What Healthcare’s Stakes Can Teach Every Industry About Safely Deploying AI

The hardest part of putting AI to work is deciding what the organization refuses to hand over. In healthcare, where an agent’s mistake can reach a patient, that line gets drawn earlier and more deliberately than in most industries, which is exactly why the rest of the enterprise world has something to learn from how health systems evaluate, deploy, and support these tools. The lesson starts with a distinction between what you let AI execute and what you never let it decide.

Advising leaders on these decisions is Roderick Harris. He's the Founder and CEO of RDH Strategic Advisors, LLC, which helps healthcare organizations govern AI, drive adoption, and tie technology decisions to measurable enterprise value. A strategic transformation executive with more than two decades inside complex provider environments, most recently in senior leadership at a major U.S. cancer center, Harris has led EMR modernization, portfolio governance, and the vetting of hundreds of technology requests. His framing of where AI belongs comes from having had to make those calls himself.

"I always remember that systems will not outsource judgment. Only execution shifts, from cost logic to control logic," he said. Holding that line, he's found, depends less on the technology itself than on the operating model an organization builds around it.

• An operating model problem, not an AI problem: In Harris's experience, the friction organizations blame on AI usually sits somewhere else. He traces it to operating models that were never redesigned for the way agentic tools actually work, with software running tasks in the background that staff can't see and haven't adjusted their own habits around. "This is an operating model problem. That’s where most of the friction derives," Harris said. He draws a parallel to the shift from paper to electronic medical records, when systems had to actively push clinicians toward the new tools. "Going back to the early days of the digital frontier, the question was, how do you change the way you work so that the system is a force multiplier?"

• Celebrate, sustain, reinvest, or lose the value: Harris sees most value-realization frameworks stop at the moment of go-live. "We celebrate. We don’t sustain it, we don’t reinvest it, and then we don’t use that data to make new decisions," he explained. In his view, this is where the real work should begin. Value has to be sustained across the full life cycle of the product, then captured and reinvested back into the organization to inform the next decision. Without that, organizations repeat the cycle blindly. What counts as success differs by organization and by use case, which means defining what good looks like before buying, then holding the business owner and finance jointly accountable for proving it.

At a time when executives are pitched AI for everything, Harris draws a sharp line between a tool that demos well and one that changes the economics of a workflow. He points to ambient documentation tools that can generate a clinical note in seconds, a genuine technical feat, that clinicians then re-read line by line because liability rides on accuracy. The two minutes spent editing can erase the promised savings and even add time. "Everybody is looking at these new shiny objects, and we’ve got to answer the question: how does it strategically align with our organizational priorities?" The questions that matter, he said, are whether a tool reduces cost and gives capacity back. He described fielding more than 700 project requests, roughly 500 of them distinct AI use cases, and warns that organizations chasing every shiny object will overwhelm and burn out their IT teams.

The distinctive challenge of AI, in Harris’s telling, isn't just what an organization chooses to implement, but what's already arrived unannounced. No pilot is truly free, since every one carries capacity, resource, and time costs that convert to money. On top of that, vendors are switching on AI features in software that's already deployed, and contracts now arrive with AI embedded, down to biomedical devices that ship with it inside. Catching all of it is difficult enough; proving a return on it is harder still. That sprawl, he asserts, is why organizations need a system built for AI rather than the technology-driven governance processes they already know. "AI is totally different, and it’s going to require a different muscle," Harris said.

• The agent that started making its own phone calls: To demonstrate what happens when oversight lags, Harris offers a concrete case involving an ambient scribe that also handled patient scheduling. "It started calling insurance companies and trying to get the insurance companies to verify eligibility for patients and payment without our knowledge," he recalled. The team only discovered it when the insurers called back, confused, reporting what sounded like robocalls from an organization that doesn't make robocalls. He asked the obvious next question: what if those calls had gone to a pharmacy and started ordering medications? "Those situations we typically didn’t think about a lot, because it was controlled in a bubble. The bubble is now burst, and those situations are everywhere," Harris said. It's why he treats data posture, loss prevention, and early cyber involvement as non-negotiable.

• Pilots are the only safe front door: The containment answer, Harris argues, is that proofs of concept and pilots should be the sole route by which AI enters an organization. A pilot controls the environment, lets users see a tool before it touches live data, and surfaces problems while they're still cheap. "If you start off that way, it also helps to improve adoption because everything that was learned in a pilot now moves right into production and implementation," he noted. Running that playbook well, though, depends on vendors willing to behave as partners, moving at the speed the organization needs, sharing in risk transfer and cost, and standing behind what they sell rather than pushing the next product. He describes holding that accountability through vendor scorecards, back-out clauses written before signing, and keeping vendors at the table through the life of the implementation.

As for what the enterprise orchestration layer has to make visible, Harris pushes the center of gravity away from IT. Security and support teams are the front door and can govern through policy and monitoring, but the larger lever is a well-informed end user who understands both the capabilities and the risks, especially among clinical operators, and that literacy has to be kept continuously in front of them. "Change management is something that lacks a lot," he said. "Everybody wants to speed to value, they want to rush and turn things on." His preferred correction is to stop letting functions operate in silos, since AI now requires intake, vetting, piloting, implementation, and continuous learning to run hand in hand rather than as a black box that emits a solution. "All groups are going to have to come together and really understand their focus and their lane so they can play their role well."

Harris' underlying point is that regulated industries like healthcare feel these constraints first and therefore work out the discipline first. Enterprise SaaS, manufacturing, and other less-regulated sectors have room to learn from that pressure, because the limits a health system operates under are what force the clarity. Whatever an agent is allowed to reach, retrieve, or trigger, the decision about what it's allowed to decide stays with people. Systems take over the execution. They don't take over the judgment.