Most Hospitals Lack Dedicated AI Testing Environments, Study Finds

Testing infrastructure lags behind AI adoption

Hospitals across the United States are deploying third-party AI tools at an accelerating pace, but the infrastructure needed to properly evaluate these systems before they reach patients is seriously lagging, according to new research from UPMC's Center for Connected Medicine and KLAS Research.

The study, which surveyed health system leaders, found that fewer than half of hospitals maintain a dedicated environment for testing AI tools prior to clinical deployment. While most organizations conduct some form of evaluation, the methods vary dramatically — ranging from formal vendor testing protocols to informal pilot programs with little structure.

Perhaps more concerning: 63% of health systems characterize their AI strategy as either still developing or ad hoc, suggesting that governance frameworks have not kept pace with the technology's rapid adoption.

Why it matters

Without proper testing infrastructure, hospitals risk committing six months or more to implementing AI solutions that ultimately fail to deliver expected value. This gap exposes patients to potential safety risks from unvalidated algorithms and wastes scarce resources that could be directed toward more effective interventions. As AI adoption accelerates — one American Medical Association survey showed clinician use nearly doubled from 2023 to 2026 — the absence of standardized evaluation methods creates mounting risk across the healthcare system.

Resource constraints drive shortcuts

The testing gap stems primarily from time, capital, and talent limitations, according to Ken Howard, vice president of technology services engineering at UPMC Enterprises. When hospitals identify a problem they believe AI can address, they often lack the resources to build proper testing infrastructure first.

"Without having a dedicated or consistent test environment strategy, they're going to go down that path just to learn that all that work potentially wasn't justified," Howard explained.

This forces organizations to default to standard IT implementation processes rather than AI-specific validation protocols.

UPMC's approach to validation

UPMC has developed Ahavi, a real-world data platform that allows the health system to validate third-party AI tools against de-identified patient data before formal deployment. But pre-deployment testing represents only part of the solution, according to Rob Bart, UPMC's chief medical information officer.

The health system has maintained a formal AI governance structure for more than two years that includes continuous monitoring of tools after they go live. This ongoing scrutiny is particularly critical for clinical algorithms that predict outcomes like hospital length of stay or readmission risk.

"We monitor that on regular intervals post-implementation to make sure that the guidance that it is intended to provide is still accurate and reflective of the original [validation]," Bart said.

UPMC also tests vendor algorithms against its own patient population rather than relying solely on vendor-provided testing data. This approach helps identify issues including algorithmic bias and model drift that more generic datasets might miss.

The equity dimension

Kate Eisenberg, senior medical director of DynaMed, an AI-powered clinical decision support tool, emphasized that evaluation protocols must specifically address equity concerns. She noted that clinical teams should receive training to assess AI responses for potential bias.

"We've always had in our web interface the opportunity for users to flag if there was an equity concern or not," Eisenberg said.

Until industry-wide standards emerge, health systems and vendors remain largely responsible for governing AI implementations on their own.

These findings were first reported by MedCity News.