Hospitals Deploy AI Faster Than They Test It

A resident on nights doesn't ask whether a hospital "has AI." Every hospital has AI now. The question that actually matters is quieter: did anyone test this thing before it started writing notes on my patients?

A report released this week by UPMC's Center for Connected Medicine and KLAS Research gives the first real numbers on that question, and they are not comforting.

What the report found

CCM and KLAS interviewed more than two dozen health system leaders in May and June 2026 for a study titled "Validation and Trust: How Health Systems Are Testing and Governing Analytics and AI Solutions." The topline: over 90% of health systems have deployed third-party AI tools. Clinical documentation and ambient scribing lead the use cases at 52%, ahead of revenue cycle and coding (36%) and medical imaging (32%).

But deployment is not the same as validation. While 92% of these organizations say they evaluate AI tools before rollout, less than 44% have a dedicated data platform or sandbox to actually test model accuracy, safety, and drift before the tool touches a patient chart. And 63% describe their own AI strategy as "still developing or ad hoc." Only a small fraction call it advanced.

Ken Howard of UPMC Enterprises put the risk plainly: without a consistent test environment, health systems "go down that path just to learn that all that work potentially wasn't justified." Rob Bart, UPMC's chief medical information officer, described what ongoing validation should look like instead: "We monitor that on regular intervals post-implementation to make sure that the guidance that it is intended to provide is still accurate."

That second quote is the whole point. A tool that was safe on the day IT signed off on it is not automatically safe six months later, after a vendor pushes a model update, after the patient population shifts, after the EHR version changes underneath it. Validation is not a gate you pass once. It is supposed to be a habit.

Why this is a workflow problem, not just an IT problem

Read against the rest of this year's coverage, this report is the missing middle piece. We've heard the adoption numbers (ambient scribes now standard at most large systems) and we've heard the courtroom numbers (consent lawsuits over recorded encounters, disputes over who is liable when an AI-generated note is wrong). What's been missing is the boring, load-bearing layer in between: does anyone actually check this thing works, on an ongoing basis, before and after it goes near a patient.

The KLAS/CCM numbers say that for most hospitals, the honest answer is "sort of, once, and then hope." That's a governance gap, but it lands on the doctor. When a scribe drafts a note or a decision-support tool flags a risk, the clinician who signs it is the one accountable for what's in the chart, whether or not their hospital built the sandbox to properly stress-test the tool first.

This is exactly the case for keeping AI legible on the ward, not just fast. A tool that shows its sourcing, that a doctor can actually check against the primary note or the lab value before signing, survives a governance gap that a black-box tool doesn't. The hospitals in this report aren't lacking ambition. They're lacking the unglamorous infrastructure, sandbox environments, drift monitoring, defined ownership, that turns "we deployed AI" into "we can prove it's still safe." Until more systems build that muscle, the burden of checking the AI's work falls on whoever is standing at the bedside when it's wrong.

Sources

• UPMC, "Research: AI Adoption by Health Systems Is Outpacing Strategy" (Aug. 6, 2026): https://www.upmc.com/media/news/080626-ai-use-in-healthcare-systems-upmc-research
• MedCity News, "Hospitals Are All In on AI, but Testing and Oversight Haven't Caught Up" (Aug. 30, 2026): https://medcitynews.com/2026/08/upmc-ai-testing-hospital-tech/
• Becker's Hospital Review, "Most health systems lack dedicated AI testing platforms: Report": https://www.beckershospitalreview.com/healthcare-information-technology/ai/most-health-systems-lack-dedicated-ai-testing-platforms-report/