Google reports early real-world feasibility for conversational diagnostic AI AMIE in primary care

What Google studied

On March 11, 2026, Google Research and Google DeepMind published results from what they describe as a first-of-its-kind prospective real-world study of AMIE, their conversational diagnostic AI, in partnership with Beth Israel Deaconess Medical Center. The system was used before ambulatory primary care visits to gather a patient history through a secure text chat and then generate a transcript and summary for the clinician. The work moves AMIE beyond simulated patient or benchmark settings into an actual care workflow, which is the threshold many medical AI projects have struggled to cross.

How the deployment worked

The study enrolled 100 adult patients with new, non-emergency complaints, and 98 of them later attended their scheduled appointment. Each AMIE interaction was observed live by a physician supervisor who could stop the session if one of four safety criteria was triggered, including signs of self-harm risk, significant distress, or potential clinical harm. Google says no safety stops were required. Before the patient saw the doctor, the system produced a pre-visit summary so the clinician could review the history in advance rather than starting from a blank chart.

What the results suggest

According to Google, blinded clinical evaluators rated AMIE and primary care providers similarly on overall differential diagnosis quality and on the appropriateness and safety of management plans, while clinicians still scored better on practicality and cost effectiveness. Google also reports that AMIE included the final diagnosis in 90% of cases and reached 75% top-3 accuracy, with strong performance maintained in the subset of cases later confirmed by diagnostic testing. Patient attitudes toward AI became more positive after using AMIE, and participating clinicians said the summaries improved visit readiness.

Why it matters

The announcement is significant because it adds prospective human-subject evidence to a field that usually relies on retrospective chart studies or controlled simulations. At the same time, Google emphasizes that this was a single-center feasibility study with a text-only interface, live physician oversight, and no controlled comparison against an alternative workflow. That means the paper is better read as evidence that supervised deployment can be safe and workable in a limited setting, not as proof that conversational diagnostic AI is ready to operate independently in routine care.