AI course-of-care notes help doctors, pose little risk of harm to patients

Agentic large-language models can draft hospital discharge summaries that are safe, useful and demonstrably effective at helping to curb physician burnout, according to research conducted at Stanford University.

The burnout-beating benefit is evident even when the time savings are not remarkable, report François Grolleau, MD, MPH, PhD, and colleagues in a study published by JAMA Network Open May 8. 

For the study, Grolleau and team analyzed course-of-treatment summaries generated by AI agents upon patient discharge over a 10-week period in 2025. 

The work was part of a prospective quality-improvement project at a single Stanford Health Care inpatient unit.

The researchers encouraged physicians to volunteer safety feedback for every AI summary they incorporated into their final notes. 

The request prompted such reviews for 100 of 219 physician-accepted AI summaries as well as 12 of 165 rejected or otherwise unused AI summaries. 

Tabulating the results, Grolleau and co-researchers found 98% of the reviewers (98 physicians) rated the likelihood of any harm occurring from the AI drafts as “unlikely” or “extremely unlikely.” 

Meanwhile 88 unedited summaries (88.0%) were rated as having “no harm potential” at all. 

One was rated as likely to cause moderate harm. 

Suitable for integration into wider clinical care 

As for technology penetration, physicians used AI content more than half the time they could (219 of 384 discharges, 57% of cases). 

Detailed feedback on the 100 reviewed AI summaries mentioned omissions in 25 cases and inaccuracies in 20. 

Hallucinations were rare, getting called out in only two cases (2%). 

No less telling was a significant drop in physician burnout scores after the AI pilot compared with before. 

Baseline data had been collected over a 10-week period earlier in the study year, and the researchers found mean physician burnout scores fell from 1.75 at baseline to 1.20 after intervention (scale: 0 to 4.0). 

In their discussion, Grolleau and co-authors state their findings “suggest that agentic LLM workflows with reasoning models may be integrated into active clinical care with minimal safety risk and may be associated with reduced physician burnout, even when objective time savings are modest.”

Goes without saying: Physicians still must review and refine 

Commenting on the latter finding, the researchers highlight as notable their finding of a discrepancy between subjective and objective efficiency measures. 

While only 71.4% of physicians showed reductions in median documentation times—“and these ranged modestly up to 2.9 minutes”—67.0% of feedback responses reported perceived time savings, with nearly one-third (32.0%) estimating savings exceeding 15 minutes per summary.

“This pattern, where perceived value and well-being improvements outpace measured time reductions, is consistent with evaluations of other generative AI documentation tools,” Grolleau and colleagues point out.

Citing prior research, they note that earlier studies of ambient AI scribes and LLM-generated draft replies have “similarly reported wellbeing improvements without proportionate time savings. These convergent findings suggest that the primary benefit of generative AI tools lies in cognitive offloading rather than clock-time efficiency.”

More: 

‘The AI serves as a scaffolding tool, providing a structured starting point that physicians review and refine rather than generate de novo. This shifts the value proposition from efficiency to sustainability, explaining why burnout improved when clock time did not.’

The study is available in full for free.