Google's AMIE took histories from 100 real patients, and supervising doctors never had to stop a chat
In a study at Beth Israel Deaconess Medical Center, a conversational AI interviewed patients by text before urgent care visits. A physician watched every chat live. Doctors still made the decisions.
- Organization
- Google and Beth Israel Deaconess Medical Center
- Industry
- Healthcare
- When
- Mar 2026
- Evidence
- Research paper
- Agents and tools
- AMIE
Reported outcome
Zero safety stops across 100 patient chats. AMIE's list of possible diagnoses included the final diagnosis in 90% of cases, with 75% top-3 accuracy.
- 100Patients who completed a chat
- 0Safety stops by supervisors
- 90% of casesFinal diagnosis in AMIE's list
- 75%Top-3 accuracy
- 56% of casesFinal diagnosis as AMIE's single most likely
Caveats
Described in a paper with a stated method. Peer review status varies.
- This was a feasibility study at one hospital with one group and no control group. Google says it does not support claims about efficacy against a normal workflow.
- A physician watched every chat live by video call. The study does not show that AMIE is safe without that.
- Primary care providers did better than AMIE on the practicality (p = 0.003) and cost effectiveness (p = 0.004) of management plans.
- The chat was text only. Google built AMIE and co-ran the study.
- We read the arXiv preprint and Google's blog post. Google says the work is now published in The Lancet; we did not open that version.
What the agent does
AMIE (Articulate Medical Intelligence Explorer) is a conversational AI built on a large language model. In this study, adult patients opened a secure web link up to 5 days before an urgent care appointment. AMIE took their clinical history by text chat and presented possible diagnoses for them to discuss with their provider.
A physician, called the AI supervisor, watched each chat live through a video call with screen sharing. The supervisor was trained to stop the chat if one of four safety criteria was met. Before the visit, the clinician received a transcript and a summary.
What happened
No supervisor had to stop a chat. Patients reported high satisfaction, and their attitudes towards AI improved after the chat (p < 0.001). Providers found the output useful for preparation.
A chart review 8 weeks later set the final diagnosis. AMIE's list included it in 90% of cases. Blinded reviewers rated AMIE and the providers as similar on overall diagnosis and plan quality. The providers did better on practical and cost-effective plans.
What to copy
- Test with real patients only under live human supervision and clear stop rules.
- Let the AI gather information before the visit. Let the clinician decide.
- Measure against the outcome weeks later, not against the AI's own confidence.
Starter prompt
Written by us, not by Google. This is not a medical device. Do not use it with real patients without clinical and ethics approval.
You collect a pre-visit history for a clinician. You do not diagnose or give treatment advice to the patient.
Ask one question at a time about: main complaint, timeline, severity, related symptoms, medicines, history.
If the patient describes an emergency sign, stop and tell them to call emergency services.
At the end, write a summary for the clinician: history, a list of possible causes with reasons, and open questions.
Sources
- A prospective clinical feasibility study of a conversational diagnostic AI in an ambulatory primary care clinic — arXiv, Mar 9, 2026
- Exploring the feasibility of conversational diagnostic AI in a real-world clinical study — Google Research, Mar 11, 2026
Added Nov 16, 2026 · Sources last checked Nov 16, 2026.
Get new use cases weekly
One short email. Sources included.
The newsletter is paused. Past issues stay in the archive.
Related use cases
Cohere Health's policy agent cut the time to digitize a clinical policy from 2 h 15 min to 1 h 35 min
Overall time spent on policy digitization fell 30 percent, from 2 hours 15 minutes to 1 hour 35 minutes per policy. Agent deployments went from 3–4 months to 2–6 weeks.
Bluesight's Prism Assistant: 20 health systems, recurring drug diversion reports in 15 minutes, not six hours
By Bluesight's internal measurements across 20 health systems: recurring reports 96 percent faster (about six hours to 15 minutes) and triage time down 90 percent.
UTMB cardiologists review an AI pre-screened list; as many as a third of heart failure patients flagged
In the first month, the platform showed that as many as a third of heart failure patients had opportunities for better guideline-directed medical therapy.