Using AI Agents to Uncover Evidence of Medicare Fraud
Coffeezilla had a line in his recent video when he said that some Medicare fraud is so blatant a 10-year-old could find it.
So instantly it got us thinking... "I bet AI could probably spot something similar." And within about 30 minutes we had a shortlist of providers with extremely suspicious billing practices, the majority of which had actually been prosecuted or banned by CMS - proving the signals we found were real evidence.
Here's what we found
1. Phantom medical supplies.
Several companies showed the signature of billing for supplies nobody ordered. The national total for one wound-dressing code ran about $23M a year, then jumped to $650M in 2024. One supplier in Florida accounted for 62% of that, going from about $0.6M for fewer than 500 patients in 2023 to almost $860 million and 127K patients in 2024. A supplier in New York went from roughly zero to $473 million. A third, in Pennsylvania, billed almost 90% of all drainable ostomy pouch volume in the country.
2. Orders the doctors never wrote.
The data show which clinician signed each order. For ostomy pouches in 2024, 271 clinicians nationwide had more than ten Medicare patients on them, and 213 are family-practice doctors. Ostomies are created and managed by surgeons; a family doctor might have a handful. Yet these family doctors each show about 175 ostomy patients, across 40 states, at nearly identical patient counts and dollar totals, roughly half a million dollars apiece. Colorectal surgeons, the people who actually do this work? Fifteen. Hundreds of unrelated family doctors don't independently become high-volume ostomy prescribers in one year. A real doctor's license number attached to orders they never wrote does produce that picture. Catheters show the same shape: 1,343 of 1,720 ordering clinicians are family practice, and the entire file contains seven urologists.
3. Labs that run every test on everyone.
Two genetic labs in the same office park billed 23% of the whole country's epilepsy gene-panel volume. The bigger one billed nearly the same patient count, about 4,300, under four different panels, plus 6,000 to 7,000 patients each on a dozen single-gene tests, and went from $4.6 million in 2023 to $60 million in 2024.
4. Some are still in business.
The most flagrant suppliers had their Medicare enrollment revoked or were excluded by HHS-OIG between 2024 and June 2026, after the 2024 money moved. Several others with the same profile, including both labs above, are still enrolled as of the July 2026 enrollment file, with no public record against them. We sent those to the OIG hotline. We are not naming them here; a pattern is a lead, not a finding.
How we worked with AI
1. Leverage the official CMS API. CMS's `data.json` lists every dataset and year. We used the "by Geography and Service" files, pulled national totals per billing code for 2021 to 2024, and flagged codes that grew 4x in three years or 2x in one. Expect false positives on the first pass: new drugs, vaccines, patented single-lab tests. We told it to filter those and explain why.
2. Look and who is billing against each code. The supplier-level and provider-level files show every biller's share, city and patient count. One biller or one office park holding most of a national code is a strong signal.
3. Figure out who ordered it. The referring-provider file shows the ordering clinician's specialty and volume. We asked for counts, states and medians only, never names; the clinicians on those rows are probably victims.
4. Check before believing anything. Every flag went against three public lists: CMS's revoked providers, OIG's exclusion list, and the current enrollment file. This step turned a list of outliers into two short lists: already caught, and still enrolled.
5. Verify twice. A second, independent run reproduced all 33 figures from the live API before we wrote this.
Run it yourself
We packaged the method as an Agent Skill: hard rules first, then the catalog lookup, the screens, the false positives, the enforcement check, a self-check table so you know your pipeline works, and a report format.
It is one file: SKILL.md. Download the skill, hand it to your agent and see what you can find.
Remember, outliers in the data are not confirmation of fraud and can come from perfectly legitimate business practices. Never name individuals. Report the specifics around what you find to the HHS Office of Inspector General at tips.oig.hhs.gov. Don't post unproven claims to social media. The skill is CC BY 4.0: copy it, fork it, improve it, and send us the patterns you found, not the names.