MAI-DxO (MAI Diagnostic Orchestrator) is a model-agnostic AI orchestrator developed by the Microsoft AI team, described in the paper "Sequential Diagnosis with Language Models" and the companion piece "The Path to Medical Superintelligence". It turns any language model into a virtual panel of physicians with diverse diagnostic approaches that collaborate to solve a case.
It works sequentially: starting from a short case abstract, it iteratively asks questions and orders tests from a gatekeeper model that reveals findings only when explicitly queried. The orchestrator proposes differential diagnoses, strategically selects high-value, cost-effective tests, runs a cost check and verifies its own reasoning before committing to a final diagnosis.
It was measured on the Sequential Diagnosis Benchmark (SDBench), which transforms 304 challenging New England Journal of Medicine clinicopathological conference cases into stepwise diagnostic encounters. Paired with OpenAI's o3 model, MAI-DxO reached 80% diagnostic accuracy (85.5% in a maximum-accuracy configuration) versus a 20% average for experienced physicians, while reducing diagnostic costs. The gains generalized across models from the OpenAI, Gemini, Claude, Grok, DeepSeek and Llama families. Microsoft AI stresses it is a research prototype, not a clinical product and not validated for clinical use.
