Microsoft’s AI Diagnostic Panel Outpaces Human Doctors on Complex Cases

Microsoft has unveiled a significant leap in its artificial intelligence capabilities, introducing a diagnostic system that outperforms human generalist physicians on complex medical cases. Mustafa Suleyman, the CEO of Microsoft AI, described the achievement as a critical step toward what the company terms “medical superintelligence.” This new system represents a departure from simple chatbots, functioning instead as a coordinated panel of AI agents designed to simulate the collaborative nature of expert medical teams.

The system, officially named MAI-DxO (or DxO for short), operates by having multiple AI agents engage in a structured process of patient questioning, test ordering, and diagnostic debate. In a recent simulation published in the journal Science, this multi-agent approach demonstrated a stark contrast in performance compared to individual human doctors. When tested against complex cases sourced from the New England Journal of Medicine, DxO arrived at the correct diagnosis 83% of the time. By comparison, primary care physicians and general internists, even when working individually with reference materials, were correct in only 36% of those same instances. It is worth noting that such complex cases are typically handled by specialists or teams in real-world settings, who were not included in this specific comparison.

The technology behind DxO leverages powerful underlying models from various providers, including OpenAI, Google, Anthropic, Meta, and DeepSeek. The study found that the system performed best when paired with OpenAI’s GPT-5 and GPT-5.5 models. However, the data suggests that much of the accuracy boost comes from the strength of the base model itself, with the orchestration system providing additional, albeit smaller, gains. For instance, when using the newer GPT-5.5 model released in April 2026, DxO improved the accuracy score on complex cases from 72% to nearly 87% while slightly reducing costs. In emergency department scenarios, however, the performance gap between the AI with and without the DxO orchestration was minimal, with both scoring around 85%.

Key figures at Microsoft AI emphasize the practical implications of these results. Dominic King, vice president of Health at Microsoft AI, highlighted that the current healthcare landscape is defined by a growing mismatch between the demand for medical expertise and the available supply. He stated, “There is something very substantial and exciting here,” noting that the technology is ready for real-world evaluations. To validate these findings outside of simulations, Microsoft is partnering with Beth Israel Deaconess Medical Center in Boston to assess DxO’s clinical reasoning across various hospital settings, including urgent care and emergency rooms.

The stakes for this development are high as Microsoft joins competitors like Google and OpenAI in claiming that AI can match or surpass human diagnostic capabilities. Suleyman, who previously co-founded DeepMind before joining Microsoft in 2024, views this work as foundational infrastructure for future breakthroughs. He is collaborating with the Mayo Clinic to develop a healthcare-specific foundation model, with a release planned for next year. Suleyman predicts that by 2027, the company will be close to a “coding style breakthrough” for medical records, allowing for highly accurate predictions of patient outcomes—a capability he describes as “gold dust” for clinicians.

Microsoft aims to move beyond benchmarking individual tasks and instead integrate these systems into broader healthcare workflows to improve overall outcomes. The company has defined “medical superintelligence” as the capacity to surpass the strongest contemporary human-led health systems in reasoning and coordination. With existing tools for dictation and radiology already in the market, Suleyman asserts that Microsoft intends to hold the “absolute number one” position in state-of-the-art clinical decision support products in the coming years, leveraging rapid iteration cycles to refine these models through real-world usage.