Microsoft’s AI Panel Outperforms Generalists in Complex Medical Diagnostics
New research indicates that a multi-agent AI system from Microsoft achieves higher diagnostic accuracy than human generalists in difficult cases, marking a significant step toward advanced medical support.
By Kevin Iyer ·
Microsoft AI has released a new diagnostic system designed to rival human medical expertise, with CEO Mustafa Suleyman describing the technology as a crucial step toward achieving 'medical superintelligence.' The system, known as MAI-DxO or DxO, functions by coordinating multiple artificial intelligence agents that simulate a panel of doctors. These agents interact with patient data by asking questions, ordering necessary tests, and debating potential diagnoses to reach a final conclusion.
The primary evidence for this capability comes from a study published this week in Science. In simulations using complex cases from the New England Journal of Medicine, the AI system correctly identified the diagnosis 83% of the time. In comparison, primary care physicians and general internists, working individually with standard reference materials, achieved a 36% accuracy rate. It is important to note that these complex scenarios are typically managed by specialists or teams in real-world settings, who were not included in this specific comparison.
Dominic King, vice president of Health at Microsoft AI, emphasized the significance of these results, stating that there is 'something very substantial and exciting here.' He noted that the current healthcare system faces a growing mismatch between the demand for medical expertise and its supply, a gap that this technology aims to address. By publishing these findings, Microsoft joins Google and OpenAI in asserting that their AI models can match or exceed the performance of human doctors in specific clinical contexts.
The study’s data suggests that the majority of the performance improvement stems from the underlying power of the AI models themselves, with Microsoft’s orchestration system providing additional gains. The company tested the system using models from various providers, including OpenAI, Google, Anthropic, Meta, and DeepSeek. The highest accuracy was achieved using OpenAI’s GPT-5 and GPT-5.5 models. When using GPT-5 on its own, the system scored 78.9% on complex cases at an estimated cost of $8,400 per case. The DxO orchestration reduced that cost by approximately $3,000 while increasing accuracy by about four points.
However, the benefits of the orchestration system are not uniform across all types of cases. On emergency department cases, there was little to no improvement in accuracy or cost efficiency when using DxO compared to running the base model alone. With the newer GPT-5.5 model, released in April 2026, the results were mixed: complex case accuracy rose from 72% to nearly 87% with DxO, but emergency department performance remained flat at around 85%. The system showed the most dramatic impact on weaker models, such as Gemini 2.5 Flash, which improved from 46.4% to nearly 69% with the assistance of DxO.
Microsoft is moving beyond simulation to test these capabilities in real-world environments. Through a new partnership with Beth Israel Deaconess Medical Center in Boston, the company will evaluate DxO’s clinical reasoning across different hospital settings. Patients will receive standard care, while those who consent will also interact with the AI system. Researchers will then compare the AI’s recommendations against those of human doctors. King described this as potentially the largest study of its kind, aiming to provide compelling evidence for the scalable deployment of these systems.
Looking ahead, Suleyman stated that Microsoft is collaborating with the Mayo Clinic to train a healthcare-specific foundation model, with a release planned for next year. He believes this infrastructure will allow for 'really enormous leaps' in medical AI. Suleyman expressed confidence that by 2027, the company will achieve a 'coding style breakthrough' for medical records, enabling highly accurate predictions of how a patient's condition might unfold. He added that the company expects healthcare models to improve by an order of magnitude in capability within the next 24 to 36 months, driven by a flywheel of real-world usage and iteration.
