When is AI “just another innovation”? A comparative conceptual analysis of artificial intelligence and evidence-based practice implementation
TL;DR
This analysis suggests that, across CFIR domains, adaptive and generative AI systems foreground issues such as opacity, probabilistic outputs, dependence on local data ecosystems, performance drift, vendor-mediated updating, regulatory uncertainty, professional identity tensions and the need for calibrated trust.
Abstract
Implementation science frameworks, such as the Consolidated Framework for Implementation Research (CFIR), are increasingly used to guide the implementation of artificial intelligence (AI) in healthcare. This assumes that AI systems can be understood using frameworks developed primarily for evidence-based practices (EBPs). However, AI systems vary substantially in their technical architecture, degree of autonomy, dependence on local data, regulatory status and capacity for change after deployment. To examine how different types of AI systems align with, extend or challenge core assumptions in implementation science, using CFIR as a diagnostic lens. We conducted a comparative conceptual analysis of foundational implementation science literature, literature on complex interventions and technology-enabled care, and empirical and conceptual studies on AI implementation in healthcare. CFIR was used as an analytic lens to identify points of fit and tension across its five domains. The analysis was not designed as a systematic or scoping review, but as a theoretically informed comparison of recurring concepts and assumptions across bodies of literature. AI systems differ substantially in their implementation implications. Fixed or locked AI tools may resemble conventional digital interventions, whereas adaptive, data-dependent and generative AI systems raise more substantial challenges. Our analysis suggests that, across CFIR domains, adaptive and generative AI systems foreground issues such as opacity, probabilistic outputs, dependence on local data ecosystems, performance drift, vendor-mediated updating, regulatory uncertainty, professional identity tensions and the need for calibrated trust. These issues can often be mapped to existing frameworks, but in some cases they challenge assumptions of intervention stability, boundedness and evidentiary closure. AI should not be treated as a homogeneous implementation object. Existing implementation science frameworks remain analytically valuable, but require refinement when applied to AI systems whose behavior depends on changing data, infrastructure, governance and use contexts. Implementation of such systems is better conceptualized as lifecycle stewardship than as a bounded rollout.