Blab AI
Stanford AI Tool Shows Safe, Low-Risk Discharge Summaries and Burnout Reduction
In a landmark pilot at Stanford Health Care, researchers demonstrated that an AI‑generated discharge‑summary tool can be deployed safely in a hospital setting while potentially easing clinician burnout. The study, published in JAMA Network Open on May 8 2026, evaluated the MedAgentBrief system during a 10‑week trial at Sequoia Hospital’s 24‑bed unit from August 1 to October 11 2025.
Discharge summaries are the bridge between inpatient care and outpatient follow‑up, yet they demand hours of clinician time. Dr. François Grolleau, a Stanford postdoctoral scholar, observed that the sheer volume of data in electronic health records overwhelms providers. MedAgentBrief was engineered to trim that burden by distilling patient records into concise, structured narratives.
During the pilot, 11 hospitalists received a secure email each morning that contained an AI‑generated summary for every patient on the unit. The summaries followed a template crafted by Stanford hospitalists: a one‑liner on the admission reason, a high‑level overview, and a structured section for each inpatient diagnosis. Physicians were asked to flag omissions, inaccuracies, hallucinations, and any potential harm.
Of the 100 summaries reviewed, 25 % contained omissions, 20 % were inaccurate, and hallucinations were rare at 2 %. Eighty‑eight summaries were judged to have no harm potential, 21 had mild harm potential, and one was rated as possibly causing moderate harm. That single moderate‑risk case involved a patient who had finished a nine‑day intravenous antibiotic course and was switched to prophylactic oral antibiotics; the summary omitted the completed course, but independent reviewers concluded the omission did not pose a risk.
Physicians reported a perceived time saving of more than ten minutes per summary, yet time‑log analysis of the electronic medical record showed a maximum of three minutes saved, varying by user. The discrepancy likely reflects the effort required to edit and verify the AI output versus writing from scratch.
Despite modest time savings, the pilot recorded a significant drop in provider burnout scores. Dr. April Liang, a Stanford hospitalist and co‑author of the study, noted that the summaries were especially helpful for complex patients with long stays, pointing out that the AI highlighted an anemia that was not central to the admission but important for outpatient care.
The pilot concluded on October 11, but MedAgentBrief remained in routine use for an additional six months before being deactivated in April 2026, ahead of the rollout of a similar tool by Stanford Medicine’s electronic medical record vendor.
The study was exempt from institutional review board review, and the authors plan to develop a system for evaluating AI tools from vendors. According to the authors, MedAgentBrief ran on secure, HIPAA‑compliant Stanford Health Care infrastructure, pulling data via FHIR and orchestrating inference through LangChain on Databricks.
The findings suggest that AI‑generated discharge summaries can be deployed safely with low risk of harm and may help alleviate clinician workload. Future work will focus on broader evaluation frameworks and integration with commercial electronic health record systems.
Discharge summaries are the bridge between inpatient care and outpatient follow‑up, yet they demand hours of clinician time. Dr. François Grolleau, a Stanford postdoctoral scholar, observed that the sheer volume of data in electronic health records overwhelms providers. MedAgentBrief was engineered to trim that burden by distilling patient records into concise, structured narratives.
During the pilot, 11 hospitalists received a secure email each morning that contained an AI‑generated summary for every patient on the unit. The summaries followed a template crafted by Stanford hospitalists: a one‑liner on the admission reason, a high‑level overview, and a structured section for each inpatient diagnosis. Physicians were asked to flag omissions, inaccuracies, hallucinations, and any potential harm.
Of the 100 summaries reviewed, 25 % contained omissions, 20 % were inaccurate, and hallucinations were rare at 2 %. Eighty‑eight summaries were judged to have no harm potential, 21 had mild harm potential, and one was rated as possibly causing moderate harm. That single moderate‑risk case involved a patient who had finished a nine‑day intravenous antibiotic course and was switched to prophylactic oral antibiotics; the summary omitted the completed course, but independent reviewers concluded the omission did not pose a risk.
Physicians reported a perceived time saving of more than ten minutes per summary, yet time‑log analysis of the electronic medical record showed a maximum of three minutes saved, varying by user. The discrepancy likely reflects the effort required to edit and verify the AI output versus writing from scratch.
Despite modest time savings, the pilot recorded a significant drop in provider burnout scores. Dr. April Liang, a Stanford hospitalist and co‑author of the study, noted that the summaries were especially helpful for complex patients with long stays, pointing out that the AI highlighted an anemia that was not central to the admission but important for outpatient care.
The pilot concluded on October 11, but MedAgentBrief remained in routine use for an additional six months before being deactivated in April 2026, ahead of the rollout of a similar tool by Stanford Medicine’s electronic medical record vendor.
The study was exempt from institutional review board review, and the authors plan to develop a system for evaluating AI tools from vendors. According to the authors, MedAgentBrief ran on secure, HIPAA‑compliant Stanford Health Care infrastructure, pulling data via FHIR and orchestrating inference through LangChain on Databricks.
The findings suggest that AI‑generated discharge summaries can be deployed safely with low risk of harm and may help alleviate clinician workload. Future work will focus on broader evaluation frameworks and integration with commercial electronic health record systems.
