Microsoft researchers present AI that diagnoses step-by-step and outperforms general practitioners
Study in Science shows a coordinator-driven AI that decides which questions, exams and tests to run and adapts the diagnostic workflow step by step.
Microsoft researchers have unveiled an artificial intelligence tool that diagnoses illnesses step-by-step, and the system proved to be more than twice as accurate as general practitioners at resolving complex medical cases, according to a study published in the journal Science on 23 October 2026.
The model guides the diagnostic process by deciding which questions to ask a patient and which examinations and tests to perform. That decision-making is supported by a coordination mechanism that structures how the AI approaches a case, using a simulated team of clinical experts that collaborate to solve diagnostic problems.
Unlike many existing AI systems for medical use, which typically examine all reported symptoms, laboratory results and scans at once before producing a diagnosis, this approach leads the diagnostic workflow stepwise and adapts as new findings arrive.
When the AI leads the process, the model first determines what to ask about the patient’s medical history and which physical-examination findings are relevant. Based on answers and observed findings, it then selects which tests to run, reviews the results and, if necessary, orders additional tests.
How the system is organized
The study describes a coordinator that orchestrates a simulated clinical team: individual expert agents focus on different aspects of a case and exchange information so the system can prioritize questions and investigations the way a clinician would during a real diagnostic encounter.
The findings, presented in Science, mark a research advance in diagnostic AI but remain a study result rather than an immediate clinical rollout; the paper documents the model’s performance on complex cases and details the stepwise process used by the system.
Photo: press material from the event
