83%: Microsoft AI Outperforms Doctors in Complex Diagnostics

Microsoft has demonstrated that its new diagnostic AI, MAI-DxO, can solve complex medical cases with significantly higher accuracy than generalist physicians. A study published this Thursday in Science reveals that the system, which coordinates multiple AI agents to simulate a panel of doctors, correctly identified diagnoses 83% of the time in simulations. By comparison, primary care doctors and general internists working individually with reference materials were correct only 36% of the time. The study notes that such complex cases are typically handled by specialists or teams in clinical practice, who were not included in this specific comparison.

The system, known as DxO, functions by having AI agents question patients, order tests, and debate potential diagnoses. Microsoft tested the architecture using models from several major technology firms, including OpenAI, Google, Anthropic, Meta, and DeepSeek. The highest performance was achieved when paired with OpenAI’s GPT-5 and GPT-5.5 models. Cost analysis showed that DxO reduced the estimated cost per case by approximately $3,000 compared to running GPT-5 alone, while increasing accuracy by about four points on complex cases. However, for emergency department scenarios, the improvements in both accuracy and cost were minimal.

Mustafa Suleyman, CEO of Microsoft AI, described the findings as a significant step toward achieving medical superintelligence. He stated that the company is currently partnering with the Mayo Clinic to train a foundational healthcare model, with a release scheduled for next year. Suleyman believes this infrastructure will allow for substantial leaps in clinical decision support, predicting that by 2027, the industry will be on the verge of a coding-style breakthrough for medical records. This would enable the streaming of highly accurate predictions regarding patient outcomes, a capability he described as gold dust for clinicians.

The research suggests that much of the performance gain comes from the underlying power of the AI models rather than the orchestration system itself. While DxO helped weaker models like Gemini 2.5 Flash improve from 46.4% to nearly 69% accuracy, the gap between strong and weak models remained larger than the boost provided by the orchestration. Microsoft researchers noted that accuracy gains naturally shrink as models approach theoretical maximums, and the system helps optimize the balance between accuracy and cost.

Microsoft is moving beyond benchmarking to real-world evaluation through a new partnership with Beth Israel Deaconess Medical Center in Boston. The study will assess DxO’s clinical reasoning across various hospital settings, including emergency departments and urgent care. Patients will receive standard treatment while consenting to parallel engagement with the AI. Microsoft and its partners will then compare the AI’s recommendations against those of human doctors. Dominic King, vice president of Health at Microsoft AI, called this the largest study of its kind, aiming to provide compelling evidence for the broader deployment of such systems.

The publication of these results positions Microsoft alongside Google and OpenAI in claiming that AI can match or exceed human medical performance. The company released a manifesto defining medical superintelligence as the capacity to surpass contemporary human-led health systems in reasoning and coordination. Suleyman, who previously co-founded DeepMind, emphasized that the use of generative AI in healthcare is still in its early stages. He expects capabilities to improve by an order of magnitude within the next two to three years, driven by a flywheel of real-world usage and iteration similar to what has occurred in coding and chat applications.

Microsoft views DxO as an expansion of its existing offerings, which include dictation and radiology tools for clinicians. Suleyman stated that the goal is to hold the number one position in state-of-the-art medical clinical decision support products in the coming years. He expressed a strong belief that the rapid progress seen in other AI sectors will now begin to manifest in medical records and clinical support, accelerating the introduction of these technologies to the global market.