Microsoft Says Its Multi-Agent Diagnostic System Beats Doctors on Hard Cases
Microsoft AI chief Mustafa Suleyman said on October 22, 2026 that the company's diagnostic system outperforms physicians on difficult cases, and he described the result as a step toward “medical superintelligence.” TechCrunch
The system, MAI-DxO, coordinates several AI agents, which are software components that carry out multi-step tasks by calling models and tools without a person directing each step. Here, the agents act as a panel of doctors. They question a patient, order tests, and debate a diagnosis before reaching a conclusion.
In a simulation built on complex cases from the New England Journal of Medicine, MAI-DxO was correct 83% of the time. Primary care doctors and general internists working individually with reference materials were correct 36% of the time.
Microsoft ran MAI-DxO on frontier models, the largest general-purpose AI systems, from OpenAI, Google, Anthropic, Meta, and DeepSeek. The best results came from OpenAI's GPT-5 and GPT-5.5. With GPT-5.5 on complex cases, accuracy rose from 72% to nearly 87%, at a slightly lower cost.
Suleyman said Microsoft is working with the Mayo Clinic to train a foundation model for healthcare, with release planned for 2027. The Mayo Clinic work is a plan, not a shipped product.
The MAI-DxO results sit within a wider effort. Reuters reported that Microsoft is forming a new team aimed at building artificial intelligence vastly more capable than humans in certain domains.
In my view, the numbers need to be read for what they measure. A case-based simulation tests diagnostic reasoning on a fixed set of published, well-documented cases. Clinical practice involves incomplete histories, conflicting test results, patients who describe symptoms poorly, and decisions with real consequences. The gap between a benchmark score and bedside performance is where most clinical AI efforts have run into difficulty, and the figures in this announcement do not close that gap.
The comparison also deserves care. The physicians in the study worked alone with reference materials, which is not how most diagnosis happens in a hospital, where colleagues, specialists, and the electronic health record all contribute. The result shows what a multi-agent system can do under these conditions. It does not show how the same system would fare against a team of clinicians working together.
The model comparison deserves a note as well. Because MAI-DxO ran on several frontier models, the orchestration layer appears to be designed to be model-agnostic, with the underlying model as a swappable component. The reported gain from GPT-5.5 suggests the scaffolding improves with each model generation, and that the cost per correct diagnosis is falling along with it.
The broader context here is a shift in what clinical AI is being asked to do. Earlier tools summarized notes or flagged anomalies on images. A system that questions a patient, orders tests, and weighs competing hypotheses is closer to a reasoning partner in the diagnostic process. If that reasoning holds up under prospective clinical validation, the practical benefit could be fewer missed diagnoses in complex cases, particularly where specialist access is limited. That outcome has not yet been demonstrated in clinical settings. The path from simulation to clinical use runs through regulatory review, clinical trials, and integration into existing workflows.
For engineers and product leaders, the open questions are concrete. How the system handles uncertainty, how it explains its differential, and how often it orders unnecessary tests will matter as much as its top-line accuracy. Accuracy and cost alone do not describe a clinical product.
