Microsoft medical diagnostic tool outperforms general practitioners

Microsoft tested an AI system capable of step-by-step reasoning on complex clinical cases. Published in Science, the study reports it as more accurate than general practitioners, with less costly assessments.

Microsoft claims to have developed an AI system that replicates a clinician's diagnostic approach and outperforms general practitioners on complex cases. Published in Science in October 2026, these results are significant because they focus less on a multiple-choice test and more on the way of reasoning, examining, then deciding. The tool was tested on over 300 unusual medical cases and on scenarios from emergency departments.

An AI that diagnoses step-by-step, not all at once

The difference with classic medical models is clear. Instead of absorbing all symptoms and examinations at once, Microsoft's system relies on an orchestrator, a central component that coordinates virtual experts and updates its hypotheses as information is gathered. It starts by choosing the right questions about the patient's history, then selects useful elements from the physical exam, requests additional tests, and adjusts its diagnosis at each step.

This sequential logic resembles a real consultation more than an instant chatbot response. Dr. Dominic King, vice president of the health unit at Microsoft AI, sums up the limit of usual benchmarks: "AI has become very good at achieving 100% success on multiple-choice tests. But no patient ever arrived with a neat, well-prepared summary." The statement is simple, but the idea behind it is serious: a good diagnosis also depends on how you search, not just the final answer.

Better results, and less costly assessments

The researchers did not just measure final accuracy. The study also evaluates the diagnostic workup, that is, the sequence of exams and tests prescribed to reduce uncertainty. On this front, Microsoft's tool also took the lead: it was more than twice as accurate as general practitioners on complex cases, outperformed human clinicians on real emergency scenarios, and designed assessments less costly than those prescribed by human doctors.

Harsha Nori, head of AI at Microsoft AI Health, emphasizes this point: "We also measure how wisely and rigorously the system progresses through the diagnostic workup." In other words, the machine does not just guess correctly; it also avoids multiplying exams out of excessive caution. In medical jargon, this detail weighs heavily, because a good diagnosis is not only a matter of accuracy but also of care pathways and costs.

Towards use in emergency rooms, not replacing doctors

The next step will take place outside the laboratory. Microsoft is collaborating with the Beth Israel Deaconess Medical Center in Boston to test the tool in real conditions, in emergency waiting rooms, to evaluate its ability to triage patients effectively. This transition from testing to use matters more than the score itself: a medical AI that works on paper is good; in an overwhelmed service, it is another story.

At this stage, the goal is not to replace a doctor with a model, but to relieve teams of the most time-consuming cases to give them more time with patients who need it most. This is precisely where this research becomes concrete for the digital health market: if the tool lives up to its promises, it could be integrated as an assistant for triage, preparation, and reasoning support, rather than as an all-knowing oracle — which is, fortunately, more reassuring for everyone.

Key points

  • Published in Science in October 2026.
  • Over 300 unusual medical cases tested.
  • Microsoft cites the Beth Israel Deaconess Medical Center in Boston.
  • The tool outperformed general practitioners on complex cases.
  • Diagnostic workups cost less than those of humans.

By the numbers

  • Over 300 cases — unusual medical cases tested by Microsoft.
  • Twice as much — accuracy claimed compared to general practitioners on complex cases.
  • October 2026 — publication of results in Science.
  • 1 hospital center — Beth Israel Deaconess Medical Center in Boston, for real-world testing.

To read