The problem
Treasury is the federal government’s central disbursing agency. It says it sends approximately 1.4 billion payments a year, worth over $6.9 trillion, to more than 100 million people, and that fraud and improper payments rose after the pandemic (U.S. Department of the Treasury).
The scale of the leak is large. The Government Accountability Office reports that for fiscal year 2024, agencies estimated $162 billion in improper payments across 68 programs, and about $2.8 trillion since fiscal year 2003 (U.S. Government Accountability Office). Not every improper payment is fraud, but every one is money paid to the wrong person, in the wrong amount, or without the right documentation.
In her written statement to a House Oversight subcommittee in January 2026, Renata Miskell of Treasury described the old approach as a reactive pay-and-chase model and said the goal is to stop improper payments before funds ever leave Treasury (House Committee on Oversight and Government Reform). Check fraud was rising, and recovering a fraudulent check from the bank that cashed it was slow.
What they did
The work sits in Treasury’s Office of Payment Integrity, within the Bureau of the Fiscal Service. According to Treasury, the office expanded risk-based screening, began identifying and prioritizing high-risk transactions before payment, and used machine learning to speed up the identification of Treasury check fraud (U.S. Department of the Treasury). It also changed its payment processing schedule and widened access to its Do Not Pay data, including a May 2024 partnership with the Department of Labor for state unemployment agencies.
Treasury’s 2026 testimony adds detail. After the losses to fraudulent payments during COVID-19, Treasury created a cross-functional “Tiger Team” to study why fraud and improper payments persist and to prototype solutions, working with federal, state and industry partners (House Committee on Oversight and Government Reform). Treasury says it tested and implemented machine learning algorithms “to mitigate check fraud in near real-time”, which strengthened and sped up the process of recovering potentially fraudulent payments from financial institutions.
The model sits inside a wider control system rather than replacing it. Since Executive Order 14249 in March 2025, Treasury has been screening payment files before agency certification: checking accounting codes and available funds, and deploying screening for deceased payees, closed or mismatched bank accounts and invalid taxpayer identifiers. Treasury has not published which vendors, models or data it uses for the machine learning work, or what the program cost.
Results
All figures are Treasury’s own, self-reported and not independently audited in the sources cited here.
According to Treasury’s October 2024 announcement, its technology and data-driven approach prevented and recovered over $4 billion in fraud and improper payments in fiscal year 2024 (October 2023 to September 2024), up from $652.7 million in fiscal year 2023 (U.S. Department of the Treasury). Treasury broke the total down by method:
- Identifying and prioritizing high-risk transactions: $2.5 billion in prevention.
- Expediting the identification of Treasury check fraud with machine learning AI: $1 billion in recovery.
- Expanding risk-based screening: $500 million in prevention.
- Efficiencies in the payment processing schedule: $180 million in prevention.
Only the $1 billion line is attributed directly to machine learning. Treasury’s later testimony says it prevented and recovered over $1.9 billion in check fraud in fiscal years 2024 and 2025 combined (House Committee on Oversight and Government Reform). The testimony also reports that the Tiger Team identified and tested 19 new data sets that could potentially prevent approximately $28 billion in improper payments, an estimate rather than a result.
What didn’t go to plan
The results are real money against a much larger problem. Treasury’s $4 billion for fiscal year 2024 compares with the $162 billion in improper payments agencies estimated for the same year (U.S. Government Accountability Office), and the machine learning line is a quarter of Treasury’s own total.
The binding constraint is data access, not models. Treasury’s testimony says Do Not Pay has been underutilized by agencies because of complex legal and administrative barriers, and that in fiscal year 2024 only four percent of federal programs had completed the legal requirements to access all its databases (House Committee on Oversight and Government Reform). Adding a data source historically took a year or more. Treasury says it still lacks statutory access to identity, income and death data it considers critical, and that legislation has been slow to advance.
Transparency is thin. Treasury has not published model accuracy, false positive rates, how flagged payments to legitimate recipients are resolved, or the cost of the program. The fiscal year 2025 breakdown is not public in the same form, so the combined $1.9 billion check fraud figure cannot be split by year from these sources.
Lessons
- Point the model at the step where money leaves. Treasury’s largest gains came from screening and prioritizing transactions before payment, not from chasing losses afterward.
- Report results by method, not only in total. Treasury’s breakdown shows machine learning delivered $1 billion of the $4 billion; the rest came from screening, prioritization and scheduling.
- Fix data access before buying more models. Only four percent of federal programs could use all Do Not Pay databases in fiscal year 2024, which capped what any model could see.
- Build a recovery path for what the model flags. The check fraud gain came from faster recovery of flagged payments from financial institutions, which needs a process, not only a score.
Adapt it in your organization
Is this for you?
- You make or receive a high volume of payments, refunds, claims or checks.
- You have labeled history of confirmed fraud or improper payments to train on.
- You can act on a flag before payment, or recover funds quickly after it.
- Your legal team can approve the data sources the model needs.
- You can measure prevented and recovered losses separately by method.
A 90-day version
Pilot: score one payment stream, such as supplier payments or customer refunds, with a fraud model run in shadow mode next to existing rules, then route the top-risk decile to investigators. Owner: head of payment operations. Success measure: confirmed fraud caught per 100 investigator reviews at least double the current rules-only rate, with no rise in delayed legitimate payments.
Metric to track
Losses prevented and recovered per method, per quarter, against the estimated total improper payment rate. Treasury moved from $652.7 million to over $4 billion in a year across all methods, with $1 billion attributed to machine learning.
Failure modes to pre-empt
- Reporting a combined total that credits the model for gains from rules and process changes.
- Training a model the legal team later blocks from the data it needs in production.
- Flagging payments with no fast path to stop or recover them.
- Leaving false positives unmeasured, so legitimate recipients wait without anyone counting them.
Questions to take to your leadership team
- What share of our improper payments do we prevent before money leaves, and what share do we chase after?
- Which data sources would most improve detection, and what stops us from using them today?
- How would we know if the model is delaying payments to legitimate customers or suppliers?
Executive FAQ
How much fraud did the US Treasury’s AI recover?
Treasury says machine learning helped recover $1 billion in check fraud in fiscal year 2024, within $4 billion prevented and recovered across all methods.
Is Treasury still using machine learning for fraud?
Yes. Its January 2026 testimony says it prevented and recovered over $1.9 billion in check fraud in fiscal years 2024 and 2025 combined.
About the curator & AI disclosure
Curated by Luciano Ayres
Luciano Ayres is Director of Artificial Intelligence at ArcTouch, with more than 20 years of experience in software and technology leadership and specialized studies in deep learning. Connect with him on LinkedIn.
AI-generated briefing. This case was produced through an automated AI research and writing workflow curated by Luciano Ayres. Sources are linked for independent verification; summaries and analysis may contain errors and do not replace professional due diligence.