Microsoft’s AI doctor works up a case step by step, and beats the physicians

Most medical AI has been tested on an exam it can already pass. Microsoft says its new diagnostic tool was built to take a harder one. In a study published in Science on Thursday, the company's researchers reported that the system was more than twice as accurate as general physicians at solving complex cases, and that it did the work at lower cost.

The test material matters here. The team gave the model more than 300 highly unusual and confusing medical cases. The physicians it was compared against were general practitioners who could use textbooks and the internet. The tool also outperformed human doctors on scenarios drawn from a major database of emergency room admissions.

How does it work?

Most existing medical AI takes an all-at-once approach. It looks at a patient's symptoms, lab results and imaging together and produces a diagnosis. Microsoft's model works differently. It uses what the researchers call an orchestrator, which organizes how the AI tackles a problem by running a simulated panel of clinical experts who collaborate on the case.

The process follows the order a clinician would. The model first decides what to ask about the patient's history and what matters in the physical exam. Based on what it learns, it chooses which tests to order. It reads the results and orders more if needed. Throughout, it updates its working hypotheses and plans, according to the research leaders.

That sequential design also changes how the tool is judged. Harsha Nori, head of AI at Microsoft AI Health, said the team measures more than accuracy. It also looks at how judiciously and carefully the system moves through a work-up.

Why does it matter?

Older systems are usually compared with doctors using multiple-choice questions. Dr. Dominic King, corporate vice president of the health unit at Microsoft AI, said AI now scores close to 100% on those. "But of course, that's not real-world," he said. "No patient ever came to see me with a nice summary."

That is a fair criticism of the field. A tool that must decide what to ask and which tests to order is closer to actual practice than one that reads a tidy case file. The cost finding is also notable. Diagnostic work-ups designed by the orchestrator cost less than those designed by human doctors, which speaks to a pressure every health system feels.

The context

For the Gulf, the timing is relevant. Saudi Arabia's Vision 2030 and the UAE's national health and AI strategies both put digital care and efficiency near the center of reform. Governments across the GCC are also managing rising chronic disease and growing demand for specialist care. A tool that can reduce unnecessary testing and support general physicians with difficult cases fits those goals well.

But this is one study, and the cases were selected for difficulty. Real clinics bring incomplete records, language barriers and patients who describe symptoms in their own way. Regulators in the region will want evidence from local populations and Arabic-language settings before any clinical use.

The next step is practical. Working with Beth Israel Deaconess Medical Center in Boston, the researchers will test whether the tool can triage patients in emergency department waiting rooms. That trial will say more about everyday value than any case series. Hospital leaders in the GCC should watch the results closely.