Patient portal messaging has quietly become one of the largest sources of administrative burden in modern clinical practice, consuming physician time without reimbursement and contributing to burnout. Any technology that durably reduces that load without compromising care quality deserves careful scrutiny — and this pilot study offers the first granular usage data on how clinicians actually behave when AI drafts are made available to them.
Across 80 clinicians at seven Pennsylvania clinics within a large academic health system, Epic's Augmented Response Technology (ART) — powered by a large language model — was made available to generate draft replies to incoming patient portal messages. Usage was voluntary and uneven: clinicians initiated only about one in five replies from an ART draft (mean 20.2%), yet when they did engage, 40% of those drafts were sent with minimal or no editing. Survey data from 52 respondents showed two-thirds found the drafts useful, though fewer than half agreed ART meaningfully improved reply quality. Notably, nurses reported more favorable views than physicians across nearly every satisfaction dimension, suggesting that perceived value tracks with baseline messaging workload.
This finding sits within a fast-moving landscape where health systems are racing to deploy generative AI for administrative relief, often ahead of robust evidence. The 20% adoption rate is modest — likely reflecting both the voluntary design and physician hesitancy around AI-authored clinical communications. The 40% low-edit rate is clinically significant: it implies a meaningful share of patient-facing messages were substantially AI-generated, raising unresolved questions about accuracy, tone, and liability. The nurse-physician satisfaction gap deserves further investigation — it may reflect differences in message complexity, autonomy, or digital comfort. This is an incremental but practically important study; replication in larger, mandatory-use deployments will be needed before generalizable conclusions about safety and quality can be drawn.