For hospitals across the U.S., launching AI pilots over the last 18 months took considerable work. But that was the easy part compared to what’s next.
Getting a vendor in the door, running a proof of concept, generating promising numbers in a controlled environment — health systems have gotten very good at that. What most haven’t figured out is what comes after: how to move AI from a carefully managed experiment into the fabric of clinical and operational workflows, with the governance structures to keep it safe, compliant and trustworthy over time.
A growing number of health system leaders say that transition is no longer optional. The question isn’t whether to scale AI, it’s whether the organization has the infrastructure to do it without the wheels coming off. Mark Mabus, chief medical informatics officer at Parkview Health, is one of those leaders. Fort Wayne, Ind.-based Parkview invested heavily in AI-powered tools to increase productivity and now aims to accelerate adoption.
“Generative AI is what everybody needs to be doing across their healthcare system to enable these tools for efficiency and ROI,” he said. “It’s here. Use it.”
Generating real results
The move from pilot to production was deliberate and fast. Dr. Mabus deployed ambient documentation systemwide in early 2025, then followed in October with a simultaneous rollout of 18 generative AI features in Epic to all providers across all care settings.
“We moved from limited pilots to broad access and made sure that we were able to deploy these tools across all areas of care delivery — ambulatory, emergency department, and inpatient,” he said.
The usage data tells the story. Before October, Parkview was processing roughly 60,000 AI tokens per month. By the end of that month, the number hit 2 million. By January 2026, it reached 3 million. The scale was real and so was the work that preceded it.
Dr. Mabus treated the October launch the way a health system treats an EHR go-live — not as a software update, but as a care delivery event requiring all-hands preparation, department-by-department training, and boots-on-the-ground informatics support. The informatics team deliberately scaled back routine clinic support to focus on the high-priority deployments.
“Think of it like a mini EHR go-live,” he says. “You want to be as prepared as possible because these are high impact tools and a different type of tool than they’ve had available to them in the past.”
At Pittsburgh-based Allegheny General Hospital, the post-pilot discipline looks different but reflects the same underlying commitment. Imran Qadeer, MD, president and CEO, deployed Care.ai for virtual nursing and virtual sitting across the hospital, and Abridge for ambient documentation — not as experiments, but as operational programs with defined outcome targets, and recorded 50 minutes saved per nurse per shift. On the revenue cycle side, the team projected 5% improvement in claim submission accuracy.
What keeps those results from being aspirational is a monitoring framework built into the deployment from day one. “We’re constantly looking at: is the technology delivering on the promise that it’s set to do,” Dr. Qadeer says. “We focus on a very result-oriented launch of these products.”
The governance problem most don’t plan for
Deploying AI at scale creates a problem that pilots never surface: you have to manage what you’ve built. Models drift and bias creeps in as patient populations shift. Tools that performed well at launch may not perform the same way six months later — and without a governance structure designed to catch those changes, a health system may not know there’s a problem until it shows up in outcomes.
Dr. Mabus is direct about what the post-deployment era actually demands.
“Now that all these AI functionalities are live, you gotta make sure there’s ongoing review, make sure there’s not bias or drift that happens with some of the models, making sure that these tools are still functioning in a safe and high quality manner,” he said.
Parkview manages that through an AI steering committee to manage changes as issues as they arise quickly. The focus on governance for 2026 reflects how completely the conversation has shifted.
“This year is the age of continuing education for users,” he said. “It’s the age of inventories, libraries, and feedback mechanisms and upkeep, all while weighing additional options with not just generative AI solutions but agentic and even perhaps autonomous AI in the coding arena.”
That framing resonates with Mayil Dharmarajan, vice president of data and analytics at UC Irvine Health. The Southern California academic medical center completed an ambitious simultaneous IT transformation this past December, integrating four acquired hospitals, bringing a new $1.2 billion facility online, and migrating to its own Epic instance, all at once. The technology infrastructure is now in place. The governance work is what Mr. Dharmarajan is focused on next.
“We are shifting ourselves from the typical AI pilot demos to how we productionize those AI projects and make sure that appropriate governance structure is put in place and enable our people across our organization, from operations to clinical, to use the technology,” he says.
For UC Irvine, the governance challenge is compounded by its academic medical center status. The institution has to manage data across clinical care, hospital operations, and research, each with different regulatory requirements, privacy obligations and access controls. For research, that means IRB compliance is layered on top of standard data governance. It also means responsibly managing access to an extraordinary asset: through its Epic instance, UC Irvine’s researchers now have access to 300 million de-identified patient records globally.
“Now we have so much access to the data,” Mr. Dharmarajan said. “How do we make sure that this is safe? That’s where the data governance for that specific real-world data comes into play. How do we make sure that the right people are having access, how do we make sure that they are trained to use the data in an appropriate way.”
The discipline of saying no
One of the counterintuitive lessons of the post-pilot era is that scaling AI well sometimes means slowing down. The pressure to deploy every new capability as it becomes available is real — and it’s a trap. Dr. Mabus names it plainly. The hardest thing he’ll have to do in the coming year isn’t deploying more AI. It’s restraint, which is a different mindset than what leaders needed during the early phases of AI piloting.
“We don’t have unlimited resources,” Dr. Mabus said. “What makes sense for us may be something that is provided through an existing vendor rather than some new latest and greatest things.”
The risk of moving too fast isn’t just financial; it’s also cultural. Health systems that overwhelm their users with tools, especially clinicians already navigating change fatigue — risk eroding the trust that made early adoption possible.
“We don’t want AI to overwhelm all of our users,” Dr. Mabus said. “We have to balance the yes’s with the no’s.”
Mr. Dharmarajan applies the same discipline to governance itself. The temptation when building a data governance program is to try to solve everything at once. His advice: don’t.
“Data governance is a very broad subject. It’s very tough for us to start all the components of the data management, data governance aspects of it,” he says. “We need to focus on some of the critical ones — what would benefit UCI Irvine for next three months, six months, one year — and then address all those specific components and create the policy, procedures, and processes for that and execute that so that we don’t boil the ocean.”
The sequencing matters as much as the structure. Dr. Mabus has built vendor intake governance into the procurement process itself — requiring any new AI request to first answer whether an existing trusted vendor already offers a comparable solution. The process catches shadow AI procurement before it starts and gives the informatics team an early opportunity to redirect requesters toward tools already in the stack.
The next frontier
Even at systems that have moved decisively beyond AI pilots, a new challenge has emerged: availability doesn’t equal adoption. Dr. Mabus is candid about the gap. Despite broad deployment, roughly a third of providers are high users, a third use AI tools occasionally, and a third remain largely disengaged.
“Our focus is going to be this year on the education of that middle and lower thirds,” he says. “Maybe we can hit 50-50 or 60-40 this year. Education and perhaps even one-on-one training on some of these things may help us move that needle.”
That gap between deployment and genuine adoption is where the real work of the post-pilot era lives. Standing up a tool is a project. Changing how an institution practices medicine is a transformation — and it runs on a longer timeline, requiring sustained investment in training, feedback loops, and leadership attention well after the go-live date has passed.
“The closer we get to the business and leverage all of our technology knowledge along with the business challenges our business team is having and then build that partnership on a regular basis and deliver value through that particular process,” said Mr. Dharmarajan. “It is also imperative to make sure that we keep in touch with our leaders on a regular basis. Whether it’s the C-suite, VPs or another senior leadership role, we want to connect every month minimum to make sure that we understand their challenges today and what are the opportunities we have to solve it through data.”
The health systems making the most durable progress are the ones that understood this before they flipped the switch. The pilot era tested whether AI could work in healthcare. The question is whether health systems can build the organizational infrastructure — the governance, the discipline, the culture — to make it work at scale, sustainably, over time.
At the Becker's 11th Annual IT + Revenue Cycle Conference: The Future of AI & Digital Health, taking place September 14–17 in Chicago, healthcare executives and digital leaders from across the country will come together to explore how AI, interoperability, cybersecurity, and revenue cycle innovation are transforming care delivery, strengthening financial performance, and driving the next era of digital health. Apply for complimentary registration now.
