Google has published a study in Nature showing that its medical AI, AMIE, can do more than diagnose: it can help manage conditions over time. In a randomized, blinded comparison against 21 primary care physicians across 100 multi-visit scenarios, AMIE matched or exceeded the doctors on overall management reasoning, and scored higher than them on plan preciseness and alignment with clinical guidelines. The crucial caveat sits right next to the result: the patients were trained actors, not real people.
AMIE, short for Articulate Medical Intelligence Explorer, began as a diagnostic conversation system, and earlier work focused on the one-off encounter of figuring out what is wrong. The new result extends it to longitudinal disease management, the harder and less glamorous work of adjusting treatment, ordering the right follow-up tests, and prescribing across repeated visits. To do that, the system leans on drug formularies and authoritative clinical guidelines, and the study built its cases around UK NICE guidance and BMJ Best Practice.
Under the hood, AMIE for management is two agents working together: an empathetic dialogue agent that handles the real-time conversation with the patient, and a deep-thinking reasoning agent that cross-references hundreds of pages of clinical knowledge before laying out a plan. The comparison was blinded, with specialist physicians scoring the management plans from AMIE and from the human doctors without knowing which was which, across the hundred scenarios.
The limits deserve as much attention as the headline. These were professional patient actors in simulated, multi-visit consultations, which means the study captures the quality of clinical reasoning in a controlled setting, not real outcomes for sick people. It is research, not a product anyone can use, and Google is careful to frame it as something that could someday support physicians and give them more time, not replace them. Testing in real care is a separate and ongoing effort, including a nationwide randomized study of AI in actual virtual care. Matching a guideline on paper is not the same as managing a real illness in a real body.
Still, the direction matters. Diagnosis is a single moment, while management is the long, repetitive grind where most of medicine actually happens and where busy clinicians most often drift from the guidelines. An AI that is precise and guideline-aligned could, in principle, hand time back to doctors. It arrives the same week as OpenAI's claim that a model helped improve a real chemistry reaction, two data points in a broader push to aim frontier models at expert work. The same caution fits both: matched-in-a-study is a genuine signal, and the jump from a controlled comparison to messy reality is exactly the part neither has yet shown.