Workflow Integration of an Ambient AI Scribe in a High-Throughput Emergency Department: A Role-Stratified Mixed-Methods Evaluation (Preprint)

<sec> <title>BACKGROUND</title> <p>Ambient artificial intelligence (AI) scribes are entering clinical practice, but their value has largely been reported as a single average effect from English-language outpatient settings staffed by a single clinician role. Whether this value is similarly observed in emergency department settings characterized by multiple clinician roles, high patient throughput, and frequent interruptions, and how it may differ across roles, requires further investigation.</p> </sec> <sec> <title>OBJECTIVE</title> <p>To evaluate the real-world implementation of an ambient AI scribe in a high-throughput emergency department and determine how its effects on documentation workflow, consultation patterns, usability, and perceived clinical value differed across clinician roles. We also aimed to assess whether system performance and user-related factors were associated with clinicians’ usability perceptions.</p> </sec> <sec> <title>METHODS</title> <p>We conducted a 12-week prospective before-and-after mixed-methods evaluation consisting of a 6-week control period and a 6-week intervention period in the emergency department of a tertiary academic hospital. A closed cohort of 18 clinicians—10 attending physicians, 3 residents, and 5 physician assistants—used the ambient AI scribe across 370 encounters. Quantitative analyses examined role-stratified changes in documentation-completion time and recorded consultation duration. We additionally evaluated speech-recognition and clinical-information extraction performance, reviewed draft errors, administered a usability survey, and conducted in-depth interviews to characterize workflow integration and perceived value.</p> </sec> <sec> <title>RESULTS</title> <p>Rather than showing a uniform documentation-time pattern, implementation was associated with role-dependent shifts in clinical work. Documentation-completion time decreased most among attending physicians, from a median of 17.5 minutes to 6.1 minutes, whereas recorded consultation duration increased most among residents, from 5.5 minutes to 6.3 minutes. Physician assistants described cognitive offloading during multitasking, representing a form of value not captured by time-based metrics. The frequency of high-risk draft errors was similar across clinician roles. Perceived usability was more closely associated with draft accuracy than with clinicians’ affinity for AI or trust in the system.</p> </sec> <sec> <title>CONCLUSIONS</title> <p>In brief, densely sequenced emergency encounters, the value of an ambient AI scribe was shaped by clinician role, workflow position, and the redistribution of documentation work rather than by a uniform reduction in documentation time. Operational friction may offset or reshape automation gains, while benefits such as cognitive offloading may remain invisible to conventional time-based outcomes. These findings indicate that average-effect summaries may be inadequate for evaluating ambient AI scribes in emergency departments and other high-throughput, multi-role clinical settings.</p> </sec>