AI in the medical practice: six applications and what they really deliver
Artificial intelligence has arrived in daily practice, but not everywhere it was promised. Six applications, each with what it takes off your desk today and what it leaves behind.
Artificial intelligence in medicine is usually discussed in the future tense, and usually about diagnostics. In the daily life of a Swiss practice, both are the wrong way round. AI is long since in use, and it almost never sits in the diagnosis but in the half hour after surgery hours, when reports get written and forms filled in.
That is the unspectacular truth behind the subject: the applications that carry weight today remove writing. The applications that make headlines have not yet arrived in daily practice. The six below are ordered by that measure, not by excitement.
1. Dictation that writes itself
Speech recognition is the one application whose benefit can be measured directly in minutes. A report that is dictated and written automatically costs a fraction of the time of typing, and today's systems recognise medical terminology reliably.
The limit lies with proper nouns and numbers. Patient names, medication doses and lab values are exactly the places where an error becomes expensive, and exactly the places where recognition is least certain. Reading it back stays mandatory, and anyone who accounts for that arrives at a realistic gain rather than a brochure.
2. The report that takes notes during the consultation
The next stage listens to the conversation between doctor and patient and proposes a draft report from it. This is the single largest time saving currently available: documentation is produced during the consultation instead of afterwards.
It is also the application with the highest hurdle. Recording happens in the room, and that requires the patient's explicit consent — health data is sensitive under Art. 5 FADP. Practices that solve this cleanly obtain consent when the appointment is made, not in the consulting room.
3. Coding that makes a suggestion
Proposing the right diagnostic codes from a documented consultation is a task language models are well suited to: plenty of text, a fixed vocabulary, a clear mapping. The suggestion arrives, a person confirms it.
The order matters. A system that codes and enters on its own shifts responsibility to a place that cannot carry it. A system that suggests and waits for confirmation still saves most of the work.
4. Answers to recurring questions
A large share of the calls into a practice is about the same ten things: opening hours, repeat prescriptions, where the report has got to, whether to come in fasting. These can be answered automatically, and it takes noticeable load off reception.
5. Scheduling and missed appointments
Systems that learn from past appointments can identify which ones are more likely to be missed and steer reminders accordingly. The effect is real but smaller than advertised — most no-shows are already prevented by an ordinary reminder two days ahead.
Before a practice invests in prediction here, the banal question is worth asking: is there a reminder at all, and through which channel? A text message that arrives beats any model.
6. Analysing large volumes of data
The area most written about: imaging data, lab series, registry analyses. It is happening, but in hospitals and research centres with their own data holdings, not in a practice with four consulting rooms.
For a practice the honest advice is to skip this point. The benefit of AI lies where typing happens today, not where research happens.
What is permitted in Switzerland
The legal question is the same for all six applications as for any other service processing patient data, and it can be answered without special rules: a data processing agreement under Art. 9 FADP is required, and it must be clear which law governs the provider. A US company falls under the CLOUD Act, even with its servers in Zurich.
The practical stumbling block is a different one: a freely accessible language model in a browser is not a tool for patient data. What is typed there leaves the practice, and nobody has signed anything about what happens to it. In daily work this line matters more than any subtlety of the statute.
The question is not whether AI is coming into the practice. It is long since here — just in different places than the marketing claims.
Where to start
Anyone starting today starts with documentation, because that is where the gain is largest and the change smallest. Dictation can be trialled in a week without altering anything else in the workflow, and the time saved is measurable after a few days rather than estimated.
Everything else follows from that. A practice that no longer types its reports sees for itself which step is the next one in the way .
