At Children’s Hospital of Philadelphia, Epic-integrated AI continues to do what humans rarely can at the tail end of long patient visits: captures every detail, improves coding accuracy, and eases the documentation burden. Greg Lawton, MD, EHR Medical Director – CHOP Primary Care, Children’s Hospital of Philadelphia, describes a measured path to deploying generative AI inside the EHR, emphasizing workflow fit, role-specific value, and disciplined change management.
Lawton also explains how CHOP is piloting AI-assisted portal replies and auto-generated patient summaries, with early signals of improved documentation quality and clinician experience.
First Attempts
At CHOP, a small early pilot with a third-party ambient tool proved instructive: the team concluded that sustained value would require tight EHR integration. Lawton’s group subsequently focused on Epic-embedded capabilities and met with the vendor weekly to tune behavior based on pediatric workflows and pilot feedback. “It really needs to be integrated into Epic,” he said, describing lessons from the initial test and the logic for shifting to platform-native features.

He added that the cadence and transparency of Epic’s roadmap allowed CHOP to advance without feeling behind more aggressive adopters of standalone tools. Lawton also pointed to a cultural posture at CHOP—innovative but cautious—that enabled pilots to run long enough to surface role-specific value, especially for nursing. He noted that this patience helped the team validate outputs, confirm an absence of hallucinations in production-grade features, and avoid overpromising benefits to clinicians before the evidence was in.
Pilot Design, Cohorts, and Change Management
Choosing the right early users is central to the method. As Lawton explained, pilots begin with clinicians who stress-test functionality and workflow integration.

“We tend to choose people who are enthusiastic but skeptical,” he said, explaining that the goal is rigorous feedback. He added that activation typically starts at the individual level and then moves to division-level rollouts when technical controls require it (for example, enabling across cardiology or pulmonary). Lawton described a playbook for scaling: concise how-to videos to fit into busy days; brief live Q&A sessions; and clear expectations that tools are optional when benefit varies by user.
He noted that ambient will not suit every documentation style—fast typists or heavy dictation users may see less gain—yet that is acceptable if the net effect across the network is positive. Lawton emphasized that CHOP’s informatics structure matters: specialty-aligned physician champions translate features into the nuances of each clinic’s workflow, while the primary-care organization focuses on simplifying steps across 31 sites and nearly 300 clinicians.
Measured Impact: Messages, Summaries, and Ambient
CHOP’s first at-scale win came from AI-assisted replies to patient portal messages. Lawton’s team found that nurses—often the first to triage inbound messages—saved 20 to 25 seconds per message, a cumulative gap-closer in high-volume ambulatory settings. He said the organization did not rush the feature; it validated draft quality, monitored error risks, and directed adoption first to the roles that benefitted most.
“We’ve not found hallucinations,” he said, a point he credits to the combination of measured rollout and close vendor feedback loops.
He added that patient summaries, generated “overnight” for the next day’s schedule, condensed roughly 15–16 notes into a five-to-six-sentence briefing. Lawton reported that nearly half of surveyed clinicians learned something they otherwise would not have known before entering the exam room unless they invested significant time in manual chart review. He noted the fit is best for those who pre-chart; for “knock-and-go” clinicians,