The cases were chosen to be hard. In a study published in Science, a Microsoft system called MAI-DxO reached the correct diagnosis in 83% of the complex cases it was given in a simulated clinical setting. The general practitioners used as a comparison group reached 36%.
On the presentations that fill U.S. emergency departments, the gap narrowed but did not close: the system’s accuracy was about 1.7 times that of the physicians it was measured against. The cases came from the New England Journal of Medicine’s published case library and from MIMIC-IV-ED, a database of emergency-department records.
Two details of the design matter more than the headline number. The comparison group was made up of general practitioners, not specialists. And the setting was a simulation, where the system was given the case material and allowed to order tests, rather than a ward where a patient is deteriorating and a family is asking questions.
MAI-DxO is an orchestrator rather than a single model. It runs a panel of models through a differential diagnosis, then has a separate component decide which test to order next, weighing the information a result would produce against its cost. The sequence repeats until the system commits to an answer. It is a rehearsal of clinical reasoning rather than a single query to a chatbot.
That architecture is why the results interest clinicians who have grown tired of benchmark claims. A language model asked to name a diagnosis from a case summary is doing recall. A system that decides to order a blood gas before a scan is doing something closer to triage, and its mistakes are the kind a hospital can audit.
Mustafa Suleyman, who runs Microsoft’s AI division, has described the objective in larger terms, saying the goal is medical superintelligence. The phrase will irritate physicians who watched earlier generations of clinical software promise decision support and deliver alert fatigue. It also sets an expectation the technology cannot meet on the timeline the phrase implies.
The caveats are substantial and mostly acknowledged in the paper. A simulated case library is not a patient. Published journal cases are selected for difficulty and narrative interest, and they have been in the training data of major models for years, which raises a contamination question the authors had to address. General practitioners working alone are also not how hospitals make hard diagnoses; a tumor board or a specialist consult is.
Cost is the quieter issue. The system’s willingness to order tests is an advantage in accuracy and a liability in a health system where imaging capacity and laboratory budgets are finite. A model that reaches the right answer by ordering eleven tests has produced a correct diagnosis and a bill, and the study design rewarded the first without pricing the second.
Then there is the question of what happens when the system is wrong. A tool that is right 83% of the time on hard cases is useful; the other 17% is where liability lives. Microsoft has been careful to describe the work as research. Regulators treat software that informs diagnosis as a medical device, and clearance for a system that sequences its own tests would be a longer conversation than a journal publication.
Hospitals have reasons to be interested anyway. Emergency departments are short of physicians, waiting times are a political issue in several countries, and a tool that helps a generalist think like a specialist addresses a real bottleneck. The 1.7-fold figure on common emergency presentations is the number administrators will remember, because that is the volume case, not the exotic one.
The competitive context is a race among the largest technology companies to become the default infrastructure for clinical work. Google has published medical reasoning work built on its models. Amazon and Oracle sell documentation and claims tools into hospitals. A long list of startups sells AI scribes. Microsoft’s advantage is distribution inside hospital IT departments that already run its software.
The strategic puzzle for Microsoft is the one it faces across its AI business. It invested heavily in OpenAI’s models and is now building its own, and an architecture that routes a case across several models is a way of depending on no single supplier. A hospital buying the system is buying Microsoft’s orchestration, not a particular model’s weights.
Suleyman’s ambition has a commercial shape as well. A diagnostic orchestrator that orders tests touches the revenue cycle, the laboratory and the pharmacy, which is where hospital software budgets sit. A model that only drafts notes is a feature.
Physicians who reviewed the design raised a methodological point that will shape the next study: the comparison should be against specialists, working with the same information and the same test budget. Until that runs, the 83% and the 36% describe two different tasks as much as two levels of skill.
The paper’s most durable contribution may be the architecture. Treating diagnosis as a sequence of information purchases, each with a price, makes a model’s reasoning legible to a hospital finance committee. Whether that becomes a product, a pilot or a cautionary tale depends on the next study, and on whether anyone runs it in a ward.
