Mayo Clinic has moved well beyond AI pilots. Now it is learning what it costs to keep AI running at scale.
The Rochester, Minn.-based health system has 128 clinical AI solutions in practice and is approaching 500 in its pipeline, according to Micky Tripathi, PhD, Mayo Clinic’s chief AI implementation officer.
“Most of the ones that we have right now, probably 70%, are internally developed solutions,” he told Becker’s. “Some of them we have a partner for, but we are ultimately, for those solutions, the legal manufacturer.”
The other 30% are a variety of vended products, with an increasing number coming from EHR vendor Epic. These groups of solutions are about removing friction in healthcare.
“They’re basically not about doing anything that a human can’t do, but they’re about helping humans do things better, faster and cheaper than we’re able to do today,” Dr. Tripathi said.
For example, the health system is a big adopter of ambient documentation, with 80% of Mayo providers using the technology across a few different platforms. The health system has also developed Record Time, an AI tool that helps clinicians process outside medical records for patients coming to the health system for care.
Record Time automatically gathers records arriving through sources such as faxes and interoperability networks, then scans, catalogs and indexes the information and creates summaries tailored to what the reviewing clinician needs.
“That’s been shown already to save something like 11 minutes per patient, which is huge when you add up how many patients we have and the amount of staff work that goes into processing those records and getting them to the right clinician at the right time,” he said.
But one of Mayo’s biggest AI ambitions is a patient-facing avatar. Dr. Tripathi said Mayo is currently testing the avatars in areas including genetic counseling and post-discharge catheter management. The idea is for patients to be able to ask questions, receive instructions and, in some cases, complete an initial interview with the AI avatar before connecting with a clinician.
For genetic counseling specifically, Dr. Tripathi said the goal is partly to expand access when specialists are in short supply. Instead of waiting months for an initial interview, a patient could choose to interact with the avatar sooner, including evenings or weekends, while still retaining the option to meet with a person.
“It’s a pretty advanced technology,” he said. “Because just getting the avatar to be able to have the right persona so that a patient feels comfortable interacting with it, and to make sure that it’s accurate and reliable, is itself a pretty big advance.”
Those use cases illustrate the clinical and operational gains Mayo is already seeing from AI. But as the health system moves more tools from development into everyday use, Dr. Tripathi said another reality is becoming clearer. Deploying an AI tool is only the beginning of its cost.
“The cost of maintaining these systems is far higher than I think any of us in the industry really thought,” he said.
Dr. Tripathi pointed to three areas where AI costs are emerging. They include the ongoing computing costs required to run AI models, the specialized workforce needed to monitor and maintain them, and the data infrastructure required to support a growing portfolio of AI tools.
Unlike traditional software that can largely remain unchanged after deployment, AI models require continued oversight. Models can drift, underlying data can change and health systems need staff with the technical expertise to continually monitor how the tools are performing.
“I already don’t have enough AI/ML engineers to create these products,” Dr. Tripathi said. “And if those engineers can never let go of the product because they’re needed to maintain and oversee things like data drift, model drift, all of those things, which are critically important, that starts to become a big challenge.”
That challenge only grows with scale. Mayo has 128 clinical AI solutions in practice today, but Dr. Tripathi said that figure could eventually grow, creating both a workforce and financial question around how health systems maintain increasingly large AI portfolios over their lifecycles.
How health systems ultimately pay for those growing costs, however, remains an open question.
Dr. Tripathi said Mayo has the benefit of philanthropic funding and other resources that allow the health system to continue investing in AI even when there is not yet a clear financial return tied to a particular tool.
“Right now, from our perspective, I think the answer is we don’t really know,” he said of how AI tools will ultimately be paid for.
Mayo’s leadership has directed the organization to continue moving technologies into clinical practice when they can improve patient outcomes without becoming overly focused on an immediate return on investment.
“Some of these things, there is no line-of-sight ROI because you’re just improving quality in a system that doesn’t compensate for higher quality right now,” he said.
That leaves a broader question for Mayo and other health systems as AI becomes more embedded in care. Who ultimately pays for the technology, infrastructure and workforce required to sustain it?
Dr. Tripathi said federal policymakers are beginning to consider how AI-enabled care could be compensated through areas such as inpatient and outpatient payment rules, but he said the discussion remains early.
“We’re just going to keep pushing really, really hard on that and hope that these questions of affordability and how these get compensated get sorted out,” he said.
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.
