Microsoft’s AI Panel Outpaces Generalists in Diagnosing Complex Illnesses

For patients facing difficult medical diagnoses, a new approach to artificial intelligence is beginning to offer a second opinion that may prove more accurate than a single doctor’s judgment. A study published Thursday in *Science* reveals that a Microsoft AI system correctly identified complex medical conditions in a simulation more than twice as often as generalist physicians. While this does not replace human doctors, it signals a shift in how technology might support clinical decision-making for the most challenging cases.

The system, known as MAI-DxO, operates by coordinating multiple AI agents that function like a virtual panel of specialists. Rather than relying on a single algorithm, these agents question the patient, order tests, and debate potential diagnoses before reaching a conclusion. In a simulation using complex cases from the *New England Journal of Medicine*, the system achieved an 83% accuracy rate. By comparison, primary care doctors and general internists working individually with reference materials were correct only 36% of the time. It is important to note that such complex cases are typically handled by specialists or teams in real-world settings, who were not part of this specific comparison.

Microsoft’s leadership views these results as a critical step toward what CEO Mustafa Suleyman terms “medical superintelligence.” He believes the technology addresses a growing mismatch between the demand for medical expertise and the available supply of doctors. However, the study’s data indicates that much of the improvement stems from the use of powerful underlying models rather than the orchestration system alone. When tested with models from various providers, the best results were achieved using OpenAI’s GPT-5 and GPT-5.5, suggesting that the quality of the base model remains a significant factor in performance.

Cost and efficiency present a more nuanced picture. On complex cases, the AI system reduced the estimated cost per case by approximately $3,000 compared to using GPT-5 alone, while raising accuracy by only about four points. For emergency department cases, the system offered little to no improvement in either accuracy or cost over the standalone model. The benefits were most pronounced when using less capable models, such as Gemini 2.5 Flash, where the system boosted accuracy from 46.4% to nearly 69%. Nevertheless, the performance gap between weaker and stronger models remained larger than the boost provided by the orchestration system itself.

To move beyond simulations, Microsoft is preparing to test the technology in real-world hospital environments. Through a partnership with Beth Israel Deaconess Medical Center in Boston, the company will evaluate the system’s clinical reasoning across various settings, including emergency rooms and urgent care. Patients will receive standard treatment while those who consent will also interact with the AI, allowing researchers to compare the AI’s recommendations against those of human physicians. This initiative follows a similar study by Google and aims to provide compelling evidence for the broader deployment of such systems.

Looking ahead, Microsoft is collaborating with the Mayo Clinic to develop a specialized foundation model for healthcare, scheduled for release next year. Suleyman predicts that within the next 24 to 36 months, healthcare models will see an order of magnitude improvement in capability. He anticipates a breakthrough in 2027 where the system can stream highly accurate predictions about how a medical record might unfold, offering clinicians valuable foresight. This effort represents a strategic move for Microsoft to differentiate itself in the AI race by focusing on deep integration with clinical workflows rather than just benchmarking individual tasks.

The publication of these results coincides with a broader manifesto from Microsoft on building medical superintelligence, defined as surpassing human-led systems in medical reasoning. By leveraging its existing tools for dictation and radiology, the company aims to make DxO a central component of its healthcare offerings. As the technology matures, the goal is to integrate these AI systems seamlessly into the full arc of patient care, ultimately enhancing the speed and accuracy of diagnosis for patients everywhere.