Microsoft has developed an AI-powered diagnostic system that achieves a success rate of 85.5 percent in medical case studies—more than four times that of the human comparison group. The "AI Diagnostic Orchestrator" was first tested with 304 particularly complex patient cases from the New England Journal of Medicine. In comparison, experienced doctors without access to tools correctly diagnosed about 20 percent of the cases.

The new technology is the first project of Microsoft's Health AI unit led by Mustafa Suleyman, who previously co-founded the AI lab DeepMind. At the center of the system is an "orchestrator," which organizes multiple specialized AI agents like a virtual medical team. These agents develop hypotheses, select tests, and discuss the best diagnosis based on a new procedure called "chain of debate.

Large language models from OpenAI, Meta, Google, Anthropic, xAI, and DeepSeek were used. While all models benefited from orchestration, the OpenAI model "o3" achieved the best performance. In addition to high accuracy, it was also noted that the system was cost-optimized: it saved hundreds of thousands of dollars in the simulation through targeted test selection.

Dominic King, former head of DeepMind Health, considers the project to be the most powerful system in the field of medical AI to date. At the same time, he emphasized the experimental nature: The results are not peer-reviewed and not currently approved for clinical use.

Nevertheless, the study underscores the disruptive potential of generative AI in healthcare—especially against the backdrop of global supply shortages and rising medical costs. Suleyman announced plans to integrate parts of the technology into Microsoft's Copilot and Bing, which process around 50 million health queries daily.

Despite the close cooperation with OpenAI, Microsoft pursues a model-independent approach. "In the long term," Suleyman said, "large language models will become commodities – the real value lies in their intelligent coordination.