AI Drafts, Doctor Edits: The Hidden Time Cost

Artificial intelligence now drafts replies to patient messages inside electronic health records. The goal was to cut down doctor workload. But a recent study found that physicians still spend extra time reviewing and changing these drafts. The research looked at over fourteen thousand responses from more than a thousand clinicians at one major medical center between April twenty twenty-four and August twenty twenty-five. Each reply was generated by AI, then touched by a human doctor before sending.

Researchers built a fifteen-part system to categorize every edit doctors made. The most common changes involved scheduling appointments and giving lifestyle tips. Many edits were moderate, shifting the message's meaning without altering the actual treatment plan. Editors frequently rephrased instructions or added details the AI left out, such as specific appointment times or diet advice. The taxonomy was created using a large language model, then refined through repeated review by clinical experts.

Every single edit type ended up slowing down how fast doctors replied. The biggest delays came from interpreting radiology images, which added about seventy percent more time to the response. Clarifying or ruling out a diagnosis added roughly sixty-four percent, and reading lab results added sixty percent. Response times got longer depending on how often edits happened and how complex the messages were. Even small changes accumulated to create noticeable delays across the board.

The findings show doctors are still very much in the loop, even with powerful AI tools. Experts say improving the AI prompts could reduce the most time-consuming edits. Targeting the categories that currently drag down response rates might help the technology live up to its original promise of saving time. Without these adjustments, the human effort required to fix AI outputs may outweigh the expected workload benefits.