Flipping a switch is easy. What happens after is where AI deployments can come apart.
Mark Mabus, MD, chief medical informatics officer at Parkview Health — a nonprofit system spanning Northeast Indiana and Northwest Ohio — spent 2025 getting generative AI into the hands of every provider, across every care setting, at scale. Ambient documentation went live systemwide in early 2025. In October, 18 generative AI features in Epic rolled out to all providers simultaneously.
The results were measurable almost immediately. Before the October launch, the system was processing roughly 60,000 tokens per month. By the end of October, that number hit 2 million. By January 2026, it reached 3 million.
“We saw just a little bit of engagement here and there, getting our pilots underground and then made it available throughout the whole health system,” Dr. Mabus said. “People are using them.”
But the token counts, impressive as they are, aren’t the point. What Parkview’s experience reveals is a more nuanced story about what it actually takes to move AI from pilot to institution and what comes next once you do.
The gap between a vendor demo and clinical reality is where AI implementations typically stall. Dr. Mabus learned early that closing that gap required something instruction manuals can’t provide.
“A lot of it is figuring out workflows and where this fits properly,” he said. “You can have your vendors give you the instruction manuals on how this actually works technically. But when you dive into the workflows of an actual user and start using these things, you get to figure out that maybe it’s not so cut-and-dry from the instruction manual.”
Parkview used what Epic calls “physician builders” — clinicians with deep system fluency — as the connective tissue between technical capability and real-world use. Critically, those builders weren’t just testing for themselves.
“Those physician builders are also in tune with non-savvy users’ workflows too,” Dr. Mabus said. That orientation toward the full adoption spectrum — early adopters, middle-of-the-road users, and those more resistant to change — shaped how best practices, training videos, and in-basket education were developed before anything went to the broader population.
The October launch was treated less like a software update and more like a care delivery event.
“Think of it like a mini EHR go-live,” Dr. Mabus said. “You want to be as prepared as possible because these are high impact tools and kind of a different type of tool than they’ve had available to them in the past.”
Ambient documentation’s value proposition is well-rehearsed at this point: less time on notes, more time with patients, reduced burnout. Dr. Mabus doesn’t dismiss those benefits but he’s also frank about the limits of soft ROI when finance teams are asking harder questions.
“If you have an extra fifteen minutes, were you actually using that fifteen minutes a day to throw another patient on?” he said. “Those questions were, of course, are big on our finance side.”
Early in the pilot, Parkview required participating providers to add an incremental patient per clinic day to help offset costs. As adoption widened and burnout metrics improved, that requirement softened but the underlying discipline around measurement didn’t disappear.
Where Dr. Mabus sees the clearest financial case is on the nonclinical side. Coding and revenue cycle applications are “no-brainers to implement.” An AI-assisted E/M code calculator that reads a provider’s note and flags the appropriate level of service, for instance, generates measurable coding accuracy improvements. Faster inpatient discharge summaries freed physician time with a quantifiable downstream effect.
The strategic directive then is to let hard-ROI use cases fund softer ones.
“Finding those key AI functionalities where you absolutely can see a financial ROI is key to getting things started,” Dr. Mabus said. “Then those help to pay for some of the other functionalities that may have more of that soft ROI based more on efficiency and time savings. I can use some of these to pay for those.”
His broader advice for budget-constrained systems is to resist the pull of novel vendors when existing contracts may already cover the need.
“There are some health systems that pay seven figures or more to embed generative AI now,” he said. “Why not use those as opposed to searching out a third party?”
The work of genuine AI adoption is just beginning for many health systems as utilization spikes and leadership teams develop governance structures. Parkview operates an AI steering committee charged with ongoing inventory, bias monitoring, and drift detection across deployed tools. New vendor requests now flow through an updated intake process that explicitly asks whether an existing trusted vendor already offers a comparable solution.
“Embedded within that, we now created our own AI review,” Dr. Mabus said. The process flags requests for informatics review before budget approvals are sought — a structural check against the sprawl of shadow AI procurement.
“Are there things that need to wait? We don’t have unlimited resources,” said Dr. Mabus. “What makes sense for us may be something that is provided through an existing vendor rather than some new latest and greatest things.”
Resource management isn’t the only limiting factor. Leadership teams are also cognizant of not pushing their teams too hard in transformation. Systems that sprint into every new capability risk overwhelming users, fragmenting governance, and eroding the trust that early AI adoption was designed to build.
“We don’t want AI to overwhelm all of our users,” Dr. Mabus. “We have to balance the yes’s with the no’s.”
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.
