicians type letters late in the day that are mostly data from the record?
- Are insurer forms or invoices filled in again from the same entries?
- Do patients wait for a reply because messages pile up?
- Is there a written rule on which patient data may go into which tool?
- Is a clinician responsible for checking every result and letter before it counts?
- Can you name one task where a draft that needs checking would already save time?
Three or more yes answers usually point to a workflow worth testing. Questions 6 and 7 are the prerequisites.
Sources
- England: GMC, Confidentiality: good practice in handling patient information; NHS England, Guidance on the use of AI‑enabled ambient scribing products; MHRA, Software and AI as a medical device
- Ireland: Medical Council, Position statement on artificial intelligence
- Germany: § 203 StGB, § 630f BGB
- Austria: § 54 ÄrzteG
- Regulation (EU) 2016/679 (GDPR), Art. 9, 22, 28 and 35: EUR-Lex; Regulation (EU) 2017/745 (medical devices): EUR-Lex; Regulation (EU) 2024/1689 (AI Act), Art. 50 and Annex III: EUR-Lex