REDMOND, WASHINGTON / RankWire.AI / – A Microsoft AI system that works through diagnoses one step at a time, like a doctor, proved more than twice as accurate as general physicians on complex medical cases. The results were published on Thursday, October 22, in the journal Science. The tool, called the Microsoft AI Diagnostic Orchestrator or MAI-DxO, is described in a paper titled “Sequential Diagnosis with Language Models.” It used OpenAI’s GPT-5 during testing.
MAI-DxO decides which questions to put to a patient and which exams and tests to order. An orchestrator directs the process by assigning roles to a simulated panel of clinical experts who work together on each case, according to the study. Most current medical AI models instead take an all-at-once approach. They review every symptom, lab result and imaging study at the same time before producing a diagnosis.
Under the sequential method, the AI first chooses what to ask about a patient’s history and what to focus on in the physical exam. It then selects tests based on what it has learned, examines the results and requests more testing if needed. Research leaders said in an interview that the model keeps updating its working hypotheses and plans along the way. The design follows the way physicians typically investigate unclear symptoms.
Tool tested on 629 difficult cases
The model was challenged with 304 rare and confusing case reports from the New England Journal of Medicine. It diagnosed 83.2% correctly, while generalist doctors with access to textbooks and the internet managed 36.1%. A second test used 325 cases drawn from a large database of emergency room admissions. MAI-DxO scored 80.3% on those scenarios. The physicians in that comparison scored 46.8%. On the journal cases, the gap amounted to more than double the doctors’ rate.
According to the researchers, earlier comparisons between AI systems and doctors usually relied on multiple-choice questions. Harsha Nori, head of AI at Microsoft AI Health, said the team measured not only accuracy but how carefully the system conducts a diagnostic work-up. Dominic King, a corporate vice president in Microsoft’s AI health unit, noted that real patients do not come with tidy summaries. The study therefore tracked each step the AI took.
Boston emergency room study comes next
The researchers acknowledged limits to the findings. The doctors in the study were generalists, and every case was presented in text form. Neither the AI nor the physicians had access to physical exams or medical imaging. Cost savings reported for budget-limited versions were estimates rather than real medical bills. Global health leaders have also called for country-led AI in disease control, reflecting wider interest in medical AI.
Microsoft plans to test the tool with Beth Israel Deaconess Medical Center in Boston. That work will examine whether MAI-DxO can help triage patients waiting in emergency departments. The system was first outlined in a preprint in June 2025, which showed it could work with models from Google, Anthropic and Meta. Microsoft has also built custom AI chips for its Azure cloud. The tool is not yet used in clinical care.