Microsoft Built an AI That Diagnoses the Way Doctors Do. In a New Study, It Beat Them.

HOUSTON, October 23, 2026 — Medical AI has spent the last few years acing its exams. Hand a model a tidy summary of a case and four possible answers, and it will outscore most doctors. The trouble is that patients never show up as tidy summaries. They show up with a vague pain, a half remembered history, and a story that only makes sense one question at a time.

Microsoft's health team has been working on that messier version of the problem: diagnosis as it is actually practiced. On Thursday, the journal Science published their results, and they are hard to wave away.

The system is built around what Microsoft calls a diagnostic orchestrator. Rather than swallowing a patient's symptoms, lab results, and imaging in one gulp and pronouncing a verdict, it stages the work. It decides what to ask about the patient's history and what matters in the physical exam. Based on what it learns, it orders tests, studies the results, and orders more where the trail points. All the while it keeps a running set of working hypotheses and revises them as evidence lands. Microsoft describes the orchestrator as a simulated panel of clinical experts, collaborating on the case step by step.

The cases it was given were not board exam fare. The study ran the system against more than 300 highly unusual and confusing medical dilemmas, the kind of cases that get written up precisely because they stumped good doctors. General physicians worked the same cases with textbooks and the internet at hand, the way they would in practice. The orchestrated system was more than twice as accurate at working up the patient and landing on the diagnosis. Presented with scenarios drawn from a major database of emergency room admissions, it outperformed the physicians again.

It was also cheaper. Because the orchestrator chooses each test deliberately instead of ordering everything at once, the work ups it designed cost less than the ones the doctors designed. The team measured that on purpose. "We're not just measuring the accuracy of the systems," said Harsha Nori, who leads AI at Microsoft AI Health. "We're also measuring how judiciously and carefully it goes through a diagnostic work up."

Dr. Dominic King, a surgeon who now runs health at Microsoft AI as a corporate vice president, is blunter about why the older benchmarks mislead. AI, he said, "has become very good at getting a 100% pass rate" on multiple choice tests. "But of course, that's not real world. No patient ever came to see me with a nice summary."

What the study does not show is worth keeping in view. These were simulated cases, worked in a lab, not real patients in real wards. The comparison doctors practiced without the colleagues, consults, and second opinions that real diagnosis leans on. And an emergency department database scenario is still a scenario. Nobody has yet watched this system perform at an actual bedside, with an actual frightened human being in front of it.

That test is next. Microsoft is working with Beth Israel Deaconess Medical Center in Boston to try the tool where the pressure is highest and the summaries are worst: triaging patients in emergency department waiting rooms. If it holds up there, the argument changes. A system that can reason through a work up carefully, order the right tests in the right order, and do it at lower cost is not a trivia champion anymore. It starts to look like infrastructure, the kind that could put a senior clinician's reasoning within reach of every small hospital and overwhelmed ER in the country.

The American Chamber of Health will be watching the Boston pilot. So, you can be sure, will every health system in America.

Source: Reuters, "Microsoft develops AI tool that can diagnose 'sequentially' like a doctor," October 23, 2026; study published in Science.