Microsoft's experimental medical artificial-intelligence system diagnosed complex cases more accurately than generalist physicians in a peer-reviewed study, achieving 83.2% accuracy against doctors' 36.1%. The results strengthen the company's case for clinical testing, but do not establish how the technology would perform while treating actual patients.
The research, published in Science, evaluated the Microsoft AI Diagnostic Orchestrator, or MAI-DxO, on 304 cases drawn from the New England Journal of Medicine. Physicians in that comparison could consult medical reference materials.
A second assessment covering 325 emergency department-to-admission encounters produced a similar advantage: the system reached 80.3% accuracy, compared with 46.8% for clinicians. Those cases came from the MIMIC-IV-ED database, using records from Boston's Beth Israel Deaconess Medical Center.
Microsoft also reported that the system selected diagnostic tests costing approximately one-half to one-third as much as those chosen by physicians. Those figures represent estimated workup costs, rather than savings demonstrated in hospitals.
"It needs clinical validation," Microsoft AI Chief Executive Mustafa Suleyman told The Verge. "That is a leap, but this is a major milestone along that journey."
Microsoft plans to begin testing the technology in hospital settings with Beth Israel Deaconess, according to Dr. Dominic King, the company's AI health lead and a study coauthor. King told The Verge the work would likely "get underway in the next couple of months," although neither executive committed to a firm timetable.
The study's approach differs from tests that ask an AI model to identify a diagnosis after receiving a complete patient history. MAI-DxO starts with an initial complaint, then requests information, considers possible explanations and chooses tests before reaching a conclusion.
Multiple AI agents work together in a structure intended to resemble a panel of medical experts. The system also considers testing expenses, allowing researchers to evaluate the route to a diagnosis alongside its accuracy.
However, the exchanges took place with an automated "gatekeeper agent" drawing information from existing case records. The system did not interview patients, conduct physical examinations or demonstrate its ability to manage the uncertainties of a live consultation.
That distinction limits what the findings establish. The records themselves reflected earlier clinical work, and the simulation did not directly assess how the system would handle incomplete explanations, changing symptoms or the interpersonal demands of care.
According to The Verge, the orchestrator used OpenAI's GPT-5 and also outperformed individual frontier models tested without its multi-agent structure. Microsoft compiled the cases into a benchmark called SDBench, which it plans to open-source.
The company has not announced when or how MAI-DxO might become commercially available. King emphasized the regulatory and compliance requirements that would accompany deployment in medicine.
Microsoft's broader healthcare plans extend beyond diagnosis to treatment decisions, coordination of care and interactions with patients. Suleyman said human clinicians should retain a central role, expressing hope that automation would free more time for patient contact.
"Next year we're going to see extremely impressive deployments of intelligent systems in production," Suleyman told The Verge. "I think that this is just going to change the way care is delivered in a big way."
