With 85% success rate and 20% lower costs, Microsoft AI can diagnose complex cases more accurately than doctors

Tech company’s AI unit, led by British tech pioneer Mustafa Suleyman, is behind the development of the diagnostic system MAI-DxO, which aims to replicate the collaborative thinking of medical teams using a structured, role-based approach

Artificial intelligence is slowly but steadily reshaping how healthcare is delivered and how diagnoses are made. Microsoft is the latest tech player to stake a claim in this evolving space, unveiling a new Microsoft AI system that it says can diagnose complex medical cases with a level of accuracy that significantly outperforms human doctors.

The tool, called the Medical AI Diagnostic Orchestrator (MAI-DxO), was tested on 304 challenging patient cases from the New England Journal of Medicine. These cases are designed to challenge even experienced physicians. In a recent study, the AI correctly diagnosed up to 85.5% of them, a success rate over four times higher than that of generalist doctors with a median of 12 years of experience, who were limited to working without external resources, unlike typical clinical settings. The study also found that the AI achieved these diagnoses more efficiently, ordering fewer unnecessary tests.

The study also found that the AI made its diagnoses at a lower cost, ordering fewer unnecessary tests along the way.

The key to MAI-DxO’s performance lies in how it approaches diagnosis. Instead of relying on pre-filled case summaries or multiple-choice answers, the system works sequentially, more like a real doctor. It starts with a limited patient description and then decides what questions to ask or what tests to order, refining its thinking as new information comes in.

Also read: Type 2 Diabetes Before 30 can Cut Life Expectancy by 15 Years, Warns Lancet Report

To enable this, Microsoft developed the Sequential Diagnosis Benchmark (SDBench), a framework that evaluates both diagnostic accuracy and cost efficiency. The AI unit behind this work is led by British tech pioneer Mustafa Suleyman, who joined Microsoft in 2025 to spearhead its consumer AI efforts.

What makes MAI-DxO unique is its structure. Rather than being a single large language model, it acts more like a team of five virtual physicians, each playing a specific role: one maintains a list of possible conditions, another selects the most informative tests, one challenges assumptions, another ensures cost-efficiency, and one checks for consistency.

In one example from the study, a traditional AI model misdiagnosed a patient who had ingested hand sanitizer, ordered expensive imaging tests, and ultimately racked up over $3,400 in costs. MAI-DxO, by contrast, asked a key question early on that led it to the correct diagnosis, at a fraction of the price.

Overall, the system reduced diagnostic costs by 20% compared to human physicians and by 70% compared to OpenAI’s o3 model working alone.

Despite the strong results, Microsoft emphasizes that MAI-DxO isn’t designed to replace doctors. Instead, it’s built to act as an assistive tool, especially useful in settings where experienced specialists are in short supply. By offering a consistent second opinion and helping prioritize tests, AI tools like this one could help reduce misdiagnoses and improve efficiency across the healthcare system.

Still, like all medical AI systems, MAI-DxO faces hurdles before it can be widely adopted. Its performance in controlled studies may not fully translate to real-world clinics, where patient presentations vary. Issues like data privacy, regulatory approval, and ensuring equitable access must be addressed. Additionally, AI cannot replicate the human empathy essential to patient care.

Also read: Herbalife Diet Plans: An Evaluation of Safety and Effectiveness in India

That said, Microsoft’s latest research offers a glimpse into what the future of diagnosis could look like: a partnership between doctors and machines, combining the scale of data-driven reasoning with human intuition and care.