Microsoft's MAI-DxO Diagnoses 83% of Complex Cases in Science Study

· Cortex Solutions

A virtual panel of AI agents conferring over medical scans and lab tests in a modern diagnostic room

A Science study published Oct 22, 2026 says Microsoft's MAI-DxO diagnosed about 83% of complex NEJM cases versus 36% for doctors. What it shows, and its limits.

The peer-reviewed journal Science published a study from the Microsoft AI team on Thursday, October 22, 2026, titled "Sequential diagnosis with language models." It reports that its system, MAI-DxO (the Microsoft AI Diagnostic Orchestrator), correctly diagnosed about 83% of complex medical cases, compared with about 36% for generalist physicians who worked on the same cases individually with reference materials. Several tech outlets covered the news, including PBX Science in a report published on October 23, 2026, and AI Weekly, which covered the study on the day it was published. Both cited reporting by TechCrunch.

What did the study test?

According to PBX Science, the paper introduces a new benchmark called SDBench. Instead of a single multiple-choice question, it turns each case into a step-by-step medical encounter: the participant, human or AI, starts with a short patient summary, asks questions and orders tests, then commits to a diagnosis that is checked against the published outcome of the case.

  • The main test used 304 case records published in the New England Journal of Medicine.
  • A second test used 325 emergency department cases from Beth Israel Deaconess Medical Center in Boston.
  • The results come from simulations of past cases, not real patients in a clinic.

How does MAI-DxO work?

The system is not a new language model. It is an orchestration layer that runs several AI agents acting like a panel of doctors with different specialties. The system decides which questions to ask and which tests to order, and it revises its assessment as new information arrives until it reaches a diagnosis. It runs on top of models from different companies: AI Weekly reported that Microsoft tested it with models from OpenAI, Google, Anthropic, Meta and DeepSeek, and that the best results came with OpenAI's GPT-5 and GPT-5.5.

Where is the real value: the model or the orchestration?

This is the most important point in the coverage. According to TechCrunch, as quoted by both outlets, the study's own data suggest that much of the improvement comes from more powerful underlying models, with the orchestration layer adding gains on top:

  • With GPT-5.5, the system raised accuracy on complex cases from about 72% to nearly 87%.
  • On emergency cases the benefit was limited: both setups scored around 85%, according to PBX Science.
  • AI Weekly reported that the system lifted Gemini 2.5 Flash from 46.4% to nearly 69%, meaning the benefit was larger for weaker models.

Caveats to read alongside the numbers

The reports themselves put the results in context:

  • Cases this complex are usually handled by specialists or medical teams, who were not part of the comparison.
  • AI Weekly noted that physicians in the test were barred from using language models and search engines.
  • According to PBX Science, all 15 authors of the paper are Microsoft employees, and the system needs clinical validation before it can be used with patients, which Mustafa Suleyman, CEO of Microsoft AI, has acknowledged.
  • No commercial release date for the system has been announced.

The next step

Suleyman described the result as a step toward what he calls "medical superintelligence." Microsoft is working with Beth Israel Deaconess to evaluate the system in a real hospital setting, where patients receive standard care and those who consent also interact with MAI-DxO so its recommendations can be compared with those of clinicians. The reports also say the company is working with the Mayo Clinic to train a healthcare foundation model targeted for 2027. AI Weekly quoted Dominic King, VP of Health at Microsoft AI, saying that the biggest challenge in healthcare delivery today is the growing mismatch between the demand for and supply of medical expertise.

What does this mean for Saudi businesses?

The news is medical, but the lesson goes beyond hospitals:

  • Orchestrating several AI agents, each with its own role, is now a scientifically recognized approach, and the same idea can be applied to work such as reviewing requests, triaging tickets and analyzing documents step by step.
  • The numbers show that the quality of the underlying model makes most of the difference, so it is wise to build solutions that are not tied to a single provider and can switch models when something better appears.
  • Healthcare providers in the Kingdom, from hospitals and medical complexes to clinics, have an opportunity to trial decision-support tools, as long as they remain an aid to the doctor rather than a replacement and go through clinical validation, regulatory requirements and patient data protection.
  • The most important lesson: measure before you scale. A trial in a controlled setting is one thing; daily operation is another.

At Cortex we help companies design practical, well-considered AI automation solutions that fit how they work. Learn about our services: https://cortex.sa/services/ai-automation

Read next