The Weekly Dispatch
By Decotechs · Published weekly

Microsoft’s AI Diagnoses 2–3× Better. What Hospitals Miss
Decotechs
October 26, 2026
On October 22, 2026, a peer-reviewed study in Science showed Microsoft’s MAI-DxO system correctly diagnosing complex cases 83% of the time. Generalist physicians using textbooks and the internet managed 36%. Without those tools, physicians dropped to 22%.
That is the headline. The more important story is what the study actually tested—and what most hospitals are still overlooking.
What the study really measured
Most earlier AI diagnosis tests handed models a complete case summary and asked for a single answer. Real medicine does not work that way. Patients arrive with limited information. Clinicians decide what to ask next and which tests are worth ordering.
Microsoft’s team built the Sequential Diagnosis Benchmark (SDBench) to match that reality. Both the AI and the physicians started with a short vignette and had to request every additional detail or test.
The system, called MAI-DxO (Microsoft AI Diagnostic Orchestrator), works like a panel of virtual clinicians. Different agents propose diagnoses, challenge one another, weigh the cost of each test, and update the working differential as new evidence arrives.
Results from the paper and Microsoft’s own summary:
- Complex NEJM cases: MAI-DxO 83% accurate vs 36% (physicians with references) and 22% (without).
- Emergency-department cases from MIMIC-IV-ED: MAI-DxO 80% vs 47% for physicians.
- In cost-optimized configurations, the system maintained strong accuracy while ordering fewer or lower-cost tests.
Sources: Microsoft AI blog (October 22, 2026) and Reuters (October 23, 2026).
Why the comparison is both impressive and limited
The physicians in the study were generalists, not specialists who normally handle these complex cases. The encounters were text-only—no physical exam, no imaging, no bedside cues. The cases themselves came from published records and a curated emergency dataset.
These limitations matter. They do not erase the result. They do change how hospitals should interpret it.
The real advance is not “AI beats doctors.” It is that a sequential, multi-agent approach can outperform single-pass models and individual generalists on step-by-step workups while paying attention to cost.
What hospitals are still missing
Many health systems are still evaluating AI the old way: feed it a complete chart note and measure final-diagnosis accuracy. That misses three practical issues.
- Information gathering is half the work. The ability to decide the next question or the next high-value test is often more useful than a polished final answer.
- Cost and accuracy can move together. In several configurations, the orchestrator reached higher accuracy at lower estimated test cost. That combination is rare and valuable under value-based care pressure.
- Workflow fit beats model score. An 83% accurate system that sits outside the EHR and requires extra clicks will be used less than a 70% accurate tool that lives inside the existing workflow and logs every action.
Microsoft has already announced the next step: a prospective study with real patients at Beth Israel Deaconess Medical Center. That will test the harder questions—safety, workflow friction, and liability.
A mini example of the difference
Imagine a patient who presents with fatigue, mild weight loss, and an abnormal lab. A single-shot model might output a ranked list of possibilities. A sequential system asks targeted questions, requests the highest-yield lab or imaging first, revises its differential after each result, and stops when the cost of more testing outweighs the expected gain in certainty.
That process looks closer to how experienced clinicians actually work. It also creates an auditable trail of decisions—something hospitals will need as regulatory scrutiny of generative AI increases.
What you can do this week
- Map one high-volume diagnostic pathway in your organization (for example, undifferentiated fatigue workups or common ED presentations). Note where information is incomplete at the start and where test decisions are made.
- Ask vendors for sequential, not just final-answer, performance data. Request how their system decides the next question or test and how it records those steps.
- Review current AI governance documents for language about multi-agent or sequential systems. Most policies still assume single-model, single-output tools.
- Identify one narrow use case where a controlled AI agent could support—not replace—clinical reasoning (for example, suggesting high-yield next tests for a specific complaint while keeping the physician in the loop).
- Measure the full cost of the current process, not just diagnostic accuracy. Include time spent hunting for information and the downstream cost of low-yield testing.
Soft next step
At Decotechs we help healthcare organizations and health-tech teams design practical AI agents and automation that fit real clinical workflows—starting with clear scope, auditability, and measurable outcomes. If you are evaluating sequential or multi-agent approaches, we are happy to share what we are seeing work in the field.
What is the biggest gap you see between benchmark AI results and day-to-day diagnostic work in your setting?

The Weekly Dispatch
Decotechs' weekly newsletter on AI agents, automation, and software engineering insights.




