AI visit notes fill 50 fields in 35 seconds in Moscow, while private clinics remain silent

During an appointment, a doctor spends half the time looking at a monitor. Time-and-motion studies of primary care physicians show that documentation takes up 47% of the time allocated to a patient. It sounds like an ideal task for AI, and Moscow has already solved it. The question is why the same system has not appeared elsewhere.

Three jobs, not one

Filling out a visit note may look like a single action, but it actually consists of three separate tasks. The first is transcribing the conversation between doctor and patient—in other words, speech recognition. The second is mapping the text to fields, which requires a language model capable of extracting complaints, medical history, prescriptions, and a preliminary diagnosis. The third is transferring the completed fields into the clinic's medical information system. That is more integration than AI.

If only the first step is completed, the doctor receives five pages of transcript instead of a blank visit note. If the first two are completed but the data must be transferred manually, copying will consume the savings. All three stages must work, with the doctor reviewing and signing the result at the end.

What already exists

Since September 9, a voice assistant developed by the Moscow Department of Health and Yandex has been available to all primary care physicians at the city's adult outpatient clinics—more than 3,000 people. SpeechKit recognizes speech, YandexGPT fills more than 50 fields, and the data is routed through EMIAS. According to the participants, the visit note is assembled in 35 seconds, accuracy exceeds 90%, and documentation time falls by up to 40%. One important caveat: all these figures come from the participants themselves; there has been no independent evaluation.

Yandex sells the same solution as Yandex Med, but does not disclose integrations with private medical information systems or its pricing. ITMO University students built a prototype using GigaAM, Pyannote, and Mistral, but it never reached the clinical record system.

Where the catch lies

The author personally tested speech recognition on her lecture about insurance. Whisper medium transcribed 10 minutes in roughly 13 minutes and confused “CASCO” with “Avkazska.” A lightweight 312 MB version of GigaAM-v3 completed the task in 25 seconds on a standard CPU and handled the terminology more accurately. But both mangled “underwriting and ratemaking.” Medicine contains more—and more complex—terminology, so testing must use recordings from the specific clinic and specialty.

Then come the organizational barriers. Doctors may see appointment recording as surveillance rather than assistance. A clinic may also lose the leverage of saying, “I am writing this for the insurer.” In the author's view, integration is no longer the main bottleneck: robotic process automation, or RPA, can fill on-screen forms for a person, as demonstrated by ProMed and Lexema-RPA. This is quick for a pilot but fragile across a network of hundreds of doctors, because the robot breaks whenever the interface is updated.

How much it costs

An in-house pilot for one clinic and two or three specialties costs RUB 1.9–3.1 million; through a contractor, RUB 3–6 million. A server with a GPU costs RUB 0.5–0.8 million a year to rent or RUB 1–2.5 million to purchase. Including support, the first year costs RUB 4.5–6 million when built in-house. The system pays for itself when used by 12 to 55 doctors. Savings per doctor are estimated at RUB 205,000–270,000 a year under mandatory health insurance appointment standards, and higher for longer private appointments.

What to do

Start with one clinic and two or three specialties. First, fine-tune recognition for medical vocabulary, implement all three stages from the outset, and measure before and after: documentation time, the share of fields corrected by doctors, and errors in terminology and codes. The tool should be given to doctors only once quality is already high: one bad experience can quickly become the opinion of the entire team.