Agentic AI discharge summaries linked to safety, clinician wellbeing
Summary
A 10-week Stanford pilot evaluated MedAgentBrief, an agentic AI workflow powered by Gemini 2.5 Pro, which autogenerated 1,274 hospital course summaries covering 384 discharges and was incorporated into final discharge documentation in 57% of cases. Clinicians rated about 88% of unedited summaries as having no harm potential; omissions were the most common error (25%), inaccuracies 20%, and hallucinations 2%. Physicians reported clinically meaningful reductions in work-exhaustion burnout scores (from 1.75 to 1.20) and perceived time savings, though measured time savings averaged only 2.9 minutes per summary and EHR closure time did not change. The study authors highlight that the tool functions mainly as cognitive offloading rather than clock-time efficiency and flag high omission rates and the need for continuous model training and safety evaluation before wider deployment.