Artificial intelligence in medicine is already built into radiologists’ workflow, supports clinical decisions, and processes medical documents — in some medical organizations this already works today. For a clinic or a diagnostic center, behind this stand specific processes: the status of software as a medical device, the vendor’s obligations to the regulator, and the unchanged place of the physician in making a diagnosis.
Below — three areas where AI is already used in clinical practice, the status of diagnostic software, the role of the attending physician, oversight by Roszdravnadzor (the Federal Service for Healthcare Supervision), and the 2026 AI law.
Where AI Already Works in Medicine
The use of AI in medicine is not limited to a single laboratory or pilot project — the technologies are already built into clinic workflows in three areas.
Diagnostic radiology. This is the most developed area: neural networks analyze computed tomography and magnetic resonance imaging scans, mammograms, and X-rays — they annotate images, highlight suspicious areas, and generate a preliminary description of the study, which a physician then checks. According to the Moscow Government as of March 2024, more than 12 million radiology studies had by then been described in the city with the help of neural networks, physicians had access to information from more than 50 computer vision services, and artificial intelligence detected signs of 37 diseases — from lung cancer to stroke and coronary heart disease. According to the same data, the technologies marked suspicious areas on images with color cues and automated some of the measurements needed for diagnosis. The time spent describing preventive mammograms was reduced eightfold over the same period.
Clinical decision support systems (CDSS). Programs of this class analyze patient data — examination results, laboratory values, medical record entries — and offer the physician diagnostic or treatment options, which the physician accepts, refines, or rejects. Unlike image analysis systems, a CDSS works primarily with text and tabular data. A CDSS does not replace the attending physician in that physician’s function: who exactly establishes a diagnosis is covered in a separate section below.
Documents and speech. The third area is speech recognition when dictating reports, structuring data in the electronic medical record, and searching a patient’s medical history. Document recognition tasks with similar mechanics are also solved outside medicine — see the article “AI-OCR in KYC: How Automatic Document Recognition Reduces Manual Review”. Such tasks do not relate to diagnosis and, as a rule, do not turn a program into a medical device — what exactly determines this status is covered next.
In brief — how these three areas differ from each other in what determines medical device status (covered in detail in the next section):
| Area | What it does | Usually a medical device? |
|---|---|---|
| Diagnostic radiology | Analyzes CT, MRI, mammography, and X-ray images; highlights suspicious areas, generates a preliminary description of the study | As a rule, yes — if the manufacturer has declared a diagnostic purpose |
| CDSS | Analyzes patient data and offers the physician diagnostic or treatment options | As a rule, yes — if the manufacturer has declared that the program is intended to support diagnosis or treatment |
| Documents and speech | Recognizes speech when reports are dictated, structures electronic medical record data | As a rule, no — such programs usually have no diagnostic or therapeutic purpose declared by the manufacturer |
When AI Software Becomes a Medical Device
Medical device status depends on the purpose declared by the manufacturer: medical devices include, among other things, products — including special software — intended by the manufacturer for the prevention, diagnosis, treatment, and medical rehabilitation of diseases and for monitoring the state of the human body (Article 38(1) of Federal Law No. 323-FZ, the law on the fundamentals of protecting citizens’ health in the Russian Federation). If the manufacturer has declared a program as a diagnostic tool, it is a medical device — regardless of what is inside: a neural network, a classical expert system, or another mechanism. A program for booking patients, electronic document management, or clinic scheduling does not become a medical device unless the manufacturer has declared a medical purpose for it. Medical devices with artificial intelligence do not differ in this sense from any other medical devices: the technology a program is built on does not determine its status.
Medical devices are additionally divided into classes depending on the potential risk of their use and into types in accordance with the nomenclature classification approved by the authorized federal body (Article 38(2)). The risk class does not determine whether a program is classified as a medical device in the first place — the purpose declared by the manufacturer is sufficient for that (Article 38(1)).
In Russia, circulation is permitted only for medical devices that have passed state registration in the procedure established by the Government, or registration under the law of the Eurasian Economic Union; an exception is made for participants in an experimental legal regime under Federal Law No. 258-FZ (Article 38(4)). An experimental legal regime is a special regime for testing new technologies, within which some of the general requirements may be modified by that regime’s own program (Article 38(3.3)). Circulation covers not only manufacture and import but also sale and use (Article 38(3)): a clinic may not use software with a declared medical purpose that has not passed registration.
What to check with a vendor. Before deploying a program with a declared medical purpose, it is worth requesting from the vendor:
- the medical device registration certificate — its number and date;
- the number of the entry in the state register of medical devices;
- whether the version of the program and the declared scope of application match the version and scope the vendor is actually offering.
The total number of AI medical devices registered in Russia differs between open sources and has no official confirmation; for a specific purchase this does not matter — what matters is the registration certificate of the exact product the vendor is offering.
Who Makes the Diagnosis
Under the law, the diagnosis is established by the attending physician: a diagnosis is a medical conclusion about a patient’s disease or condition, based on a comprehensive examination and drawn up using medical terminology (Article 70(5) of 323-FZ). Consultants’ recommendations are implemented only in agreement with the attending physician, except when emergency medical care is provided (Article 70(2)). The attending physician is the physician entrusted with organizing and directly providing medical care to the patient (Article 2(15) of 323-FZ). Article 70 and Article 2 of 323-FZ say nothing about artificial intelligence: these provisions determine only who establishes the diagnosis.
In spring 2025, the head of Roszdravnadzor, Alla Samoylova, said that artificial intelligence does not diagnose patients on its own — its task is reading images and video sequences and analyzing information. The final decision remains with the physician.
What Roszdravnadzor Sees
If AI software that is a medical device has a built-in function for automatic data transfer, the manufacturer or its authorized representative must transmit information about its operation to the Roszdravnadzor automated information system (Roszdravnadzor Order No. 123 of 10 February 2026, clauses 1 and 5 of the Procedure):
- the name of the program and the number and date of its registration certificate;
- the version of the program;
- the scope of application;
- the type of solution;
- the number of studies or objects processed;
- data on the results of the program’s operation;
- data on the number of errors, including network errors related to data quality and errors related to the operation of the program itself;
- the TIN of the medical organization that uses the program.
The obligation rests on the manufacturer of the medical device. Under this order, the clinic has no obligations of its own — only its TIN appears in the transmitted information. The obligation arises only under one additional condition — the program has a built-in function for automatic data transfer; without it, no obligation arises under the order. The order is in force until 31 December 2027.
The 2026 AI Law: What It Regulates
Separately from 323-FZ, Federal Law No. 243-FZ of 26 July 2026 “On Supporting the Development of Artificial Intelligence Technologies in the Russian Federation” is in force — it regulates the development, deployment, and use of large foundation models of artificial intelligence (Article 1(1)). The subject of the law is defined by the model’s architecture, and the features in this definition apply only all together: it is a program intended to perform intellectual tasks at a level comparable to or exceeding human results, using algorithms and learning (or already trained) on data sets, at the same time serving as a basis for creating and refining various types of software, containing at least 1 billion parameters, and used to perform a large number of different tasks (Article 3(2)). A narrow diagnostic program that does not serve as a basis for creating other software does not fall under this definition, even if it formally exceeds the threshold of 1 billion parameters. The law has no separate “high-risk” category with a list of areas of application — there is a general principle of a risk-based approach that takes into account the likelihood and scale of harm to people’s life and health when such models are used, along with other values protected by law (Article 4(5)). The law is in force from 1 September 2026; some of its articles — from 1 March 2027 (Article 13).
323-FZ and 243-FZ define their subject on different grounds: 323-FZ — by the purpose declared by the program’s manufacturer, 243-FZ — by the model’s architecture. These grounds do not exclude each other: a program for which the manufacturer has declared a medical purpose and which at the same time meets the features of a large foundation model under 243-FZ falls under both laws at once.
Dates that are worth keeping apart from one another because they relate to different acts:
| When | What |
|---|---|
| From 1 September 2026 | The AI law (243-FZ) enters into force, except for certain articles |
| From 1 March 2027 | Article 5(2)(3)–(5), Article 6(2)–(5), and Articles 8–10 of 243-FZ enter into force |
| Until 31 December 2027 | Roszdravnadzor Order No. 123 on data transfer to the Roszdravnadzor AIS is in force |
For a clinic, these requirements for AI in healthcare add up to one practical check: before deploying a program with a declared medical purpose, it is worth making sure it is registered as a medical device, clarifying with the vendor whether the program provides for data transfer to Roszdravnadzor, and keeping in mind that the final diagnostic decision remains with the attending physician.
Recognizing a patient’s documents at the registration desk is a task separate from diagnosis, and medical device status usually does not extend to it.
Automatic document recognition for the registration desk is a separate task, unrelated to the status of diagnostic software
FAQ
