A prospective randomized study of patient-centered AI for clinical visit preparation
A · Strong evidence Google Research prospective randomised study with blinded clinician review and patient surveys, prompts published; strong design for a consumer AI tool, though funded by the maker, run in its own app and measuring preparation rather than health outcomes.
Summary based on the full text·Original: about 50 minutes to read
In short
A preprint (arXiv, 30 October 2026) by Xincheng Tan and colleagues at Google Research. Participants preparing for a doctor's visit in the Google Health app were randomly assigned either a structured agent (gather context, set expectations, give relevant information, build a prioritised question list, write a summary) or a minimally prompted chatbot on the same model, Gemini 2.5 Flash. The structured agent did better on both clinician ratings and patient surveys, with similar safety. Clinician and patient scores did not correlate.
Key numbers
- 802US adults randomised across a structured visit-preparation agent (two variants) and a plain chatbot on the same model
- 67.2% vs 50.6%completed sessions rated comprehensive by blinded clinicians, agent versus control (n=272)
- 73.8% vs 57.0%patients reporting they felt prepared for their visit, agent versus control (n=460)
- 98.9% vs 90.4%sessions rated aligned with the patient's health context, agent versus control
- 42.3%included conversations that ended before completing the preparation flow
As reported by the source.
English·Web page·opens arxiv.org
What it reports
- 802 adults held 866 conversations across three arms (standard agent, concise agent, control).
- Blinded clinicians, completed sessions: comprehensive preparation 67.2% versus 50.6% (P < 0.05); alignment with the patient's health context 98.9% versus 90.4% (P < 0.01). Medical safety and appropriateness stayed high in all arms.
- Patient post-surveys: felt prepared 73.8% versus 57.0%; confident about what to ask 81.4% versus 67.1%; satisfaction 74.0% versus 62.0%.
- No survey item correlated significantly with any clinician rating, so the two perspectives measure different things.
- The standard and concise variants each beat control by similar margins, but neither pairwise difference alone reached significance.
How they got it
Prospective, consented, three-way randomisation inside a consumer app. The arms differed only in system prompts, which are published in the appendix. Clinical evaluators rated completed sessions blind to arm on safety, appropriateness, comprehensiveness and patient alignment; participants completed an immediate survey and were asked to complete a 14-day follow-up. Spearman correlations tested clinician-patient agreement.
Limits
- 42.3% of conversations ended early; clinician ratings cover completed sessions only, though completion was similar across arms.
- Few follow-up responses, and fewer participants had a visit in the 14-day window, so effects on actual visits are untested.
- Android users of the app who opted in, likely more health-literate than average patients.
- Funded by Alphabet; authors are employees.
Why it matters for companies
This is one of the few randomised tests of a structured agent against a plain chatbot on the same model, and the structure alone moved results by about 16 points. For any firm building customer- or staff-facing assistants, workflow design is a lever separate from model choice. The non-correlation between expert and user ratings is a practical warning: user satisfaction scores alone do not show whether the output is good, so evaluation needs both.
Summary based on the full text. Read the original for the full argument and data.Read the original on arxiv.org
- Tags
- Healthcare, Rct