MMOOZZGG.com

AI against corruption

Grant fraud detection: checking subsidies, grants and state support before payment

7 min read3 sources

Grant fraud detection pays off most before the payment: fraud with subsidies, grants and other state support is visible in the application itself. Signs include a document number that belongs to someone else, a bank account shared by different applicants and a company with no activity at all. AI matches each application with registries and with other applications and sets a risk flag, while the decision on payment stays with an official. Below is the experience of the US, the UK and the EU, with figures and dates.

On this page · 6 sections

In short

Grant fraud detection is cheaper before payment than recovering money after it. AI matches applicant data with registries and with other applications and flags the risk in advance. In the US, checking applications against data before payment could have prevented about USD 79 billion of possible fraudulent pandemic aid payments.

Why checking after payment is expensive

Support programs usually launch fast, and checks are postponed so that the money is not delayed. The price of that choice shows later: the recipient must be found, the violation proved and the funds returned, and by then the company may be dissolved.

After payment, an auditor works with the consequences:

1. Recovery needs demands, lawsuits and enforcement proceedings.

2. Some recipients no longer exist or have no assets.

3. An investigation takes years, and all that time the money does not serve the program’s goal.

Before payment, the same check is cheap. The application is not yet approved, and an official can ask for documents without starting a recovery.

There is also a cost for honest recipients. When the check comes after payment, all recipients of the program fall under suspicion at once, not only the risky ones, and the proceedings drag on for years.

The US experience shows the difference well. The PRAC committee estimates that checking applications against data before payment could have prevented about 79 billion dollars of possible fraudulent pandemic aid payments. Now this money can only be recovered through investigations.

USA: USD 79 billion without a check before payment

In June 2025, the US Pandemic Response Accountability Committee (PRAC) issued a fraud alert. Its main conclusion: checking applications against data before payment could have prevented about 79 billion dollars of possible fraudulent payments.

How the estimate was made. The committee randomly selected 662,000 records from 67.5 million funded applications in three programs: COVID-19 Economic Injury Disaster Loans (EIDL), the Paycheck Protection Program (PPP) and pandemic unemployment benefits. The Social Security Administration checked the Social Security numbers, names and dates of birth.

The check flagged almost 24,000 records in the sample as possibly fraudulent. A statistical model extended the result to the whole population and gave an estimate of about 1.4 million questionable Social Security numbers.

The committee’s conclusion is not about how many fraudsters there were but about where the check stood. Identity matching could have been done before payment by comparing agency data with what the applicant stated: IP addresses, dates of birth, bank accounts and home addresses.

Each of these records is only a risk sign so far. Investigators and courts decide case by case whether fraud took place.

UK: GBP 480 million in a year

The UK government reported GBP 480 million recovered from fraudsters from April 2024 to April 2025.

Along with this figure, the government presented an AI tool, the Fraud Risk Assessment Accelerator. It checks not applicants but the drafts of new programs: it looks for weak spots in the rules before launch. The tool is to be shared with other countries.

For subsidies, this is an important shift. Some grant fraud becomes possible because of the program’s own rules, and checking the rules before launch closes such gaps before the first application arrives.

What to check in a draft program before launch:

1. Which applicant data is matched with registries, and at what point.

2. How the program learns about a repeat application from the same applicant in this or another program.

3. Who confirms, and with which documents, that the money went to its purpose.

More on British control systems is on the page about the UK experience.

EU: EPPO cases worth EUR 24.8 billion

The European Public Prosecutor’s Office (EPPO) investigates fraud against the EU budget. According to its annual report for 2024, at the end of the year it had 2,666 active investigations with damage to the EU budget of about EUR 24.8 billion.

VAT cases make up 18% of investigations and 53% of the damage. The other cases cover other schemes with EU money. VAT schemes are reviewed on a separate page on tax schemes.

For comparison: assets worth EUR 849 million were frozen. Even a successful investigation returns money neither quickly nor in full, so checking an application before payment costs less than any investigation.

The European experience shows something else too. For an investigation to start at all, it needs a signal: a complaint, an audit or a flag from an analytics system. The earlier the signal comes, the better the chance that the money has not yet been moved out.

How the European Commission checks recipients of EU funds is covered on the page about the European Union.

What AI checks in each application

An application check has three layers: identity and the right to support, links with other applications, and whether the application matches real activity.

1. Identity: the document number and tax number are valid and belong to the applicant, the person is alive and the company is active.

2. Right to support: the applicant meets the program’s terms by region, industry, size and time in business.

3. Repeats: the same applicant, account or address has already appeared in this or another program.

4. Links: different applicants share a bank account, phone, email, IP address or director.

5. Real activity: the company has employees, revenue, tax payments and a history before the program.

6. Use for the stated purpose after payment: the money went to the declared goal, not to the accounts of related parties.

The PRAC committee names a similar set of data for matching: applicants’ IP addresses, dates of birth, bank accounts and home addresses. The signs add up to a risk score, and an official sees the applications with the highest score first.

A made-up example: Company A and Company B applied for a grant in different regions but gave the same bank account and the same IP address. The system gives both applications a high score, and the official asks for statements and contracts and finds out how the companies are linked. If the link is lawful, for example both use the same accounting firm, the flag is removed and the applications move on.

Applications in the names of dead and non-existent people are covered on the page about ghost recipients.

How not to delay honest applicants

If you check every application by hand, the program loses its point: help arrives when it is no longer needed. So the machine passes applications without risk signs, and manual review goes only to the flagged ones.

Rules that reduce delays:

1. The matching runs automatically while the application waits in the queue.

2. The flag goes to a specific application with a sign, not to a whole category of applicants. Citizenship and place of residence are not part of the model.

3. The applicant sees which document is needed and can send it without filing again.

4. Flagged applications have a review deadline, after which a senior official decides.

5. The outcome of each manual review goes back into the system, and weak signs that are often cleared lose weight.

A flag means risk, not proof. Refusing a payment, demanding the money back or passing materials to an investigator remains the decision of an official, an oversight body or a court.

The British approach of checking the program rules before launch also helps honest applicants: the more precise the rules, the fewer applications go to manual review. The goal of a check before payment is for an honest applicant to get the money fast, and for a questionable application to get questions before the money.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. Pandemic Response Accountability Committee, 4 June 2025 (pandemicoversight.gov)
  2. TechRadar, 25 September 2025 (techradar.com)
  3. EPPO, Annual Report 2024, 3 March 2025 (eppo.europa.eu)

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