Microsoft published new research on October 22, 2026, showing that its AI Diagnostic Orchestrator (MAI-DxO) achieved high accuracy in simulated medical diagnosis tests. The findings, published in a peer-reviewed study in Science, suggest that artificial intelligence could eventually help doctors investigate complex health conditions. However, the results do not mean the technology is ready to diagnose patients independently.
Microsoft’s AI Diagnostic Orchestrator (MAI-DxO) uses large language models, a form of machine learning, to investigate medical cases. Its main reported results used GPT-5. Instead of giving one immediate answer, the system considers possible diagnoses, requests additional clinical information, and evaluates which medical tests could help reach a conclusion.
Microsoft’s researchers evaluated 304 difficult medical cases from The New England Journal of Medicine. MAI-DxO achieved 83.2% diagnostic accuracy, while participating generalist physicians averaged 36.1% when using textbooks and online references and 22.3% without those resources. The physician comparison involved a smaller set of 56 held-out cases, so the findings should not be treated as a direct clinical trial comparing AI with specialist medical teams.
These findings suggest that AI could help doctors investigate complex medical conditions, but the results come from controlled, text-based simulations using historical case records. The evaluation did not involve direct patient examinations or real treatment decisions. It also compared the system with individual generalist physicians rather than specialist teams. Further clinical testing is needed to determine whether MAI-DxO can deliver similar results in hospitals.
The researchers evaluated MAI-DxO using SDBench, a benchmark that turns written medical cases into step-by-step diagnostic exercises. The system coordinates several simulated physician roles to examine possible diagnoses, request additional information, and select useful medical tests. A separate AI component provides findings from the medical case record or generates case-consistent information when certain details are missing.
The research also examined whether MAI-DxO could reduce the cost of medical testing. However, the savings depended on how the system was configured. The version that achieved 83.2% accuracy had estimated diagnostic workup costs of about $5,430 per complex case, compared with $4,725 for physicians using reference materials. A more cost-focused version reached 80.9% accuracy at an estimated cost of $2,241 per case. These figures represent calculated diagnostic costs, not actual savings achieved in hospitals.
Microsoft also tested MAI-DxO using 325 previous emergency-department admissions drawn from MIMIC-IV medical records. In this evaluation, the AI system achieved 80.3% diagnostic accuracy, compared with 46.8% for participating physicians. The researchers have documented the related emergency-department benchmark dataset through PhysioNet, allowing qualified researchers to examine how the medical records were prepared for testing.
Despite its promising results, MAI-DxO cannot replace healthcare professionals. The study did not establish how the system handles physical examinations, treatment decisions, or long-term patient outcomes. These limitations also reflect wider concerns about AI-generated medical advice, particularly when automated answers are used without appropriate clinical supervision.
According to Microsoft’s research announcement, the company plans to evaluate MAI-DxO through a prospective study with Beth Israel Deaconess Medical Center in Boston. Patients will interact with the system in clinical settings, and researchers will compare its diagnostic assessments with those of physicians. The planned research will help determine whether the system’s promising benchmark results translate into reliable medical performance.
Microsoft’s research is part of a growing AI medical diagnostics race involving major technology companies. OpenAI, Google, and Anthropic have also introduced healthcare-focused AI systems, although these products serve different purposes and cannot be directly compared with MAI-DxO’s diagnostic benchmark results.
For patients, the practical benefits remain unproven. If future clinical studies confirm the findings, systems like MAI-DxO could help doctors review challenging cases and choose more appropriate medical tests. However, the current research has not established whether the technology can improve treatment outcomes or reduce waiting times in real healthcare settings.
Microsoft’s 83% accuracy result marks an important development in AI-assisted medical research, but it does not establish how the system will perform in everyday clinical practice. The next challenge is testing whether MAI-DxO remains accurate, safe, and cost-effective when working with real patients and medical professionals.
