JMIR Publications

ABSTRACT

Background:

Psychiatric documentation is lengthy and narrative, and Swiss mental health services report rising administrative burden alongside growing demand for care. Artificial intelligence (AI) ambient scribes—systems that transcribe consultations and draft clinical notes using large language models—have reduced documentation time and burnout in randomized trials in primary and somatic specialty care, but psychiatric evidence is limited to simulation studies and conceptual analyses. Psychiatry differs in the length of assessments, in a mental status examination that is observed rather than narrated, and in the central role of the therapeutic relationship. No controlled study has evaluated an ambient scribe in routine psychiatric care.

Objective:

This study evaluates whether an AI ambient scribe (AlpineAI Listener Tool) in routine psychiatric care (1) reduces clinician documentation time, (2) changes the quality of initial psychiatric assessment reports, and (3) affects patient-perceived therapeutic alliance and empathy, relative to usual documentation practice.

Methods:

Nonrandomized, open-label controlled trial with a difference-in-differences (DiD) design at the level of mental health care providers at Psychiatrie St. Gallen, a public psychiatric institution with 9 sites in Eastern Switzerland. The intervention group (n=24) uses the ambient scribe and an integrated general-purpose chatbot from August 17, 2026; the control group (n=20) continues usual documentation; outcomes are assessed in both groups before and during the intervention. Primary outcomes are documentation time (electronic health record audit logs), quality of initial assessment reports (psychiatry-adapted Physician Documentation Quality Instrument [PDQI-9], 8 blinded specialist raters), and patient-rated working alliance (Working Alliance Inventory–Short Revised [WAI-SR], collected anonymously). Secondary outcomes are note creation-to-completion interval, patient-rated empathy (Consultation and Relational Empathy [CARE] measure), quality of raw AI-generated reports (8-item psychiatric adaptation of the Provider Documentation Summarization Quality Instrument [PDSQI-8]), large language model–scored coverage of guideline-relevant information against a 71-item checklist, and clinician burnout (Burnout Assessment Tool). Analyses use linear mixed-effects DiD models with covariate adjustment. The Ethics Committee of Eastern Switzerland confirmed nonapplicability of the Human Research Act (BASEC Req-2026-00885).

Results:

Of 49 clinicians enrolled, 44 (24 intervention, 20 control) participate actively across 8 sites. Baseline questionnaires from 38 providers (81%) showed the self-selected groups to be balanced on sociodemographic, professional, and burnout characteristics. The pre-intervention patient questionnaire wave (July 13 to August 14, 2026) yielded 354 questionnaires from 26 providers (median 13 per provider versus 50 planned). Under realized conditions, estimated power to detect the prespecified WAI-SR DiD of 3.5 points is 63% (minimum detectable effect 4.3 points); power for CARE exceeds 99% owing to a ceiling-compressed SD. The intervention runs until November 30, 2026; analysis is planned for December 2026 to February 2027.

Conclusions:

This study will provide the first controlled evaluation of an AI ambient scribe in routine psychiatric care with real patients, jointly covering efficiency, documentation quality, and the therapeutic relationship, and will inform implementation decisions in mental health services. Clinical Trial: ClinicalTrials.gov NCT07777224 URL: https://clinicaltrials.gov/study/NCT07777224?lat=47.4657227&lng=9.0366273&locStr=Psychiatrie%20St.Gallen,%20Standort%20Wil,%20Z%C3%BCrcherstrasse,%20Wil,%20Switzerland&distance=50&viewType=Card&rank=1&tab=study