Health systems stood up AI governance for launch. Now comes the art of maintenance.

Ask a health system about AI governance and you’ll hear about the front end: the intake committee, the evaluation framework, the validation study that runs before anything touches a clinician or a patient. That machinery now exists almost everywhere, and it works. Ask what happens to the same tool eighteen months after go-live — who checks whether the model has drifted, whether it still earns its keep, what it is quietly doing to the people who use it — and the answers get thin.

That gap, between governing AI to launch and governing it to maintain, was the theme that kept resurfacing when roughly 50 attendees — health system executives, faculty, and sponsors — gathered in July at Cedars-Sinai’s space in the Pacific Design Center for CHIME’s 2026 Innovation Summit in Los Angeles. Nobody presented it; it surfaced from the room, panel after panel, until the closing discussion finally gave it a name.

Governance to launch is solved; governance to maintain is not

Intake, one panelist noted, is a key milestone whether the technology is brand new or an AI feature added to something the organization already owns — and the industry handles that milestone well. The scribes are even getting decent ongoing quality checks now. But ambient scribes are one tool among a fast-multiplying many, and across the panels, leaders kept naming the specific places where post-launch governance is missing and the bill is already arriving:

• Early deployments came with permanent obligations nobody priced. Radiology triage tools were among the first AI implementations in one leader’s health system: algorithms that flag urgent findings and reorder the worklist so those cases reach the radiologist’s eyes first. The tools work, but each one arrived trailing commitments that never end. Each tool required a training process for physicians on the notification workflow, plus governance and compliance structures that hadn’t existed before. And every missed notification carries potential legal consequences. All of that costs money, one leader pointed out, and none of it appeared in the business case that got the tool approved.
• AI is entering through products no one thinks to govern. The intake committee sees the clinical AI tool that arrives through the front door. It does not see the AI features switching on inside the ERP, the HR system, and dozens of vendor products already under contract — many of which carry no FDA approval and no post-market surveillance obligation. Jennifer McCaney, chief innovation officer at UCLA Health, argued this is about to reshape what vendor partnership means. Nearly every product a health system buys will soon have a software component built on an AI model, updating on the vendor’s schedule rather than through the usual checks and balances — and no health system can monitor all of it. Surveillance, she predicted, becomes a component of collaboration itself: systems will have to choose partners partly on the strength of who is watching on their behalf.
• The maintenance bill lands on IT, and nothing funds it. McCaney put a number on the trajectory: an IT workforce that stood under 300 before UCLA Health’s 2013 Epic go-live now exceeds 1,000, a budget line that simply did not exist a decade ago — while reimbursement has not risen to meet it. Each new AI implementation adds its own increment of maintenance, training, and surveillance cost on top. Without a “meaningful use 2.0,” she asked — invoking the federal incentive dollars that helped carry the EHR transition — how do health systems sustain the burden?

The summit’s most clarifying moment took about ten seconds. In a Thursday fishbowl, a moderator polled the room — picking up on a point Eftekhari had just made about sanctioned tools losing to personal favorites. How many of you have an enterprise-approved AI? Most hands went up. How many prefer a different AI in your personal life? Almost everybody. How many would be tempted to use the personal favorite at work if you could get away with it? Almost everybody. Governance on paper, the room had just demonstrated, is not governance in practice.

Nasim Eftekhari, chief AI and analytics officer at City of Hope, described where that realization leads. Her team ran five or six AI 101 sessions across the organization and stood up governance and education together, then accepted that neither holds on its own. Her physicians love OpenEvidence because it is good and free; telling people to stop only works when the sanctioned alternative is “as good or better.” City of Hope’s target is an enterprise platform strong enough to pull 80 to 90 percent of one-off AI use back inside governance — treating shadow AI as a product problem to out-build rather than a policy problem to enforce.

The maintain-side question extends to minds, not just models. In the Friday takeaways discussion, one panelist predicted the industry’s real regret won’t be AI that made mistakes but outsourcing too much thinking — pointing to the much-discussed idea of “cognitive surrender” — and observed that as AI absorbs successive layers of expertise, nobody yet has an answer for how the next generation of the workforce gets trained.

The next profession: Minding the machines already at work

The summit’s most practical forecast arrived half-joking. Reviewing everything the Jetsons predicted correctly, one panelist pointed out that George Jetson’s job was quality assurance on computers — perhaps the first human in the loop — and predicted an entire profession will form around validating AI that is already live, because organizations will rely on its output whether they like it or not. It was a laugh line that doubled as a to-do list. The room spent a day and a half agreeing that AI governance doesn’t end at go-live — and that the budgets, vendor contracts, and job descriptions for everything after still have to be written.