Lancet study finds AMIE needed no safety stops in 98 real urgent care conversations

A study published in the Lancet on 8 October 2026 tested a patient-facing conversational AI system, the Articulate Medical Intelligence Explorer (AMIE), with real patients before urgent primary care appointments, and found no conversation had to be stopped on safety grounds ¹. The work was funded by Alphabet ¹.

What the study found

The design was a prospective, single-centre, single-arm feasibility study ¹. English-speaking adults spoke to AMIE up to 5 days before an appointment for a single complaint ¹. Physician safety supervisors watched every interaction and had been trained to step in against predefined safety criteria, and the primary care provider (PCP) received the AMIE transcript and summary before seeing the patient ¹.

Between April and November 2025, 114 patients enrolled and 98 completed both the AMIE conversation and the appointment ¹. Supervisors needed to halt none of the conversations under the prespecified criteria ¹. They did record one hallucination, and they added clinical information in five interactions ¹.

Clinical evaluators rated the conversations favourably in 87% to 100% of cases across 17 criteria ¹. Patients were less uniform, rating them favourably in 48% to 96% of cases across 16 criteria ¹. Patients' attitudes towards AI improved after using AMIE and stayed higher after they had seen their physician ¹.

The clinician side of the evidence is thinner. PCPs returned post-visit surveys in 60 of the 98 cases, and in 44 of those they had reviewed the transcript beforehand ¹. In that group, they judged AMIE helpful for preparing the visit in 33 of 44 cases and said it might have changed what they did in 25 of 44 ¹.

What it means for the market

For health systems weighing AI that talks to patients before a clinician does, the study tested AMIE in real-world urgent primary care appointments ¹. The result supports safety and acceptability under close supervision, which is a narrower claim than safe unsupervised use. Every conversation in the study had a physician watching in real time, and the single hallucination and five clinical additions show why that layer was there ¹.

The authors describe the work as showing "initial feasibility" and say further research is needed before clinical translation ¹. The study did not measure diagnostic accuracy, clinical outcomes, appointment length or cost, so buyers and regulators have no evidence yet on whether pre-visit AI history-taking saves clinician time or changes care ¹.

Caveats

The study ran at one centre, had no control arm and enrolled 114 patients, all English-speaking adults with a single complaint ¹. Clinician feedback covers 60 of 98 visits, and the 44 cases with a reviewed transcript are a small base for the helpfulness figures ¹. Alphabet funded the study ¹. The trial is registered as NCT06911398 ¹.