Microsoft says its Diagnostic Orchestrator outperformed human physicians in tests built from past medical cases. The AI system reached a diagnosis by coordinating multiple agents to review symptoms and medical histories, consider possible explanations, and decide which tests to order. The results point to a potential role for AI in clinical reasoning, while leaving a major question open: how well will it work with patients in real settings?
Strong results on two sets of cases
The study compared the system with physicians using scenarios drawn from medical case reports and emergency department data. Microsoft assembled the cases into a benchmark called SDBench, which it plans to open-source. The system is powered by OpenAI’s GPT-5.
For cases drawn from hundreds of reports in The New England Journal of Medicine, Microsoft reported diagnostic accuracy of around 83 percent. Generalist physicians reached roughly 36 percent when they could use references and 22 percent without them.
The tool also scored higher on emergency department cases based on patient admissions at Beth Israel Deaconess Medical Center in Boston. It achieved 80 percent accuracy, compared with an average of 47 percent for physicians. Microsoft also said the AI tended to reach diagnoses at a lower estimated cost than human counterparts.
Those comparisons are striking, but they describe performance on selected past cases. They do not establish how the system would perform when helping to assess patients as care unfolds.
A team of AI agents works through a case
Diagnostic Orchestrator is designed to resemble a panel of medical experts. Its agents work through the information available, weigh possible diagnoses, and choose tests before settling on an answer. That process is intended to reflect how doctors investigate a case more closely than a system that simply receives a question and returns a diagnosis.
The researchers say it also outperformed frontier AI models running on their own. However, the peer-reviewed paper does not include newer models named in the article. The system was first developed when GPT-4.1 was the most advanced model available, and later testing suggested that stronger individual models do not necessarily erase the benefits of combining agents.
In practice, the approach is an attempt to organize AI reasoning into steps that resemble clinical work. But a benchmark result cannot show whether the system will ask useful questions, handle incomplete information, or fit smoothly into a clinician’s decisions during actual care.
Real-world testing remains ahead
Microsoft acknowledges that clinical validation is still needed. The company plans to test the system in hospital settings with Beth Israel Deaconess Medical Center, according to its AI health lead, Dr. Dominic King. King said the work would likely “get underway in the next couple of months,” but neither he nor Microsoft AI CEO Mustafa Suleyman gave a firm timeline.
The company has not said when or how the technology might become commercially available. Any rollout in medicine would have to meet high regulatory and compliance requirements, King said. Those demands matter because diagnostic tools can influence treatment decisions and patient care.
Microsoft’s broader health ambitions extend beyond diagnosis. King described other areas AI would need to master, including coordinating clinical care, making treatment decisions, and handling the human side of medicine. The company’s stated goal is what it calls “medical superintelligence,” an AI that can outperform the best human-led healthcare systems.
Microsoft expects clinicians to remain central
Suleyman has predicted that AI could reach performance indistinguishable from human performance in medical superintelligence “in the next couple of years.” He has also said human clinicians should continue to play a central role in patient care.
One proposed benefit is reducing routine work so clinicians can spend more time with patients. Suleyman also recognizes that AI could leave doctors with additional work. Whether the technology lightens workloads will depend on how it is introduced and how well it fits into clinical practice.
For now, Microsoft’s study offers evidence of strong results on a constructed benchmark and a plan to move toward hospital testing. The next step is to learn whether those results carry over to real clinical environments, where the system’s usefulness and limits can be evaluated alongside the people responsible for care.