Seven steps
1. Choose a process with a measurable loss: no-shows, missed requests, minutes per document. 2. Set a baseline before the start: the same measure for a past period. 3. Separate service AI, which needs no registration, from medical AI, which goes through the register of medical devices. 4. Arrange the data: a processing agreement, consent, servers in Russia. 5. A pilot of 14 days on one process, with the metric in the contract. 6. Train the staff: in a clinic network in Kenya, doctors corrected errors more often on AI signals after training and mentoring. 7. Scale one agent per process: in Moscow, radiologists already have 70 AI services.
What breaks most often
Too many alerts: in a Canadian hospital the system warned only about high-risk patients, and according to the study, unexpected deaths in the ward fell by 26%. No single data source: the Kuzbass and Sverdlovsk regions started with a central image archive. A pilot without a metric: then there is nothing to scale. The agent answers medical questions: this is a risk for the clinic, and such a question must go to a doctor.
Examples
A clinic network: a robot confirms visits, and no-shows, which were about 2–3% of bookings, went down. A laboratory: test result explanations validated by experts. A university clinic in the USA: AI documentation, and 73% of surveyed clinicians spend less time on paperwork after a shift. In each example, AI works on one process.