From Pilot Trap to Institutional Capacity: A Governance Framework for Sustainable Clinical AI Implementation in Health Systems

Abstract

Clinical artificial intelligence (AI) applications frequently fail to transition from short-term pilot projects into sustained components of routine clinical care, a phenomenon referred to in this viewpoint as the pilot trap. This persistent gap reflects not only technical or regulatory limitations but also insufficient governance capacity within health care organizations. We argue that such capacity is not fully established before deployment; rather, it develops through implementation as real-world operational tensions clarify organizational ownership, accountability boundaries, and coordination mechanisms. Drawing on an 18-month implementation of a provincial clinical AI platform in China, we develop a 6-module governance framework encompassing institutional carrier formation, infrastructure governance, regulatory and ethical governance, interdisciplinary coordination, translational scaling, and lifecycle evaluation and oversight. These modules represent functional governance conditions observed during implementation rather than a prescriptive institutional architecture to be installed prior to deployment. We further introduce the concept of functional transferability and position the framework as an upstream complement to existing international governance standards, which typically specify what governance should achieve, but often assume that the organizational capacity to implement it already exists. Advancing clinical AI beyond demonstration, therefore, depends less on model performance alone than on the ability of health systems to develop and sustain the institutional capacity required for routine clinical use.