MMOOZZGG.com

AI against corruption

AI fraud detection in US government: checking before the money leaves

6 min read3 sources

AI fraud detection in US government moved the check from after payment to before it. With machine learning, the Treasury prevented and recovered more than 4 billion dollars in fiscal year 2024, and the PRAC committee calculated that matching applications before payment could have prevented about 79 billion dollars of possible fraudulent pandemic aid payments.

On this page · 5 sections

In short

AI fraud detection in US government works before payment: machine learning flags risky payments, and an employee decides whether to stop them. In fiscal year 2024, the Treasury prevented and recovered more than 4 billion dollars of fraudulent and improper payments, against 652.7 million a year earlier. The GAO estimate sets the scale of the task: 233–521 billion dollars of losses a year.

The scale: the GAO estimate

In April 2024, the US Government Accountability Office (GAO) estimated the direct annual losses of the federal government to fraud at 233–521 billion dollars. The estimate is based on data for fiscal years 2018–2022.

That is 3–7% of average federal obligations. GAO notes that this percentage cannot be applied to single agencies or programs: each program has its own risk. The White House Office of Management and Budget (OMB) disagreed with the estimate.

The wide range says the main thing: nobody knows the exact sum, because fraud stays hidden until it is found. And GAO counts only direct financial losses; indirect damage is not in the estimate.

For a manager, the conclusion is practical. If losses are measured in percentages of spending, sample checks after payment do not cover the risk, and the check has to move to the point before payment.

Treasury: $4 billion in a year

In fiscal year 2024, the US Department of the Treasury prevented and recovered more than 4 billion dollars of fraudulent and improper payments. A year earlier the result was 652.7 million dollars.

The sum comes from several tools:

1. Machine learning on check fraud recovered 1 billion dollars.

2. Screening and prioritizing high-risk transactions stopped 2.5 billion.

3. Risk-based screening prevented 500 million.

4. Changing the payment schedule gave another 180 million.

The work is done by the Office of Payment Integrity at the Treasury. According to an agency representative, machine learning reviews large data sets and flags possible fraud schemes. The Treasury gives no details, citing the nature of the schemes.

The scale explains why the check sits here. Every year the Treasury makes more than 1.4 billion payments worth more than 6.9 trillion dollars to more than 100 million recipients, so it sees the flows of all programs at once. How such signals look in the data is covered on the page about duplicate and false payments.

An important caveat: the 4 billion covers both fraud and improper payments. An improper payment is not always intentional: the money may have gone to the wrong recipient or in the wrong amount, including through the agency’s own error.

PRAC: $79 billion that need not have been paid

In June 2025, the Pandemic Response Accountability Committee (PRAC) calculated how much matching applications before payment could have saved. Its conclusion: about 79 billion dollars of possible fraudulent pandemic aid payments.

The method is simple. Analysts randomly selected 662,000 records from 67.5 million approved applications and asked the Social Security Administration (SSA) to check the Social Security number, name, date of birth and death indicator.

SSA flagged almost 24,000 records in the sample as possible fraud: the number had never been issued or did not match the name or date of birth. Across all applications, this is about 1.4 million questionable numbers.

More than 11,000 other records in the sample belonged to people who were dead as of the date of the SSA reply. PRAC did not include them in the estimate, because these people may have been alive when they applied. This is an example of the caution without which matching turns into accusation.

Three programs are involved: the EIDL loans and the PPP program of the Small Business Administration, and the Department of Labor’s pandemic unemployment benefits worth 888 billion dollars. As early as January 2023, PRAC showed 69,000 questionable numbers that had received 5.4 billion dollars in loans and grants. How such signs are found in support programs is covered on the page about subsidies and grants.

Checking before payment

Both stories lead to one conclusion: it is cheaper not to pay than to recover the money later. PRAC says directly that you need to verify the applicant’s identity and flag anomalies in the application before the money leaves.

To do this, PRAC matches data from different agencies against what the applicant stated: IP addresses, dates of birth, bank accounts, home addresses. For example, one account shared by many applicants or someone else’s date of birth becomes a risk sign.

The Treasury has the Do Not Pay service for the same task. The agency checks a recipient against several data sources before paying a vendor, a grant recipient, a borrower or a benefit recipient. One of the sources is the Social Security Administration’s register of deaths (Death Master File).

The Treasury’s services are open to other federal agencies and to programs that states run with federal money. In 2024, the Department of Labor and the Treasury announced simpler access to Do Not Pay for state unemployment agencies.

Each such check gives a signal, not a verdict. A data mismatch may be a typo, so the decision to hold or reject a payment stays with a program employee. Good practice: give the applicant a way to confirm the data before refusing.

What the US experience teaches

The US experience gives four rules for any public organization that handles payments, subsidies or contract payments:

1. Measure the scale first. The GAO estimate of 233–521 billion dollars a year turned talk about fraud into a number you can compare with the cost of control.

2. Match before paying. The PRAC example shows that checking the number and date of birth at intake could have prevented about 79 billion dollars. Getting money back after payment is always harder than not paying it.

3. Make tools shared. The Treasury keeps the check in one place and opens it to other agencies, so each program does not need to build its own.

4. Separate fraud from error, but count both. The Treasury reports one total, and that is honest: a mistaken payment also leaves the budget.

How to calculate such an effect for your organization is covered on the page about the effect of AI in control.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. GAO-24-105833, 16 April 2024 (gao.gov)
  2. Nextgov/FCW, 18 October 2024 (nextgov.com)
  3. Pandemic Response Accountability Committee, 4 June 2025 (pandemicoversight.gov)

Found a mistake? Tell us

Publications

Mike Mozg

Agentic AI use cases at AT&T: nearly 500 processes now run on agents

AT&T no longer uses AI only to help staff work faster: it is rebuilding whole business processes around agents. SiliconANGLE reported on 14 November 2026 that nearly 500 agentic AI use cases at AT&T are already in production. Here is how the company rebuilds them, where a human still checks the agent, and what other large companies can take from it.

News4 min

For government bodies

We build such systems for government bodies and state companies, from data analysis to a pilot on one scheme. The data stays on the organization's own servers.

Contact us