Physician edits to artificial intelligence (AI)-generated messages to patients were associated with different response time burdens
1. Poursoltan and colleagues examined physician modifications to AI-generated patient messages and associated time burden.
2. Messages about clinical interpretations had the greatest per-message time burden, while frequent scheduling edits accounted for the most cumulative burden.
Evidence Rating Level: 2 (Good)
Study Rundown: Widespread adoption of electronic health records (EHRs) has increased physician workload, with a significant increase in patient message volumes. AI has the potential to alleviate this burden by drafting replies to patient messages, but physician review of these drafts still generates time burden. Poursoltan and colleagues examined the time burden associated with physician modifications to AI-drafted patient messages. Researchers developed an expert-reviewed taxonomy of 15 edit categories using a large language model and assessed their frequency, extent, and association with response time. The initial dataset contained 14,350 AI-assisted responses from 1,131 physicians. The study found that messages related to scheduling, lifestyle guidance, and emotional support had the most common edits. Meanwhile, radiology interpretation, diagnostic clarification, and laboratory interpretation had the largest per-message associations, whereas scheduling, lifestyle advice, and referral coordination contributed to the greatest cumulative burden. This study identified potential priorities for improving AI drafting systems to reduce physician workload.
Click here to read the study in NEJM AI
Relevant Reading: Utility of Artificial Intelligence–Generative Draft Replies to Patient Messages
In-Depth [retrospective cohort]: Patient messages from UC San Diego Health between April 2024 and August 2025 were analyzed. AI message drafts and the final messages were analyzed, and response time was measured from reply initiation to sending, serving as a workload proxy. Outliers exceeding the 95th percentile of response time were removed to exclude system timeouts and interrupted sessions. A total of 13,632 responses from 1,110 physicians were included in the study. The most frequent edits involved scheduling or rescheduling appointments (38.5%), followed by providing lifestyle or nonpharmacologic guidance (18.2%) and providing empathy or emotional support (16.1%). The least frequent category was discontinuing or tapering prescription medications (2.4%). Edits related to radiology interpretations were associated with a 70.1% longer response time (95% confidence interval [CI], 57.3-83.9%), diagnostic clarification with 63.9% (95% CI, 54.7-73.5%), and laboratory interpretation with 60.8% (95% CI, 50.4-72.0%). Most edits were moderate, meaning that they reflected differences in clinical meaning but not in actual patient care recommendations. The largest proportion of edits that changed care recommendations were related to scheduling changes (5.03%). Frequent scheduling edits produced the greatest cumulative burden. This study was limited by the single-center setting and response time incompletely reflecting cognitive and emotional effort. Nonetheless, this study highlighted priority areas for improving AI message drafting systems to minimize physician edits and workload.
Image: PD
©2026 2 Minute Medicine, Inc. All rights reserved. No works may be reproduced without expressed written consent from 2 Minute Medicine, Inc. Inquire about licensing here. No article should be construed as medical advice and is not intended as such by the authors or by 2 Minute Medicine, Inc.
