During a recent conversation hosted by the Coalition for Health AI (CHAI), Brenton Hill, the Head of Operations and General Counsel at the Coalition for Health AI (CHAI) and Brittney Harrell, Head of Information Security at Nabla explored how healthcare organizations can improve AI evaluation by asking better questions—not simply collecting more information. While model cards have become an important foundation for AI transparency, the discussion emphasized that transparency extends far beyond documentation alone.
Transparency Is More Than Documentation
A recurring theme throughout the discussion was that transparency isn't a single artifact.
While model cards help establish a common starting point by outlining intended use, limitations, validation approaches, and known risks, they don't tell the entire story. Organizations also need visibility into how vendors communicate, monitor performance, validate models over time, and support customers after implementation.
Brittney Harrell noted that true transparency is reflected not only in what vendors publish, but also in how they engage with customers, support implementations, and validate the claims they make.
The Biggest Bottleneck Isn't Review—It's Information Gathering
One of the more surprising insights shared by CHAI came from analyzing AI intake processes across multiple health systems.
Organizations reported spending anywhere from two weeks to two months evaluating individual AI solutions. The greatest source of delay wasn't reviewing documentation—it was tracking down missing information from vendors. Validation evidence, monitoring plans, workflow integration details, incident response processes, and clearly defined product boundaries were consistently identified as the information reviewers struggled to obtain.
Standardizing how this information is presented has the potential to significantly reduce evaluation time while giving governance teams greater confidence in procurement decisions.
AI Governance Doesn't End at Procurement
Another important takeaway was that AI governance is an ongoing operational discipline—not a one-time approval process.
As models evolve, organizations need confidence that vendors have structured approaches to continuous monitoring, change management, bias mitigation, and product updates. Brittney demonstrated this through highlighting the change management processes outlined on Nabla’s model card. Rather than viewing model cards as static documents, they should evolve alongside major product releases and governance practices, with standardized monitoring and updating processes.
For healthcare organizations, that also means defining expectations early. Some health systems have the resources to perform independent monitoring, while others rely more heavily on vendor-provided evidence. Neither approach is inherently right or wrong, but both require clear roles, responsibilities, and shared expectations.
Evidence Matters More Than Claims
Throughout the conversation, one message remained consistent: healthcare organizations should look beyond high-level promises.
Strong AI governance is supported by external validation, peer-reviewed research, structured monitoring, and measurable outcomes—not simply vendor assertions.
As AI adoption accelerates, organizations should look for vendors that can demonstrate how models are evaluated, monitored, and continuously improved over time. The goal isn't simply transparency for transparency's sake; it's providing healthcare organizations with the evidence they need to make informed decisions.
