The U.S. Treasury announced on October 17, 2024 that it had prevented or recovered more than USD 4 billion in fraud and improper payments during fiscal year 2024, and its own breakdown names machine learning for only one of four items: USD 1 billion recovered from Treasury check fraud. That split is useful for anyone weighing a claim that AI is reshaping an industry, because a headline total usually covers a whole program while the AI share is one slice of it.
Key takeaways
- Treasury's Office of Payment Integrity reported more than USD 4 billion prevented or recovered in fiscal year 2024, up from USD 652.7 million in fiscal 2023.
- The release ties machine learning to one item only, USD 1 billion in recovered check fraud. The other three items (USD 500 million, USD 2.5 billion, and USD 180 million) are not described as AI.
- Prevented dollars and recovered dollars are different kinds of result, so adding them into one "AI savings" figure hides an assumption.
- The Government Accountability Office (GAO) counted USD 162 billion in improper payments across federal programs for the same fiscal year, which is a separate measure and not a denominator for Treasury's total.
- Five questions, listed in a table below, can be applied to any report of AI changing a sector.
What Treasury actually reported
The announcement came from the Office of Payment Integrity, which sits inside Treasury's Bureau of the Fiscal Service and works on federal payment integrity. Its press release of October 17, 2024 gave one total and four lines beneath it. The fiscal year ran from October 2023 to September 2024, and the comparison year, fiscal 2023, was reported at USD 652.7 million. A sixfold jump in one year is the kind of figure that tends to be repeated without its components, so the components are set out below.
| Line in the release | Amount | Type of result | Release names machine learning or AI? |
|---|---|---|---|
| Expanding risk-based screening | USD 500 million | Prevention | No |
| Identifying and prioritizing high-risk transactions | USD 2.5 billion | Prevention | No |
| Speeding up identification of Treasury check fraud | USD 1 billion | Recovery | Yes, machine learning |
| Changes to the payment processing schedule | USD 180 million | Prevention | No |
Some secondary coverage described the entire USD 4 billion as the work of AI. The primary text supports a narrower reading: a mix of screening, prioritization, scheduling, and one machine learning tool, with AI explicitly credited for a quarter of the total.
Prevention and recovery are not the same result
Recovery means money that had already left the government and was brought back. It can be traced to specific payments. Prevention counts payments that were stopped before release, so the figure describes something that did not happen and depends on a judgment about what would have been paid otherwise. Three of the four lines in the table are prevention. The four lines are totals only, and a reader cannot tell from them how much of the USD 2.5 billion would have been caught by older methods anyway.
None of this means the figures are wrong. It means each one answers a different question, and a sentence such as "AI saved USD 4 billion" combines answers that were never measured the same way.
How the USD 162 billion GAO figure fits
GAO's March 11, 2025 release reported that agencies identified USD 162 billion in improper payments for fiscal year 2024, across 68 programs at 16 agencies. About USD 135 billion of it, roughly 84 percent, was overpayments. Medicare, Medicaid, the Earned Income Tax Credit, the Supplemental Nutrition Assistance Program, and the Restaurant Revitalization Fund together made up 75 percent. GAO defines improper payments as those that should not have been made or were made in the incorrect amount, which covers clerical error and missing documentation as well as fraud.
Treasury's total and GAO's total measure different things. One counts dollars stopped or recovered in Treasury's payment channels; the other counts errors estimated by program agencies. Placing one over the other to get a percentage would mix two scopes, and neither release invites that calculation.
Five questions for any "AI is reshaping this industry" claim
The Treasury case suggests a short screen that works on press releases, annual reports, and vendor case studies alike. The questions are illustrative, built from the gaps seen above rather than taken from a standard.
| Question | What the Treasury release showed |
|---|---|
| Does the total belong to the AI tool or to the whole program? | The total covered four initiatives; AI was named in one. |
| Is the result a prevention estimate or money actually recovered? | Three lines were prevention, one was recovery. |
| Is there a comparison year, and what else changed between the years? | Fiscal 2023 was given as USD 652.7 million; the four lines carry no controlled comparison. |
| Who issued the number, and can the primary document be opened? | A dated Treasury release exists and can be read in full. |
| What is the denominator, if a share is implied? | GAO's USD 162 billion uses a different scope, so no share can be computed. |
What the case does and does not show
A federal payment office reported large dollar totals after adding screening, prioritization, scheduling changes, and a machine learning check-fraud tool. The release supports the statement that machine learning was part of a USD 1 billion recovery. It does not support a statement that AI alone produced the whole result, and it does not show how the same tool would perform in a private bank, a hospital, or a retailer. The October 2024 figures are the ones verified here, and later fiscal years may differ.
FAQ
Did AI recover USD 4 billion for the Treasury?
No. Treasury reported more than USD 4 billion prevented or recovered through several measures. Only the USD 1 billion recovery from Treasury check fraud is described as using machine learning.
What is an improper payment?
GAO describes it as a payment that should not have been made or was made in the incorrect amount. Fraud is one cause, and administrative mistakes are another.
Is the USD 162 billion in improper payments all fraud?
No. The figure is an estimate of payment errors reported by agencies, and it includes overpayments and underpayments from many causes.
Why use a government example in an article about industries?
A dated public release with line items can be checked by any reader, which is rarely true of company announcements. The same questions can then be carried over to claims in other sectors.
Related reading
- which industries data show as least exposed to AI
- how network architectures match tasks
- Canada's survey results for health care and other sectors
- an open-source planning and routing optimizer
More in AI and Society.
Last verified: October 2026
Sources
- Treasury Announces Enhanced Fraud Detection Processes, Including Machine Learning AI, That Prevented and Recovered Over $4 Billion in Fiscal Year 2024, U.S. Department of the Treasury, October 17, 2024
- GAO Reports Estimated $162 Billion in Improper Payments Across the Federal Government for Fiscal Year 2024, U.S. Government Accountability Office, March 11, 2025