Microsoft has unveiled an experimental artificial-intelligence system, the Microsoft AI Diagnostic Orchestrator (MAI-DxO), that it says can diagnose complex illnesses far more accurately than human physicians. In an internal study the tool correctly solved 85.5% of 304 difficult cases taken from the New England Journal of Medicine, compared with a 20% success rate achieved by 21 experienced doctors—roughly a fourfold advantage. The benchmark required the system to mimic real clinical practice by iteratively questioning patients, ordering laboratory and imaging tests, and updating its assessment until a final diagnosis was reached. MAI-DxO orchestrates several frontier large-language models—including versions from OpenAI, Google, Anthropic, Meta and xAI—so that they behave like a virtual panel of specialists. Microsoft says the approach also reduced simulated diagnostic spending by about 20% because the software chose fewer, lower-cost tests. Mustafa Suleyman, who heads Microsoft’s AI health unit, called the result a “step toward medical superintelligence” that could ease workloads and speed treatment decisions. The project is the first public initiative from the consumer-health team formed late last year after the company hired several high-profile researchers from Google. The software has not been cleared for clinical use. Microsoft plans further trials with healthcare organisations and will seek regulatory approval before any deployment. External experts described the early results as promising but cautioned that physicians in the study lacked the reference materials they would normally consult, and that real-world validation is needed to confirm the system’s safety, accuracy and cost benefits across diverse patient populations.