From AI pilots to practical scale: How healthcare leaders are balancing productivity, governance and trust

BECKER’S HEALTHCARE // HP

Healthcare leaders no longer need to be convinced that AI matters. The technology is already reading radiology studies, drafting clinical notes, working denials in the revenue cycle and reaching patients at home.

“AI has moved from being a future opportunity to implementation, period,” said the vice president and chief medical officer of an academic safety-net health system in the Northeast.

The question is no longer whether the technology matters, but which problems it should solve, how to prove value in real workflows and how to scale without creating new clinical, operational, financial or data risk.

To understand how leading organizations are navigating that shift, Becker’s Healthcare and HP brought together CIOs, CTOs, chief medical officers and clinical informatics leaders from nine health systems across the country for a peer-to-peer conversation. For HP, the conversation reflects a broader priority: helping healthcare organizations modernize the technology their teams rely on and create the secure, manageable environment responsible AI adoption requires.

Despite differences in size, geography and resources, the organizations surfaced many of the same priorities for putting AI into practice. Across their experiences, five priorities emerged: use-case discipline, governance, adoption, responsible scale and human-centered care.

Together, they show how healthcare organizations can improve workforce productivity, protect data and preserve the human relationships at the center of care.

1

BECKER’S HEALTHCARE // HP 2

He also noted that when clinicians bring in tools they have heard about elsewhere, each one is vetted against a concrete use case and value-add before moving forward. Both the ambient AI and coronary CT tool deployments cleared that bar because each answered a defined clinical need, not because the technology happened to be available.

The CTO of a Catholic health system in the Midwest warned about what happens when that discipline is missing. Leaders often ask IT to automate a broken workflow, expecting technology to fix it. “What you wind up with is a very bad solution that is now automated and will fail the same way every time,” he said. AI amplifies whatever process it is layered onto, which is why fixing the workflow must come first.

Educating the organization about what AI is and what it is not has become a core leadership responsibility. Setting realistic expectations about which problems AI can solve, and which it cannot, is what keeps the problem rather than the product at the center of the decision.

Governance determines scale Across the discussion, leaders treated formal AI governance as a prerequisite for scale rather than an obstacle to it. One system splits oversight between clinical and operational boards under a single governance umbrella; another built a comprehensive AI policy jointly across its university, hospital and physician practice plan.

Across organizations, the goal was the same: create a safe, repeatable path for AI to enter the organization.

The key, leaders agreed, is right-sizing the process. A cybersecurity leader at a cancer research and care organization in the Northeast described the volume of AI requests from researchers as a forcing function for innovation within governance itself. “We can’t be asking 400 questions

Start with the problem, not the technology Leaders that approach AI deliberately share one discipline: they are not chasing tools. They start with a real problem in a clinical, operational or financial workflow, define what good looks like and only then decide whether AI is the right fit.

That discipline came up more than any other theme in the conversation, and the pressure behind it is easy to see. New AI products arrive weekly, and clinicians are bringing tools to IT “hot and heavy,” as one leader put it. In that environment, use-case discipline has become the foundation on which everything else — governance, adoption and scale — depends.

“You can’t use AI as a solution looking for a problem. You really have to be dialed in on what the challenges are in your organization,” the CIO of a safety-net health system in the Southwest said. Her organization serves a vulnerable patient population under mounting Medicaid pressure, and it has taken a platform-first, fast-follower approach. Her teams adopt AI capabilities only when they reduce cognitive load on physicians or automate low-complexity, rules-based work. The capabilities arrive through the core EHR, ERP and collaboration partners already in place. Those criteria are the discipline she described in practice: identify the problem, then find the capability that fits it.

The IT director of a community medical center in Texas described a similar discipline. His organization runs ambient AI in patient encounter rooms and recently went live with AI that reads calcium scores in coronary CT studies, but deliberately resists the hype cycle.

“We have a governance model in place, so we are very conservative in our AI approach,” he said. “We are intentional about what we do with AI. We are not trying to drink from the AI fire hose. I tend to take a slow approach to it.”

like we would on a regular risk assessment. We have to come in with something a little lighter, to at least get folks set up in a safe way, and then put the overlay of governance on top,” he said.

That risk-tiered mindset, with light-touch review for low-risk tools and deeper scrutiny for high-risk clinical applications, surfaced repeatedly. So did the recognition that governance does not end at go-live.

“It’s really taking it from ideas and pilots and concepts to something that is repeatable and reliable,” the safety-net CIO said. A clinical informatics program director at a multi-hospital health system in the Mountain West emphasized that organizations budget heavily to implement AI, but few plan for the ongoing “care and feeding”: monitoring model performance, identifying model drift, checking for bias, adjusting workflows and maintaining ethical standards over time.

Adoption is a human challenge If use-case discipline is the paper’s first lesson, this is its most human one. Turning on an AI tool is the easy part. Whether it scales depends on whether clinicians and staff trust it, understand its value and can use it inside real workflows. This makes adoption a change-management challenge, not a technology deployment.

“A lot of the problem with getting AI scaled is not turning on the technology. It’s getting humans to change their behavior and that is not easy,” said an innovation leader at a large distributed health system in the South. “Most of the things that I’ve been able to scale this year have been inside of the EHR, primarily because it’s just easier to integrate it into the workflow.”

The Mountain West informatics leader, a nurse of 18 years, underscored the importance of making AI decisions with humans in mind.

“A tool is just a tool. Without all the people and the process, it’s going to be a very expensive tool that doesn’t make much of a difference,” she said. Her team does not deploy AI tools straight out of the box. Instead, clinicians who are experts in conditions like sepsis review model outputs, data scientists tune thresholds and rule-based logic so nurses are not over- triggered and workflows are built around what frontline teams can realistically act on.

3BECKER’S HEALTHCARE // HP

BECKER’S HEALTHCARE // HP 4

The chief medical officer of the Northeast academic safety-net system added that his system saw utilization of an early ambient documentation tool decline steadily because integration friction made it impractical in daily use.

“Having the frontline staff involved from the beginning and putting staff satisfaction in as a metric, is so important in the successful implementation of any AI product,” he said.

This is where the technology experience becomes part of the adoption strategy. AI tools are only as useful as the environments in which clinicians and staff access them: secure devices, reliable performance, simple management and workflows that reduce friction rather than add to it.

Responsible scale requires control of risk, cost and data exposure Leaders agreed that they are ready to scale AI but not at the expense of data security, budget discipline or operational trust.

“We also have to manage the expenses. I have several friends who budgeted so much for token use and blew through their budget in half a year. That would not sit well with my CFO,” said the CTO of the Catholic health system.

Data exposure is an equally live concern. The same CTO described discovering that a development partner had placed code in a public repository, creating a potential PHI exposure that his team moved quickly to contain.

Legacy data presents another risk. The CTO noted that many long-established organizations still have substantial data- governance and cleanup work ahead.

“We’ve been around for 150 years and our data governance has not been really good. It’s kind of like a really messy closet. We had to go back and do a lot of data cleanup and secure that data to make sure the AI doesn’t get into where it’s not supposed to be.”

Where AI workloads run is becoming part of the same calculus. Leaders described a mix of vendor-delivered SaaS, cloud environments and homegrown models built with local partners.

For HP, this is where responsible AI scale becomes as much an operational technology question as a strategy question. As more AI-enabled workflows move closer to the point of care, healthcare organizations need endpoint environments that are secure, manageable and resilient enough to support sensitive data, budget discipline and frontline trust.

BECKER’S HEALTHCARE // HP 5

AI should support human care For all the discussion of governance, budgets and infrastructure, the leaders kept returning to the same core principles. AI should give time, capacity and attention back to human care. The chief medical officer of an integrated health system in the Midwest framed AI as “a complement, not a replacement for manpower.” He described it as a toolbox for optimizing access, quality, safety and financial performance as healthcare builds its next operating model.

The examples leaders were proudest of reflected that framing: AI that screens medical records and reaches out to patients about medication adherence, cancer screenings and post- discharge follow-up, improving value-based performance and readmission rates for a safety-net population; ambient documentation that reduces cognitive load so physicians can be present with patients; and predictive models that identify sepsis, deterioration and fall risk earlier and connect those insights to workflows nurses can act on.

“Using AI for language translation is a huge opportunity where we can remove some barriers and increase the dignity of care for our patients,” said the Southwest safety-net CIO, whose health system serves patients speaking 80 languages.

The goal, she and others emphasized, is closing care gaps while staying vigilant that new technology never creates them for vulnerable populations.

Building the foundation for responsible AI The organizations that will lead in healthcare AI are not those piloting the most tools. They are the ones choosing the right problems, governing with discipline, bringing their people along and scaling only what they can secure, afford and sustain.

As AI moves from pilots to everyday use, success depends on secure, reliable technology that clinicians can trust and IT  can manage. For AI use cases operating closest to the point of care, that foundation includes a secure, manageable endpoint layer.

HP helps healthcare organizations strengthen that layer to support responsible AI adoption.

To learn how HP supports healthcare organizations with secure, modern technology experiences, visit HP Healthcare.

https://discover.hp.com/healthcare-solutions

CENTRALIZED: TOP AI: Button 3: Button 4: Button 5: Button 6: