Medicare loses an estimated $100 billion a year to improper payments. In just over 12 months, AI helped the Centers for Medicare and Medicaid Services prevent $2.1 billion in improper payments from going out the door.
That sounds like a lot until you do the math. It's about 2% of the annual problem.
CMS processes roughly 5 million claims every day. Human reviewers used to examine less than 1% of them. Everything else went through, and the agency would spend months or years trying to claw the money back. Kim Brandt, CMS's Deputy Administrator and COO, described the old approach to Congress as moving from a "pay and chase" strategy to a "caught and stopped" one, which is a polite way of saying the previous system was structurally designed to lose.
The new system runs AI that screens claims in real time before any money moves. Brandt described it as a "Netflix-type algorithm" that scores incoming claims against historical patterns and flags anything suspicious, like one patient supposedly receiving thousands of procedures across multiple states in a matter of days. Within the first year, she said the AI helped eliminate 90% of some categories of bad actors.
The AI doesn't make the final call, though. "AI is great to help us say, 'Hey, here are areas you want to focus on,'" Brandt said at HIMSS, "but then we need to actually validate that. A doctor or clinician is the one who reviews the data."
Getting that validation process to work required CMS to change how different parts of the agency talk to each other. Bethany Messick, Acting Deputy Director of CMS's Center for Program Integrity, put it simply: "All we did was get everybody in a room together. We got our investigators, our policymakers, our data scientists, our legal counsel."
CMS also built a training operation around the technology. Brandt told attendees at the AI for Government Summit that senior leadership now completes an hour of AI training every week, designed by University of Chicago economist Anup Mulani. Thirty employees were trained as "AI ambassadors" who fan out across the 6,000-person workforce. CMS's internal AI Slack channel grew from fewer than 80 members in 2021 to over 900 by last year. Separately, the agency deployed a contract analysis tool called CLAW that Brandt said has saved "hundreds of millions" across $8.3 billion in annual contracts by catching overpriced labor rates before deals get signed.
The results are real, but so is the gap. Two out of three Medicare beneficiaries say they don't trust AI, according to Brandt. Medicare Part A's trust fund is projected to deplete in roughly seven years. "If we can extend that, thanks to some of these efforts," Brandt said on the Medicomp podcast, "then to me, AI has definitely proven its worth."
CMS built an actual playbook for deploying AI in any operation where you're processing more transactions than your people can physically review. Mandatory executive training, internal ambassadors, AI that flags and humans who decide. Most organizations haven't built that scaffolding and probably need to. But the math is still severe, with seven years of trust fund runway and roughly $98 billion in improper payments still walking out the door annually. Whether this turns out to be a proof of concept or a very expensive rounding error depends entirely on how fast CMS can close the gap.
