Separating AI Hype from Clinical Reality AI Workshop | September 2026-What GIs Need AI to do

Pdf Summary

The presentation argues that the “approval economy” in large health systems strongly shapes which AI tools actually reach clinical use. Using UCSF’s experience with two hepatology tools—LiVersa note drafting and LiVersa e-consults—it shows that the same team and institution could deploy one tool quickly while another took years because of governance and integration requirements, not necessarily because of clinical merit.<br /><br />LiVersa was built from guideline-based, peer-reviewed, benchmarked retrieval-augmented generation and performed better than ChatGPT and LLaMA on hepatology questions. A separate e-consult system was also validated and worked well, but its path to deployment was slowed by repeated review steps. UCSF’s AI governance process requires many stakeholders, multiple presentations, and committee votes at several stages. This process was designed for traditional predictive machine learning, but generative AI is harder to evaluate because outputs are free-text, there may be no single correct answer, and human judgment varies.<br /><br />The talk’s main point is that governance creates a fixed cost that can take 12–36 months per tool, favoring incumbents and already-cleared platforms such as internal enterprise AI systems or Epic products. As a result, what gets deployed is often what is compatible with governance, not necessarily what is best clinically.<br /><br />To improve this, the speaker recommends division-level clinical AI leadership—such as a gastroenterology AI officer—who can translate clinical needs, prepare acceptable evaluations, and guide products through the regulatory and governance pathway from the start. The overall message: successful clinical AI deployment requires not only good models, but also strategic navigation of health system approval structures.

Asset Subtitle

Jin Ge, MD

Keywords

approval economy

health systems

clinical AI

governance

generative AI

UCSF

LiVersa

hepatology

e-consults

AI deployment