9 October 2026 · Medows
AMIE in Clinic: 98 Patients, 90% Match
A Lancet study put Google's AMIE chatbot in a real clinic with 98 patients. What the 90% match does and does not tell a doctor on rounds.
A diagnostic chatbot just left the lab and walked into a real clinic. The result is encouraging. It is also small, and the details matter more than the headline.
What happened
On October 8, 2026, Google researchers and Beth Israel Deaconess Medical Center (BIDMC) published a study in The Lancet on AMIE, a research diagnostic AI chatbot. According to Google's own write-up, 98 patients talked to AMIE before urgent care visits at BIDMC's ambulatory primary care clinic. Supervising physicians watched the interactions in real time.
Three figures from that write-up stand out:
- "Not a single conversation needed to be interrupted based on predefined safety criteria."
- AMIE's differential diagnoses matched the doctors' final diagnoses 90% of the time.
- Clinicians said AI summaries helped them prepare for visits in 75% of cases, and influenced their approach to care in more than half.
Why this is more interesting than a benchmark
Most clinical AI news is a score on an exam. This one is a live clinic, real patients, and a doctor in the loop. That is the setting that counts.
The most useful number is not the 90%. It is the 75%. The value reported is that the doctor walked into the room already oriented. The AI did not replace the encounter. It shortened the cold start.
Anyone who has started a ward round knows the cold start. You open a chart, scroll, piece together the story, and only then begin to think. A summary that is ready before you arrive changes where your attention goes.
What it does not show
Be a skeptic here. Google says it plainly: larger clinical trials are needed to assess patient-facing AI at scale. Ninety-eight patients is a first look. The lead authors include the company that built the system. A 90% match with the final diagnosis also says nothing yet about the 10%, or about what happens when a differential is wrong and nobody is watching.
And supervision was part of the design. Physicians monitored in real time. That is a safety feature of the study, and it is not guaranteed in every deployment.
The workspace angle
The pattern we keep seeing is that the AI earns its place when it feeds the doctor's thinking and stays visible to the doctor. Pre-visit summaries, sourced and checkable, fit that. A black box that hands over a conclusion does not.
For a doctor on rounds, the question is not whether a model can match a diagnosis. It is whether you can see why, check it quickly, and stay in charge. That is the bar we hold ourselves to at Medows: verifiable output, inside the workflow you already use, with you making the call.
What to watch next
Watch for the larger trials. Watch for independent replication outside the developer's own team. And watch how much of the benefit comes from the chatbot's reasoning, and how much comes simply from having a structured history ready before the visit.
Sources
- Google, "Study in The Lancet suggests AI could improve patient-physician relationships": https://blog.google/innovation-and-ai/technology/health/amie-clinical-study-lancet
Medows is a clinical AI workspace for the doctor on rounds. Learn more or write to alapan@medows.ai / alapanx@gmail.com.