Artificial intelligence (AI) offers potential to improve diagnostic accuracy, risk assessment, and patient care decisions. Prior work has largely emphasized algorithm development in controlled settings; less is known about embedding AI into routine clinical practice. The Augmented Intelligence in Medicine and Healthcare Initiative (AIM-HI), led by Kaiser Permanente and funded by the Gordon and Betty Moore Foundation, was created to evaluate and support integration of AI tools into everyday clinical workflows. This report summarizes early insights from AIM-HI’s five funded projects intended to inform real-world AI implementation.
AIM-HI distributed funding to five projects through a national, multistage review process. Selection used a structured scoring rubric to evaluate proposals. The initiative prioritized projects that moved beyond model development to address implementation in clinical environments. The authors synthesized cross-project findings to identify implementation processes, recurring challenges, and lessons learned during early deployment efforts.
The five AIM-HI–funded projects targeted distinct clinical use cases across diverse health care settings. Collectively they spanned acute, chronic, and screening contexts. The clinical focuses reported were:
The portfolio was intentionally varied to test AI deployment across different workflows, data types, and care settings, reflecting AIM-HI’s emphasis on real-world application rather than confined experimental settings.
Across the five projects, several common barriers to real-world AI deployment emerged.
Electronic health record (EHR) integration: Integrating AI tools with local EHR systems was repeatedly cited as a significant obstacle. Practical integration issues included data access, interoperability, and embedding model outputs into clinicians’ workflows in a usable way.
Data complexity: The projects encountered challenges related to heterogeneous, incomplete, or noisy clinical data. Data complexity complicated model inputs, validation, and generalizability when moving from development datasets to live clinical environments.
Regulatory requirements: Navigating applicable regulatory frameworks and organizational governance for clinical decision support and AI tools was a shared challenge. Regulatory and compliance expectations influenced deployment timelines and implementation design.
Variation in clinical workflows: Differences in local practice patterns and workflow structures required adaptations of AI tools and their delivery. A one-size-fits-all deployment approach was not effective across diverse clinical settings.
Although the projects addressed different clinical problems, the authors identified recurring themes associated with successful implementation.
Stakeholder engagement: Early and sustained involvement of clinicians, operational leaders, IT teams, and other stakeholders was essential for aligning goals, designing workflows, and encouraging uptake.
Local adaptation: Implementations that allowed for tailoring to local workflows and contextual needs were more feasible. Flexibility in how AI outputs were presented and acted on helped teams integrate tools into care processes.
Quality assurance and performance monitoring: Ongoing evaluation—both before and after deployment—was necessary to ensure safety and to detect performance drift. The authors emphasize continued monitoring of model performance and clinical impact as a core component of real-world AI use.
Partnerships and governance: Strong partnerships between implementation teams, clinical staff, and institutional governance structures supported practical problem solving, compliance with regulatory requirements, and alignment with organizational priorities.
The AIM-HI experience indicates that deploying AI in routine care is feasible across varied clinical environments, but feasibility alone is not sufficient for sustained adoption. Success depends on thoughtful systems integration, stakeholder alignment, and infrastructure for evaluation and oversight. Projects benefitted from structured selection and support but faced common operational and technical barriers that slowed or complicated deployment.
These early projects illustrate that the transition from model development to routine use requires investment in implementation activities often outside the traditional scope of data science, including EHR engineering, workflow redesign, clinician engagement, and regulatory navigation. The need for sustained performance monitoring highlights that AI in health care is not a one-time implementation but a continuous process of evaluation and adaptation.
AIM-HI’s early insights underscore key priorities for organizations seeking to adopt AI at scale: integrate tools thoughtfully into local clinical environments; engage stakeholders across clinical, technical, and operational domains; permit local adaptation of tools and workflows; and establish robust quality assurance and ongoing performance monitoring.
The authors conclude that initiatives like AIM-HI are essential to build the evidence base for scalable, real-world AI implementations. Continued work to address EHR integration, data complexity, regulatory constraints, and workflow variability will be needed to translate AI promise into routine clinical benefit.