Microsoft’s AI diagnostic system gave clinicians a stronger model for deciding which questions and tests to pursue in difficult cases. In controlled tests, the system was more than twice as accurate as general physicians at identifying the underlying condition, according to a study published in Science on October 23, 2026.
Researchers said the system could also reduce diagnostic costs, while clinical testing in emergency departments will determine whether its performance carries over to everyday medical practice.
How does Microsoft’s AI approach medical diagnosis?
Microsoft researchers designed the system to treat diagnosis as a sequence of decisions. It begins with a patient’s medical history and physical examination, identifies relevant follow-up questions, selects investigations that may narrow the possibilities and reassesses its diagnoses as new evidence becomes available.
An orchestration mechanism directs the process and determines which steps to take. The system also draws on a simulated group of medical specialists, coordinating their contributions as it investigates a case. This differs from many existing medical AI systems that analyze symptoms, laboratory findings and medical images in one pass to produce a diagnosis.
The sequential design reflects the uncertainty clinicians face when initial symptoms fit several conditions. Later examinations and test results can rule out some explanations, strengthen others, and determine the most useful next step.
Dr Dominic King, corporate vice president of Microsoft’s health unit, said conventional AI assessments can give an incomplete picture of a system’s clinical capabilities. He said AI has become very good at getting a 100% pass rate on multiple-choice medical questions, but “No patient ever came to see me with a nice summary.”
How did the system perform against physicians?
The researchers evaluated the approach using more than 300 uncommon and diagnostically challenging cases. General physicians investigated the same cases with access to textbooks and online resources. Microsoft’s AI system achieved more than twice their accuracy in working through the available evidence and reaching a diagnosis.
The system also performed better than human physicians on cases based on records from a major emergency-department admission database. The results suggest that organizing an investigation can be as important as recognizing patterns in medical information.
Harsha Nori, head of AI at Microsoft AI Health, said the evaluation considered both diagnostic accuracy and how efficiently the system gathered evidence.
What potential cost benefit did the study find?
The researchers reported that diagnostic investigations cost less when the AI selected the next steps than when doctors conducted the work-ups. Prioritizing relevant questions and tests could limit unnecessary investigations, although the controlled results do not establish how much hospitals might save in routine practice.
Microsoft is working with Beth Israel Deaconess Medical Center to assess whether the system can help triage patients in emergency-department reception areas. That work will test its usefulness when information is incomplete and patients need different levels of attention.
The current findings do not show that the technology can replace doctors. Researchers have not yet shown how the technology performs across the full range of patients encountered in hospitals.
What to do now
- Keep AI as decision support rather than a replacement for clinical judgment. The study tested the system in controlled assessments, so healthcare organizations should evaluate it under clinical conditions before changing care pathways.
- Measure both the answer and the investigation. When evaluating an AI diagnostic tool, review the accuracy of its conclusion alongside the questions, tests and evidence it uses to reach that conclusion.
- Emergency-department assessments matter because clinicians may lack information and face varying urgency and levels of patient attention.
- Evaluations should assess the efficiency of each work-up because a system that reaches the right conclusion but orders unnecessary investigations may offer limited value.
- Build trust through transparent sourcing and oversight. A related systematic review in the British Journal of Ophthalmology identified data hallucination, authorship and data privacy as recurring concerns in large language model research and called for standardized evaluation frameworks.
FAQ
What did Microsoft’s AI diagnostic system do better than general physicians?
In controlled tests covering more than 300 uncommon and difficult cases, the system was more than twice as accurate as general physicians at investigating the evidence and reaching a diagnosis. It also performed better on cases based on emergency-department admission records.
Could Microsoft’s AI diagnostic tool reduce healthcare costs?
The researchers reported that diagnostic investigations cost less when the AI determined which steps to take than when doctors conducted the work-ups. The study did not establish how much hospitals might save in routine practice.
Is Microsoft’s AI diagnostic system ready to replace doctors?
No. The findings came from controlled assessments, and the technology still needs to demonstrate comparable performance in real-world clinical settings. Microsoft is collaborating with Beth Israel Deaconess Medical Center on emergency-department triage research.
Sources
See our editorial disclaimer for how our articles are produced.
Run a free scan to see your AI Visibility Score, SEO rating, and local citation accuracy.