IA · 23 October 2026 · 4 min read
Diagnosis as an Agent Board: Microsoft's DxO Outperforms Doctors on Complex Cases
In brief: Microsoft has published results in Science detailing MAI-DxO, a multi-agent AI framework capable of diagnosing complex medical cases with 83% accuracy, compared to 36% achieved by individual generalist physicians. The platform coordinates specialized agents that interview patients, order tests, and debate differential diagnoses. Microsoft AI CEO Mustafa Suleyman framed the milestone as a foundational step toward domain-specific medical superintelligence.
by Team Mocchi's
A virtual medical board tested on the NEJM
Diagnosing rare conditions and convoluted clinical presentations requires collaborative deliberation, disciplined methodology, and the ability to order only the most informative diagnostic tests. Replicating that workflow, Microsoft's AI organization—led by CEO Mustafa Suleyman—has developed MAI-DxO (Diagnostic Orchestrator), a multi-agent framework whose peer-reviewed validation was just published in the journal Science.
Tested against simulated cases sourced from the New England Journal of Medicine's famed Clinicopathologic Conferences (CPC), the system reached the correct diagnosis 83% of the time. In identical testing conditions, licensed general practitioners and internal medicine physicians working individually with full access to clinical reference materials scored 36%. As reported by TechCrunch, this performance gap indicates that structured multi-agent orchestration enables complex, multi-hop medical reasoning that single-turn prompting has consistently failed to deliver.
From inquiry to verdict: how MAI-DxO works
Rather than relying on a single large language model generating an end-to-end response, DxO mimics a multidisciplinary panel of specialists. The diagnostic workflow is segmented into distinct operational stages: an initial cluster of agents takes the patient's history by asking targeted questions to a simulated patient, a second group selects and orders appropriate laboratory and imaging tests, and an analytical board deliberates on conflicting hypotheses before certifying the final diagnosis.
Crucially, the study's data reveals that while underlying model capability matters—Microsoft evaluated foundation models from OpenAI, Google, Anthropic, Meta, and DeepSeek, with top marks achieved on GPT-5 and GPT-5.5—it is the orchestration layer that drives the diagnostic leap. The multi-agent debate enforces procedural discipline, cuts out redundant diagnostic workups, and guards against premature clinical closure.
The race toward medical superintelligence and study limitations
Mustafa Suleyman characterized the breakthrough as a tangible step toward what he terms «medical superintelligence». In an interview with The Verge, Suleyman noted that Microsoft aims to deliver tools within the next few years that are practically indistinguishable from human clinical excellence. Dominic King, vice president of Health at Microsoft AI, highlighted the practical stakes, pointing to the acute global mismatch between skyrocketing healthcare demand and a shortage of specialized clinical labor.
Clinical researchers, however, emphasize the need to interpret the benchmarks carefully. The simulation compared a collective multi-agent system against solitary primary care doctors, whereas in real-world hospitals NEJM-grade cases are referred to multidisciplinary tumor boards and expert teams. Furthermore, transitioning from curated medical transcripts to clinical practice introduces non-standardized noise, missing records, and bedside complexities that structured simulations cannot fully replicate.
Mocchi's take
Microsoft's findings reinforce an architectural principle we observe daily in enterprise AI deployments: in high-stakes domains, monolithic prompts inevitably yield to modular, multi-agent frameworks featuring checks and balances. For European healthtech and life sciences organizations, this shift clarifies strategic priorities: competitive advantage will not stem from hosting the largest raw parameters, but from designing resilient, verifiable decision workflows. The goal is not displacing clinicians, but providing them with diagnostic co-pilots that mitigate cognitive fatigue and eliminate diagnostic blind spots.