Microsoft's MAI-DxO system runs AI agents like a team of doctors

Summary

Microsoft published a Science paper demonstrating that its MAI-DxO multi-agent AI system mimics a panel of physicians. The system achieved 83% diagnostic accuracy on complex NEJM cases, whereas general practitioners working alone remained at 36%. However, specialists were not included in the comparison, and all results were derived from simulations rather than real patients.

Why it matters

This study is one of the highest-profile efforts to position AI not as a single source of prediction, but as a collaborative diagnostic team. The real-world trial planned at Beth Israel Deaconess will be the first serious test of whether the gains observed in simulation hold up when interacting with actual clinical workflows. A cost reduction of approximately $3,000 per case, combined with a modest increase in accuracy, could make multi-agent diagnosis economically attractive at scale if the results remain consistent.

Highlights

• 83% accuracy in complex cases versus 36% for general practitioners working alone; however, specialists and full multidisciplinary teams were not included in the comparison
• TechCrunch concludes that much of the improvement stems from the underlying large language models (GPT-5/5.5) rather than multi-agent orchestration
• Real-world validation is planned at Beth Israel Deaconess; a health foundation model co-developed with Mayo Clinic is expected to be ready within the year

Critical take

The 83% figure in the headline is misleading because it excludes specialists and multidisciplinary teams, which are the true clinical standard for complex cases. Furthermore, the entire evaluation is a simulation based on previously published cases, not forward-looking patient data. Moreover, a marginal increase of only four points compared to using GPT-5 alone, coupled with a cost of approximately $5,400 per case, raises serious questions about whether the multi-agent 'team' architecture adds meaningful value beyond simply deploying a stronger foundation model.