News Research Research and Evidence Ethics, Regulation, and Responsible Use Personalized Treatment

Conversational AI system tested in primary care

November 10, 2026 By Matthew Solan 5 min read
Share Share via Email Share on Facebook Share on LinkedIn Share on Twitter

A conversational artificial intelligence system completed supervised previsit clinical conversations without requiring any safety stops under prespecified criteria, according to a study published in The Lancet. Clinical evaluators also rated the system’s conversation quality favorably. 

The prospective feasibility study evaluated the Articulate Medical Intelligence Explorer (AMIE), a patient-facing artificial intelligence (AI) system, before urgent primary care appointments.  

“Our findings show that AMIE is conversationally safe within this physician-supervised, single-complaint, urgent-care setting study design,” the authors wrote. 

Study design and criteria 

Researchers enrolled patients at Healthcare Associates, an academic primary care practice at Beth Israel Deaconess Medical Center in Boston, from April through November 2025. Eligible patients were aged 18 years or older, had English listed as their primary language, and had a single urgent complaint that clinic staff had determined did not require emergency care. 

Patients needed an active patient portal account and access to a laptop or desktop computer. Of 114 patients who started an AMIE encounter, 100 completed it; 98 also attended the scheduled primary care appointment and were included in the analysis. 

AMIE conducted synchronous text conversations, gathered information about the presenting complaint, and discussed potential diagnoses. If requested, it also described the next steps the primary care provider (PCP) might discuss at the visit. The system had no access to the patient’s medical chart, physical examination findings, or office-based diagnostic tests. PCPs included attending physicians, nurse practitioners, and internal medicine residents working with attending oversight. 

A board-certified internal medicine physician monitored each interaction through a video call with screen sharing and debriefed the patient afterward. Prespecified stopping criteria included immediate concern for harm to self or others, substantial emotional distress related to the interaction, potential clinical harm identified by the supervisor, or a patient request to stop. Transcripts and automatically generated summaries were sent to PCPs before the appointment. 

Safety findings 

Physician supervisors identified one hallucination, in which AMIE incorrectly described a past surgery date as being in the future. Supervisors added clinical information in five cases. In three cases, supervisors reassured patients after AMIE included a malignancy among possible diagnoses. The other two cases involved clarifying symptoms to rule out a potentially emergent condition and clarifying when to seek emergency care. 

Patient and provider experiences 

Clinical evaluators rated conversation quality favorably in 87% to 100% of cases across 17 criteria. Patients gave favorable ratings in 48% to 96% of cases across 16 criteria, with a majority doing so for all but two: trust in information confidentiality and perceived honesty and trustworthiness. Patient attitudes toward AI became more favorable after the interaction and remained more favorable than at baseline after the PCP visit. 

PCPs completed surveys for 60 of 98 cases; in 44, they reported reviewing the AMIE transcript or summary before the appointment. Among those 44 cases, PCPs found AMIE helpful for visit preparation in 33 (75%) and said it might have changed their behavior in 25 (57%).  

One PCP rated an interaction as somewhat harmful because of possible patient anxiety after AMIE included lymphoma among potential diagnoses, although the supervisor had not observed pronounced distress during the chat. In three cases, PCPs raised concerns that AMIE’s output could lead them to settle on a diagnosis prematurely. 

Management-plan assessment 

In a secondary analysis, three independent clinical evaluators assessed AMIE and PCP management plans for each case in masked, randomized order. Aggregated ratings indicated equal quality in 47 of 98 cases, favored AMIE in 25, and favored PCPs in 26. Favorable or very favorable ratings were higher for PCP plans on cost-effectiveness (97% vs 85%) and practicality (100% vs 94%). The authors cautioned that the plans had been reformatted by a large language model for masked assessment and that PCPs had substantially more clinical and health-system context. 

Limitations and next steps 

Limitations included the modest sample, single-center setting, and lack of a control group for the previsit intervention. Participants tended to be younger and have higher health and technological literacy, limiting generalizability. Novelty, financial incentives, and real-time supervision could have inflated acceptability ratings. Safety oversight extended only to the AMIE encounter and might have missed distress before the appointment. The modified questionnaire used to assess attitudes had not undergone independent psychometric validation, and repeated administration could have influenced responses.  

The authors wrote that the evidence does not yet support autonomous use of patient-facing AI for diagnosis, triage, or management. They called for randomized trials assessing patient outcomes, diagnostic safety, patient-clinician communication, clinician workload, equity, cost, and the amount of oversight required. 

“Before evaluation as workload interventions, AI tools must first be shown to operate safely and acceptably in real care workflows,” the authors wrote. 

The authors acknowledged that Google funded the study. Researchers at Beth Israel Deaconess Medical Center designed the study and completed data collection. Analysis, interpretation, and manuscript preparation were conducted collaboratively with Google researchers. Numerous authors reported Google or Google DeepMind employment, Alphabet stock compensation, travel support, and patent interests. Adam Rodman, MD, disclosed a visiting researcher role at Google, other research support, and educational roles. Other authors reported additional interests, including research funding, honoraria, consulting or expert-testimony payments, advisory or leadership roles, and stock options.  All other authors declared no competing interests. 

AACE Endocrine AI is published by Conexiant under a license arrangement with the American Association of Clinical Endocrinology, Inc. (AACE®). The ideas and opinions expressed in AACE Endocrine AI do not necessarily reflect those of Conexiant or AACE. For more information, see Policies.

Related Content