$3.3M, 20,000 hours, 10 points: What AI is actually returning at 5 health systems

Advertisement

Health systems have spent heavily on AI, but specific numbers on what individual tools return are still hard to come by. Much of the evidence is anecdotal, and some benefits, such as clinician time and satisfaction, don’t show up neatly in financial results.

At Becker’s 14th Annual CEO + CFO Roundtable in Chicago this month, leaders from five health systems shared what specific AI tools have delivered, including the extra work and limits that came with them. Some figures are measured results, and some are projections, as noted.

1. Rush University System for Health (Chicago)

Rush’s quickest AI payback has come from SmarterDx, which reviews coded charts before billing to flag missed documentation, said Jeff Gautney, senior vice president and CIO. Rush pays the vendor a contingency fee based on what it finds. Since going live in March, the tool has identified about $5 million in gross revenue, or roughly $3.3 million after the vendor’s fee, well above what Rush had budgeted for the year, he said.

The tool also created more work for revenue cycle staff, who now review charts a second time after coding, though SmarterDx sorts the flagged charts into work queues.

“Now the payback is worth it, but it’s an unintended consequence that we originally didn’t plan for,” Mr. Gautney said.

Separately, a generative AI tool for registry abstraction let Rush end a $1 million outsourcing contract for a couple of service lines, he said. The five employees who had done part of that work were not redeployed.

2. Sutter Health (Sacramento, Calif.)

Sutter’s ambient documentation tool from Abridge has more than 4,000 active users and has saved its physicians about 20,000 hours, said Rick Moreland, CEO of Sutter Delta Medical Center in Antioch, Calif. Sutter also tracked how mentally demanding clinicians found charting, and that score fell from 12.2 to 6.3, he said.

“I was in a group meeting with physicians and literally out of 1,000 physicians, I almost got a standing ovation,” Mr. Moreland said. “Well, not me, but the system for helping our clinicians really give time back to them.”

3. Boston Medical Center Health System

BMC piloted Abridge’s ambient documentation with about 80 physicians before scaling it to all clinicians in October 2025, said Tinu Tadese, MD, who until recently was vice president and system CMIO. The pilot cut the time it took clinicians to close their notes, but that difference largely disappeared once the tool scaled, because many factors affect when notes get closed, she said.

Coding moved more clearly. Before Abridge, 61% of clinicians’ evaluation and management visits were billed at levels 4 and 5, the two highest complexity levels. After, that rose to about 71%, she said.

“We haven’t tracked that to revenue, but we actually wanted to see what will happen there,” Dr. Tadese said.

4. Emory Healthcare (Atlanta)

Emory expects some of its revenue cycle AI deployments to return between $15 million and $25 million annually, said Nabile Safdar, MD, chief AI officer of Emory Healthcare and Emory University. Revenue cycle is the easiest area to measure because the results show up in dollars, he said. Calculating returns across the full AI portfolio is much harder, partly because AI is embedded in software such as the EHR.

“In other areas, frankly, it’s going to be very hard to measure, and it’s kind of more like how do you measure the ROI of electricity? AI is just coursing through the veins of the system now,” Dr. Safdar said.

5. HonorHealth (Scottsdale, Ariz.)

HonorHealth’s clearest returns have come from predictive models, said Jim Whitfill, MD, senior vice president and chief transformation officer. Its EHR vendor’s deterioration index has produced a measurable clinical benefit by identifying patients at high risk of deteriorating over the next six to 12 hours, he said. Work with a care operations company has cut excess inpatient days and freed up capacity, and the system has also used AI for operating room optimization.

“Those are two areas where, again, we’re not using agentic AI. We’re not using the latest LLMs,” Dr. Whitfill said. “We’re just doing, I won’t even say simple, but predictive models that definitely are creating return.”

Advertisement

Next Up in Innovation

Advertisement