A software panel can now work through a medical case step by step, and on hard, published cases it reached the right diagnosis far more often than generalist doctors.
Microsoft published a peer-reviewed paper in Science describing MAI-DxO, a diagnostic orchestrator that coordinates multiple AI agents to propose diagnoses, choose questions and tests, evaluate new results, and then revise its hypotheses.
Why this matters: on complex cases drawn from The New England Journal of Medicine the system was correct about 83% of the time. The paper reports that generalist physicians, tested on the same cases, were correct far less often. That gap is large enough that this is not just an incremental improvement; it is a clear demonstration that sequential, team-like AI reasoning can change diagnostic performance in challenging cases.
How it works in plain terms: think of MAI-DxO as a small team of specialists. For each new symptom it calls in the few experts best suited to interpret that clue, decides the next question or test, and updates the diagnosis as evidence arrives. The process is deliberate and stepwise rather than a single guess.
What changes now: researchers and health systems have stronger, peer-reviewed evidence that multi-agent sequential reasoning can boost accuracy on hard cases. That makes clinical trials and regulatory review the sensible next step, because retrospective performance on curated cases is not the same as safe use in live hospitals.
What’s next: the open question is whether independent teams can reproduce these results in real clinical workflows and whether regulators and clinicians will adopt this style of AI reasoning at scale. If they do, the next few years could reshape how difficult diagnoses are handled.
