Separating AI Hype from Clinical Reality AI Workshop | September 2026-AI-enhanced Clinical Documentation

Pdf Summary

This presentation describes the rapid development of a transplant hepatology AI note co-pilot at UCSF and contrasts it with a slower, more complex AI e-consult project. The central problem is that hepatology and liver transplant documentation is unusually long and clinically complex, requiring synthesis of many scores, risks, and comorbidities.<br /><br />The speaker outlines a spectrum of generative AI options for clinical documentation, from off-the-shelf ChatGPT to custom GPTs, retrieval-augmented generation (RAG), and eventually EHR-embedded copilots and agents. The UCSF solution used ChatGPT Enterprise with a custom GPT built from exemplar notes and knowledge files. Key design principles included style imitation of real hepatology notes, a “zero-hallucination” approach that inserts *** when data are missing, a rigid but modular structure, and intelligent extraction of pasted Epic data into the expected note format.<br /><br />A major theme is how the team moved “from 0 to 1 in days” by avoiding common implementation barriers. They did not need API integration, EHR write-back, new vendors, procurement, or engineering queues because the platform had already been approved at the enterprise level. Governance, security review, and HIPAA compliance were inherited from ChatGPT Enterprise.<br /><br />The tool has been in daily clinical use since December 2025. Reported outcomes include zero marginal deployment cost, use by 22 users, and 385 notes written, with self-reported pre-charting time reduced to 5–30 minutes. However, the speaker emphasizes that formal validation is still pending, the workflow remains copy-paste based rather than EHR-integrated, and every note still requires physician review.

Asset Subtitle

Jin Ge, MD

Keywords

transplant hepatology

AI note co-pilot

liver transplant documentation

ChatGPT Enterprise

custom GPT

retrieval-augmented generation

clinical documentation

zero-hallucination

Epic data extraction

EHR-embedded copilots