Microsoft's diagnostic AI beats doctors in study: what it means for India
Microsoft's multi-agent Diagnostic Orchestrator scored 83 percent against doctors' 36 percent in a Science study. Here is why Indian hospitals and regulators should pay attention.
The News
Microsoft says a multi-agent diagnostic system built by its AI division correctly identified conditions in around 83 percent of complex cases, against roughly 36 percent for generalist physicians who were allowed to consult references and 22 percent for those working without them. The results appeared on Thursday in the journal Science.
The system, called the Diagnostic Orchestrator, splits a case across several AI agents that behave like a panel of specialists. Together they probe the history, choose tests and settle on a diagnosis. The test cases came from hundreds of reports published in The New England Journal of Medicine, packaged into a Microsoft benchmark named SDBench that the company intends to open-source.
A second test used emergency department admissions from the Beth Israel Deaconess Medical Center in Boston. There the Orchestrator scored 80 percent accuracy, while human physicians averaged 47 percent. Microsoft also says the system usually reached its answer at a lower estimated cost. It runs on OpenAI's GPT-5, and the company says it beat frontier models working alone.
Why It Matters
The method is the real signal. Rather than asking one model for an answer, Microsoft wraps models in an orchestration layer that copies the clinical workflow of history, hypothesis and testing. The researchers say later checks suggest that "improvements in model capability do not necessarily diminish the benefits" of agents working together. If that holds, the value in health AI may shift from whoever owns the best model to whoever builds the best workflow around it.
Caution is warranted. The system has only been tested on historical cases, and Microsoft AI chief executive Mustafa Suleyman concedes "it needs clinical validation". Health AI has overpromised before: IBM spent years marketing Watson as a cancer-care adviser before selling its Watson Health assets to Francisco Partners in 2022. Dr Dominic King, Microsoft's AI health lead and a coauthor, says hospital trials with Beth Israel Deaconess will likely begin "in the next couple of months", with no firm timeline for commercial release.
Indian Angle
For India, the more relevant comparison is not AI against a specialist in Boston, but AI against no doctor at all. Primary health centres in many rural districts run short of physicians, and a cheap triage layer that narrows a diagnosis and avoids unnecessary tests would matter most in exactly those settings. The cost finding matters as much as accuracy, since many Indian patients pay for care out of pocket.
Indian companies are already working in adjacent areas. Mumbai-based Qure.ai sells AI for reading chest X-rays and CT scans, and Bengaluru's Niramai focuses on breast cancer screening. A multi-agent reasoning layer from a global hyperscaler could either become a platform these startups build on or a direct competitor, depending on how Microsoft packages it for its Azure customers.
Regulation will decide the pace. Diagnostic software can fall under the Medical Devices Rules, 2017, overseen by the CDSCO, while patient data handling sits under the Digital Personal Data Protection Act, 2023. ICMR's 2023 ethical guidelines for AI in healthcare already stress human oversight, which lines up with Suleyman's view that clinicians keep a central role.
FAQ
Is the Diagnostic Orchestrator available to hospitals now?
No. Microsoft has not said when or how it will be sold. The next step is a hospital trial with Beth Israel Deaconess Medical Center in Boston, which Dr Dominic King says should start within a couple of months.
How accurate was the system compared with doctors?
On cases drawn from New England Journal of Medicine reports, it scored around 83 percent, against roughly 36 percent for generalist physicians with references and 22 percent without them. On Boston emergency department cases it reached 80 percent, compared with an average of 47 percent for human physicians.
Could Indian hospitals use a tool like this?
Potentially, but not soon. Diagnostic software would likely need CDSCO clearance under the Medical Devices Rules, and patient data would fall under the DPDP Act, 2023. Hospitals already digitising records under the Ayushman Bharat Digital Mission would be natural early adopters once a clinically validated version exists.
Does this mean AI will replace doctors?
Microsoft says no. Suleyman argues human clinicians should always keep a central role in patient care, and hopes the technology removes routine drudgery so doctors spend more time with patients.
This story was reported by The Verge. Read the full original coverage at The Verge.