Microsoft's MAI-DxO system correctly diagnosed 83% of complex cases in a simulation, compared with 36% for generalist physicians working individually with reference materials. The findings, published in Science, mark a step in Microsoft AI CEO Mustafa Suleyman's effort to develop what he calls "medical superintelligence."
MAI-DxO coordinates multiple AI agents in a virtual clinical team. The agents gather information, consider possible tests and work through diagnoses together. The simulation used challenging cases, but the comparison did not include specialists or physician teams, who commonly handle cases of this complexity. The results therefore do not establish how the system would perform against specialist-led care.
From simulation to clinical evaluation
Microsoft's analysis found that the system's benefits varied by AI model and case type. On complex cases, the coordination approach improved some models' accuracy and, in certain comparisons, reduced estimated costs. Gains were smaller when the underlying model was already more capable, and the system showed little added benefit on emergency department cases in some tests.
The company is developing a healthcare foundation model with Mayo Clinic, with a release planned for 2027. It also plans to assess MAI-DxO's clinical reasoning with Beth Israel Deaconess Medical Center across hospital settings. In the proposed evaluation, patients would receive standard care while consenting participants also interact with the AI; its recommendations would then be compared with clinicians' decisions.
Microsoft sees the system as a potential addition to its existing clinical tools, with the broader aim of supporting healthcare professionals rather than relying only on isolated diagnostic benchmarks. The next phase of testing will help show whether simulated gains can translate safely and reliably into real clinical settings.
If validated in practice, coordinated AI systems could help clinicians navigate complex information and extend access to decision support as healthcare needs evolve.
