Source: JAMA Network Open, published online October 30, 2026, observational EHR implementation study, original paper · DOI 10.1001/jamanetworkopen.2026.41364. Evidence: large real-world observational comparison; association, not randomized efficacy. Clinical state: operations measurement, not autonomous communication.
It sounds obvious that if a language model can prepare a good portal reply, the clinician should spend less time responding. The actual experience of a large health system was more complicated. Investigators analyzed 983,154 patient-portal message responses, involving 1,542 clinicians and 272,091 patients, to examine an EHR-integrated AI draft tool. Adoption was associated with an 8.6% increase in turnaround time. For messages on which clinicians actually used the tool, turnaround time was associated with a 29.5% reduction. The headline gains did not translate to an overall reduction in after-hours responding.
Those results are not contradictory. Clinicians may select AI for some messages, may use extra time reviewing others, and may respond to more messages because the work becomes easier to initiate. One cannot infer the causal effect of using the tool from comparing selected messages with other messages in an observational deployment.
What this changes: A productivity demonstration needs to track more than draft-generation speed. It needs elapsed patient wait, time spent editing, message count, response quality, missed clinical issues, after-hours work, and whether the task has moved to another staff member.
Keep reading on this subject
AI in medicine · Care & independence