Artificial intelligence carries costs that are visible, growing and increasingly tied to token-level consumption. What it delivers in return is harder to quantify and health system leaders say that gap, more than any technology limitation, is the central challenge of getting AI investments approved and sustained.
At Becker’s 11th Annual Health IT + Revenue Cycle Conference in Chicago, leaders from UT Health San Antonio, Northwestern Medicine, Endeavor Health and Aspen Valley Health described how their organizations are building the business case for AI across burnout reduction, risk mitigation, margin improvement and direct revenue, and the difficulty of making those numbers legible to finance committees operating on razor-thin margins.
Nirav Shah, MD, associate CMIO for AI and innovation at Endeavor Health, said the macro financial context makes every conversation with leadership harder. Hospital margins are averaging around 0.75% this year, he said, which intensifies demand for hard return on investment figures that AI investments often cannot immediately produce.
“It is a challenge, but it’s something that we have to continue to push,” Dr. Shah said. “Saving a single patient’s life, while you may not get a hard ROI from that, is significant and is why we kind of go into medicine, why we are in this mission-driven industry.”
Dr. Shah frames hard financial returns as a mechanism for reinvestment rather than an end in themselves. When AI generates measurable cost savings or efficiency gains, those dollars go back into clinical and operational transformation, a pitch he has found more effective with health system leadership than ROI in isolation.
Edward Sankary, vice president and chief health information officer and chief value officer at UT Health San Antonio, the largest academic medical center in South Texas, said his organization uses four value dimensions to evaluate AI investments.
1. Burnout reduction. Dr. Sankary said workforce turnover costs are a more credible ROI proxy than health systems typically use. Replacing a nurse costs roughly $60,000, he said, and replacing a physician is estimated at more than $300,000. Ambient dictation, for example, cannot yet be directly tied to a revenue line but its impact on clinician experience translates into those dollar figures.
“If you take ambient dictation technology by helping physician burnout — while we can’t yet translate it entirely into a direct revenue stream — if you look at the value of it, it’s that improvement in clinician experience that’s bringing value by reducing burnout,” Dr. Sankary said. “There’s worth to that that you can actually translate into those hard dollars.”
2. Risk reduction. Dr. Sankary said UT Health San Antonio built a secure, HIPAA- and FERPA-compliant enterprise platform giving clinical and administrative staff access to multiple large language models — a project partly justified by the cost of not having it. Without a sanctioned internal platform, clinicians will use consumer AI tools for work purposes, exposing the organization to compliance liability.
“We know that clinicians and others will go put HIPAA [information] into open ChatGPT, which places the organization at risk,” he said. “If we can roll this platform out, make it seamless, make it within their normal workflow, we’re reducing the risk to the institution by significant dollars, millions of dollars it’s been estimated.”
3. Financial gain through margin improvement and direct revenue. Dr. Sankary pointed to automated fax referral processing as an example of the margin case: an AI model that extracts information from an incoming referral, verifies insurance coverage, identifies the appropriate physician, creates the referral order and routes it into the EHR.
“We’ve been able to build out this technology that’s actually increasing margin by having our staff, who spend 15 to 20 minutes per referral, be deployed to doing other things for the organization,” he said. “You can cut that down to two minutes per referral.”
On the direct revenue side, Dr. Sankary cited denial appeals and prior authorization as areas where AI is beginning to generate measurable returns, including ambient-assisted prior authorization built directly into the clinical workflow.
Doug King, senior vice president and chief digital information officer at Northwestern Medicine, said the value framework his organization applies has five categories: hard ROI, patient experience, quality, patient safety and revenue. Rather than defining success around a single metric, the categories are agreed upon before a pilot launches.
“AI has got to be additive by removal,” he said. “If it’s on the patient side, it needs to remove friction. If it’s on the physician side, it needs to remove cognitive burden. If it’s within revenue cycle, it needs to remove the manual review of the denials.”
Remove enough friction across enough touchpoints and the downstream gains become measurable: shorter length of stay, higher throughput, reduced pajama time for clinicians, better patient access.
But Mr. King warned that demonstrating value at a moment in time is not sufficient. As token costs evolve and the technology changes, an AI tool that was cost-effective at deployment may not be six months later, or a new capability from a core platform may make a previously valuable third-party tool redundant.
“Just because [the AI tool] doing it now doesn’t mean it’s going to be doing it in the future,” he said. “You need to revisit it. And I think historically that’s something that healthcare has done poorly. We just implement something and then we’re off to the next thing.”
Michelle Gelroth, chief transformation officer at Aspen Valley Health — a 25-bed critical access hospital in Aspen, Colo. — said for smaller organizations, the ROI conversation is less about financial modeling and more about data accessibility. Leadership at her organization has struggled to build trust in data itself, let alone AI-generated insights drawn from it.
“We have a really hard time teaching people to trust data and data literacy for folks that have kind of moved up through nursing or as a physician and teaching them what that data means,” Ms. Gelroth said. “And AI can now — like ChatGPT can just explain this is what’s going on, or you have an opportunity in this service line. I think bringing that data together from a financial and a clinical perspective is going to be really powerful for us in the future.”
