JMIR Publications

ABSTRACT

Background:

Artificial intelligence (AI) is rapidly reshaping healthcare, offering tools to enhance diagnostic accuracy, streamline clinical workflows, and personalize care delivery. However, real-world AI implementation remains limited, hindered by organizational, technical, and sociocultural barriers that implementation science has only begun to address systematically.

Objective:

This scoping review maps the intersection of AI and implementation science in healthcare, examining the types of AI technologies deployed, their intended use, and the processes by which these tools are implemented into practice.

Methods:

Following PRISMA-ScR guidelines, we synthesized empirical evidence from 65 studies published between December 2011 and March 2025. Searches were performed across the databases CINAHL, PubMed, PsycINFO, Scopus, and Web of Science using the terms Artificial Intelligence, Healthcare, implementation, and empirical, combined with relevant synonyms.

Results:

AI implementation research has expanded rapidly, predominantly in high-income countries, raising important questions about global equity. The most common application areas were automation and optimization (40%), computer vision (34%), and human language technologies (20%), primarily targeting clinical care (68%) and health systems management (25%). Most systems were designed for low-action autonomy (62%), emphasizing human-in-the-loop decision-making. Intended users were physicians (43%), nurses (26%), and radiologists (25%), while patients appeared as intended users in only 11% of implementations. Across the 65 studies, 40 barriers and 55 facilitators were identified across five themes: the AI system itself, healthcare professionals, patients, organizational context, and the macro level. Organizational factors and multidisciplinary stakeholder engagement emerged as the most critical enablers of successful adoption. Key barriers included insufficient AI performance, lack of transparency and explainability, limited IT infrastructure, and inadequate workflow integration. Patient-level and governance-level barriers, including data privacy and regulatory uncertainty, remained underexplored. Only 20% of studies applied theoretical implementation frameworks, and most analyses were conducted retrospectively. Mapping via the AIGENT framework revealed a disproportionate focus on workflow alignment and outcome evaluation, with comparatively little attention to early-phase activities such as needs assessment, adaptation planning, and stakeholder approvals.

Conclusions:

The current literature predominantly focuses on implementation evaluation and workflow alignment, while patient perspectives, governance conditions, and early implementation activities are underexplored. The finding that only 20% applied theoretical implementation frameworks, mostly retrospectively, reflects a gap between theory and practice and points to a need to apply them prospectively across the full implementation process. From a practitioner perspective, AI implementation should be seen as a sociotechnical and governance process that requires technical, contextual, and system knowledge, rather than merely a technical deployment.