JMIR Publications

ABSTRACT

Background:

Documentation burden contributes to physician burnout and electronic health record (EHR) dissatisfaction. Ambient artificial intelligence (AI) scribes were developed for general ambulatory care and are used across specialties. Psychiatry documentation differs. The clinical data include how the patient speaks and appears, not only what is said. Consent depends on capacity that varies with illness, substance use information is sensitive, and the record informs involuntary care. A 2026 Ontario Auditor General review found 17 of 20 AI scribes missed key mental health details in at least one test. Implementation experience in psychiatry is scarce.

Objective:

This report describes the implementation, governance preparation, and early adoption of an enterprise-integrated ambient AI scribe in outpatient psychiatry, and identifies psychiatry-specific risks.

Methods:

The scribe was deployed within the institutional EHR at an academic mental health hospital in Toronto, Canada, using a People, Process, and Technology framework. Fourteen months of governance preparation preceded launch at this single site: a Privacy Impact Assessment, a Threat Risk Assessment, standardized consent developed in consultation with patient and family advisors and a lived experience focus group, and prelaunch testing across at least 57 component and script tests, including eight simulated psychiatric scripts. Rollout began January 27, 2026 with 70 physicians in seven subspecialties. Adoption was measured from backend EHR analytics to May 31, 2026. Outcomes will be assessed against a five-year EHR-burden baseline. Reporting follows iCHECK-DH.

Results:

Of 69 physicians analyzed, 45 (65%) used the scribe and 36 (52%) reached meaningful adoption (≥10% of eligible encounters). Physicians reported better eye contact and less screen time. Structured psychiatric assessments did not populate; pronouns were inconsistent with gender identity; uncommon names were misspelled or omitted; medication content was incomplete; and draft language occasionally framed diagnoses prematurely, raising concern about diagnostic anchoring.

Conclusions:

Prelaunch testing identified psychiatry-specific and identity-related documentation failures before any live patient encounter. The scribe ran on a primary care model, the only integrated one then available in Canada. Mental status, risk, forensic, substance use, medication, and identity-sensitive content warrant testing in psychiatric deployments. Ambient documentation is a low-risk setting in which to build trust, and the governance, consent, and engagement established before launch are intended to carry forward to higher-risk agents such as order entry and chart review. Outcome data are pending.