Microsoft's MAI-DxO diagnostic system achieved 83.2% accuracy on a set of complex medical cases, compared with 36.1% for generalist physicians using reference materials, according to research published in Science on Oct. 22. The system also outperformed physicians on a separate set of emergency department cases.

The results are promising, but they do not show that AI is ready to replace doctors. The evaluation used written case records rather than live patients, and the complex-case comparison involved different-sized sets for the AI and physician assessments. The study also leaves open how the system would perform across the full range of conditions doctors encounter.

Microsoft's Medical AI Diagnostic Orchestrator, or MAI-DxO, is designed to simulate a panel of clinicians working through a case. It can ask follow-up questions, request tests, consider competing diagnoses and assess whether further investigation is worth the cost.

The system coordinates existing AI models rather than functioning as a standalone medical model. In the reported complex-case test, GPT-5 alone achieved 78.9% accuracy, while MAI-DxO running on GPT-5 reached 83.2%. That roughly four-percentage-point improvement suggests that organizing the diagnostic process can improve an already capable model's performance.

The system is intended to make diagnosis a step-by-step process rather than rely on a single answer to a complete case description. Its performance, however, depends on both the underlying model and the orchestration method.

Researchers tested MAI-DxO on 304 complex cases from The New England Journal of Medicine and 325 emergency department encounters. It achieved 83.2% accuracy on the complex cases and 80.3% on the emergency cases. Generalist physicians averaged 36.1% and 46.8%, respectively, in the reported comparisons.

Those numbers need context. The NEJM cases are selected for their diagnostic complexity, often requiring extensive investigation and specialist knowledge. They are not representative of every patient who visits a clinic.

The physician comparison for the NEJM cases also used a held-out subset of 56 cases, while MAI-DxO was evaluated across all 304. The headline percentages, therefore, should not be read as the result of every doctor and the AI independently diagnosing an identical set of 304 cases.

The evaluation was also text-based. Neither the AI nor the doctors had access to a full clinical encounter involving physical examination and imaging. The benchmark did not include healthy patients or people with benign conditions, so it could not establish how often the system might wrongly flag a serious disease.

That matters because diagnosis involves more than identifying rare or difficult conditions. Doctors must also recognize common illnesses, decide when testing is unnecessary, and avoid investigations that could expose patients to needless cost or risk.

MAI-DxO asked an average of 25.5 questions before ordering its first test, compared with 5.6 for physicians. That more extensive history-taking may have helped the system identify clues that would otherwise be missed, although its practical value in a busy clinical setting remains untested.

The system's strongest accuracy result was not its cheapest configuration. On the complex cases, the standard setup cost a simulated $5,431 per case, compared with $4,725 for physicians using reference materials.

Microsoft also tested configurations with spending limits. One reached 80.9% accuracy at an estimated $2,241 per case, while another achieved 79.6% at $1,303. These results suggest that the system could retain much of its diagnostic performance while reducing modeled costs.

The figures are estimates based on the benchmark's testing-cost methodology, not actual hospital bills. They also do not establish whether the same savings would occur in clinical practice, where costs depend on the healthcare system, available tests, and the patient's circumstances.

Microsoft has announced plans to evaluate MAI-DxO with Beth Israel Deaconess Medical Center in Boston. Testing in clinical settings could help establish whether the system's performance carries over when doctors deal with real patients, incomplete information and changing symptoms.

The results will need to show more than diagnostic accuracy. Researchers will also need to assess false alarms, missed conditions, unnecessary tests, and whether the system improves clinicians' decisions.

Microsoft has described the findings as progress toward what it calls medical superintelligence. That is the company's ambition, not a conclusion established by the study. For now, MAI-DxO has demonstrated strong performance on selected diagnostic benchmarks. Whether it can improve patient care safely, reliably, and cost-effectively remains a question for clinical testing.