Health System AI Governance Case Studies: How MSK and Duke Structure Oversight

Both built a named, standing committee rather than an ad hoc review process, both use a model registry that tracks every AI system through defined stages, and both put a number on their own oversight activity rather than describing it only in qualitative terms. That combination, a named committee, a working registry, and a published or disclosed track record, is what separates an operating governance program from a policy document that was written once and never run against.

The initiative evaluates clinical value, fairness, usability, regulatory compliance, and accountability before a tool goes live, and Duke is a founding member of both the Coalition for Health AI and the Trustworthy and Responsible AI Network. Duke's most concrete recent commitment is the Health AI Maturity Model Project, a joint effort with Vanderbilt University Medical Center funded by a 1.25 million dollar grant from the Gordon and Betty Moore Foundation.

The project is building a maturity model that lets a health system assess its own readiness across governance, data quality, workforce competencies, and ongoing monitoring, rather than relying on a vendor's assessment of its own product. Duke has also published policy guidance through the Duke-Margolis Institute for Health Policy, including a white paper on aligning AI innovation with accountability and trust.

That means a named governance committee with defined authority, a model registry that actually gets updated as systems move through review, and a documented pathway that differs by how a model was built or acquired. From there the engagement produces the committee charter, the registry structure and required fields, the review criteria for each deployment path, and a rollout plan for getting existing AI systems retroactively registered rather than only gating new ones.

The goal is not to copy either institution's program wholesale. It is to bring a health system's own governance structure to the point where it can be described in the same concrete, published terms MSK and Duke used, a named committee, a working registry, and a defined review path per model, before an auditor, a board, or a vendor's own sales team asks for it first.

Bring the number of AI tools already deployed or piloted across clinical, operations, and research workflows, whether any governance committee or review board already exists, and how it currently makes decisions. That is enough to scope a fixed engagement or a retainer against the rates above rather than the floor.