Microsoft researchers have developed an AI tool called MAI-DxO that diagnoses patients step by step, using an orchestrator to coordinate questions, exams, and tests, and it solved complex medical cases with more than twice the accuracy of general physicians in the study.

However, the validation is currently based on historical cases and simulations; real-world clinical testing in live hospitals is pending, with plans to test at Boston’s Beth Israel Deaconess Medical Center in the coming months.

The system updates diagnostic hypotheses as new information emerges and chooses the next steps in evaluation, following a sequential workflow rather than simply combining data.

Industry caveats include that newer AI models (GPT-6 Astra, Muse Wonder Full, Gemini 4 Argon, Opus 5.5) were not evaluated in the study, and regulatory and compliance considerations will shape future rollouts.

The article discusses potential impacts on junior clinicians’ learning and stresses maintaining learning value through coaching, simulations, and rotations alongside automation.

Beyond healthcare, the study notes potential applications in fields like insurance claims and IT support, and cites research indicating AI will more likely augment high-skilled work, while some administrative roles remain vulnerable.

Ongoing exploration of healthtech developments and upcoming healthcare conferences provide broader context for the field.

Reuters contributed to the report, and statements from Microsoft leadership emphasize AI safety and regulatory considerations in the technology’s deployment.

Microsoft’s health AI lead, Dr. Dominic King, underscores regulatory and compliance requirements for responsible rollout and notes that live-site deployment timelines remain uncertain.

The study stresses that widespread deployment is premature and highlights the need for clinical validation before hospital use.

Regulatory and governance themes are highlighted, calling for stricter rules and clinical testing controls before deploying AI in sensitive health applications.

The study notes real-world performance is untested, the work used simulations with no real patients, and much of the gains come from the underlying AI models (GPT-5).