For years, the promise of artificial intelligence in healthcare has been tempered by a recurring operational wall: the gap between a clinician’s bedside insight and the technical ability to deploy a functional tool. While front-line providers are often the first to identify where AI could save time or improve patient outcomes, those ideas typically enter a long queue for IT development, validation, and governance.
University of Utah Health is attempting to dismantle this bottleneck by shifting its strategy from building individual AI applications to creating a “tool to build tools.” By implementing an EHR-integrated system known as the AI Workbench, the institution is empowering clinicians to develop and deploy their own AI-driven solutions without requiring one-off engineering for every new use case.
This approach marks a significant transition in how academic medical centers scale clinical AI. Rather than relying on a centralized IT team to hand-craft every application, the AI Workbench allows the people closest to the patient—the clinicians—to analyze medical record data using custom prompts and standards-based interoperability.
Overcoming the IT Bandwidth Bottleneck
In many health systems, clinical AI programs stall because the volume of ideas from front-line staff far outpaces the capacity of IT teams to build and govern them. This creates a backlog that can stifle innovation and leave valuable efficiency gains on the table.
Ken Kawamoto, MD, the Chief Health AI Transformation Officer at University of Utah Health, notes that the institution recognized this pattern early. To address it, the health system developed the AI Workbench to decentralize the creation of AI tools. By providing a common framework, the system reduces the need for bespoke engineering for every single clinical request, allowing the organization to scale its AI capabilities across the entire enterprise.
The Technical Architecture: Epic and SMART on FHIR
The effectiveness of the AI Workbench relies on its deep integration with the electronic health record (EHR). The tool sits directly inside Epic, the EHR system used by the institution, ensuring that clinicians do not have to leave their primary workflow to utilize AI capabilities.
To achieve this seamless integration and ensure data can be moved securely and consistently, the Workbench utilizes SMART on FHIR (Substitutable Medical Applications, Reusable Technologies on Prompt Healthcare Interoperability Resources). This standards-based approach allows the tool to pull specific data from the medical record and analyze it through custom prompts, bypassing the need for custom-coded interfaces for every new application.
From Rapid Prototyping to Production
The acceleration of these tools was facilitated by the launch of a health system innovation lab dedicated to rapid prototyping. This lab is designed to move AI concepts from the initial idea phase to a working prototype within one or two days.
According to Dr. Kawamoto, this shift in pace was made possible by a “step change” in the technology available. While earlier large language models (LLMs) showed promise in demonstrations, they often proved unreliable when moved into a production environment. However, the emergence of reasoning models in the fall of last year provided the reliability and safety necessary for the organization to deploy tools for use by both providers and patients.
For a detailed walkthrough of the AI Workbench and its implementation, you can view the discussion led by Dr. Kawamoto here:
Note: The AI Workbench demonstration illustrates the “tool to build tools” philosophy in a clinical setting.
Institutionalizing AI Transformation
Recognizing that scaling AI is as much an organizational challenge as a technical one, University of Utah Health created the role of Chief Health AI Transformation Officer earlier this year. Dr. Kawamoto leads this effort, overseeing the integration of AI across clinical care, research, and education.
By combining a specialized leadership role with a decentralized technical workbench, the institution is moving away from a model of “AI as a product” toward “AI as a capability.” This ensures that as new reasoning models and AI capabilities emerge, the health system can integrate them rapidly without needing to rebuild its infrastructure from scratch.
Key Takeaways of the Workbench Approach
- Clinician Empowerment: Shifts the role of the clinician from a “requestor” of IT tools to a “creator” of AI applications.
- Interoperability: Uses SMART on FHIR to ensure secure, standards-based data retrieval from the Epic EHR.
- Reduced Latency: Rapid prototyping in the innovation lab allows ideas to become working models in 24 to 48 hours.
- Reliability: Leverages newer reasoning models to overcome the production-level unreliability seen in earlier LLMs.
As healthcare systems worldwide grapple with clinician burnout and administrative burden, the move toward clinician-led AI development offers a potential roadmap for sustainable scaling. By removing the engineering bottleneck, University of Utah Health is positioning its staff to solve their own workflow challenges in real-time.
Further updates on the deployment and clinical outcomes of the AI Workbench are expected as the institution continues to expand the tool’s use across its research and education sectors.
Do you believe clinician-led AI development is the key to reducing healthcare burnout, or does it introduce too many governance risks? Share your thoughts in the comments below.