Microsoft MAI-DxO and the future of medical artificial intelligence

What a patient needs is a reliable diagnosis and a clear path to treatment. Microsoft's latest medical AI research, published in the journal Science on October 22, 2026, asks healthcare leaders a practical question: how should an organisation evaluate a system that gathers evidence over many steps before deciding?

Paired with GPT-5, MAI-DxO achieved 83.2 per cent diagnostic accuracy on 304 difficult NEJM cases, against 36.1 per cent for general practitioners using reference resources and 22.3 per cent without them. In 325 emergency-department admissions it reached 80.3 per cent against doctors' 46.8 per cent. These are controlled evaluations on case records; the doctors' figures reflect the study's tasks, not everyday practice.

MAI-DxO orchestrates a virtual medical panel that proposes diagnoses, questions assumptions and chooses tests, updating as new information arrives, while the benchmark also scores the path taken, including estimated test costs. Microsoft has announced a prospective study with Beth Israel Deaconess Medical Center in which patients will interact directly with the system; that is planned validation, not evidence that outcomes have already improved.

The design lesson is to separate proposing an answer from reviewing its evidence: before a high-impact action, a reviewer should spot missing information and test the leading explanation. Multiple AI roles can repeat the same mistake, so agreement among agents must be checked against external sources.

Compare systems on three dimensions: the final answer's score against the quality of the decision process; estimated task cost against total service cost; and performance on old records against outcomes in a prospective study.

Before funding a pilot, chief executives should define the expected benefit, appoint a clinical owner, agree outcome measures and stop conditions, and avoid workforce decisions based on benchmarks. Technology officers should evaluate around the intended workflow, beginning with a retrospective evaluation and moving to a supervised prospective study only if the evidence suffices.

Evidence behind recommendations must be traceable, patient-data access controlled, and model and workflow versions recorded, with evaluations repeated when either changes. If you approved a medical AI pilot, what evidence would you demand before letting its recommendations affect patient care?