Microsoft unveils AI diagnostic system

Microsoft has introduced MAI-DxO, an AI diagnostic system that uses multiple AI agents to assist doctors in making diagnoses. The system was described in a study published in the journal Science, according to TechCrunch. Microsoft's head of AI, Mustafa Suleyman, views this as a step toward "superintelligent medicine." The system functions as a panel of doctors, with multiple AI agents that question patients, prescribe tests, and debate diagnoses.

AI outperforms general doctors in simulations

Microsoft evaluated MAI-DxO in a simulation using complex cases from the New England Journal of Medicine, where the system achieved a correct diagnosis rate of 83%. In comparison, general practitioners and internists working alone with reference resources achieved 36%. However, specialists and medical teams, who are typically assigned these difficult cases, were not included in the comparison. The evaluation was based on a simulation rather than real patients.

AI performance varies by model and case type

Microsoft presented MAI-DxO in a preprint in June 2025, where the system solved 85.5% of 304 cases from the same journal using the OpenAI o3 model, compared to 20% for 21 doctors. The two sets of results are not directly comparable, as they depend mainly on the models used. Microsoft tested MAI-DxO with models from OpenAI, Google, Anthropic, Meta, and DeepSeek. The best results come from GPT-5 and GPT-5.5.

Cost and accuracy improvements vary by model

With a short prompt, GPT-5 alone achieves 78.9% success, at a cost of about $8,400 per case. With MAI-DxO, accuracy improves by about four points and the cost decreases by about $3,000. GPT-5.5 improves from 72% to nearly 87%, at a slightly lower cost. In contrast, in emergency situations, MAI-DxO does not provide significant improvements in either accuracy or cost, with results of about 85% in both cases. Gemini 2.5 Flash improves from 46.4% to nearly 69%.

Real-world testing planned with Boston hospital

TechCrunch concludes that most of the improvement comes from the underlying models rather than the system itself. Microsoft plans to evaluate MAI-DxO in real conditions with the Beth Israel Deaconess Medical Center in Boston. Consenting patients will be treated as usual, while also answering MAI-DxO's questions. Researchers will then compare the tests and diagnoses proposed by the AI with the judgment of doctors.

Microsoft expands health AI partnerships

The group is also working with the Mayo Clinic on a foundation model for health, with a release expected next year. Mustafa Suleyman expects health models to gain an order of magnitude in capability within 24 to 36 months. Dominic King, Vice President of Health at Microsoft AI, considers the technology ready for real-world evaluations and highlights the growing gap between the demand and supply of medical expertise.

Other companies develop similar AI tools

Microsoft is not alone in this area: Google published a comparable study on its tool AMIE in The Lancet, and OpenAI also makes similar claims.