The scale of losses: 3–7% of spending
The US Government Accountability Office (GAO) estimated that the federal government loses between $233 and $521 billion a year to fraud. That is 3–7% of average federal obligations, based on data for fiscal years 2018–2022.
According to the same report, other organizations give a similar order of magnitude: the UK Public Sector Fraud Authority 0.5–5%, ACFE 5%, the Portsmouth Centre 6.4%. Their methods differ, so this is a guide to scale, not a norm.
The estimate applies to the federal government as a whole. GAO warns directly that these percentages cannot be applied to a single agency or program, because the level of risk varies a lot with controls, budget growth or cuts, and new schemes.
Hence the first rule of the calculation: do not multiply your budget by someone else’s percentage. Your own data gives your baseline: how many purchases and payments went through, how many of them already raised questions with inspectors, and what the last confirmed violations cost.
GAO’s second caveat: a fraud estimate cannot be compared with improper payment estimates, because they are counted by different methods. In an effect report, these sums are also kept on separate lines.
USA: six times more in a year
In fiscal year 2024, the US Treasury prevented and recovered more than $4 billion of fraudulent and improper payments. A year earlier the sum was $652.7 million.
Two tools drove the growth. Machine learning on check fraud brought $1 billion of recoveries, and prioritizing high-risk transactions stopped $2.5 billion.
The difference between them matters for the calculation. A check recovery comes after payment and needs work with banks, while a stopped transaction never leaves the budget at all.
Prioritizing means that inspectors look first at payments with a high risk score. The rest go through without delay, and an honest recipient does not wait longer than usual for the money.
The lesson of this experience: most of the money comes not from searching after payment but from sorting payments by risk before it. More about the US systems is on a separate page.
Brazil: stopped tenders
The ALICE robot of Brazil’s audit body CGU has worked since 2015. Every day it collects more than 500 procurement notices, checks them against about 40 types of risk and sends signals to more than 500 auditors.
In 2019–2022, tenders worth more than 9.7 billion reais were suspended or canceled in Brazil. This money did not have to be recovered: the tenders were stopped before contracts were signed.
In 2023, the robot analyzed 190,923 purchases. Such a flow cannot be checked in full by hand, and auditors would look only at a sample.
The robot does not cancel tenders by itself: it sends signals to auditors, and people decide whether to suspend or cancel. So the sum of stopped tenders reflects the work of the whole chain, not of one program.
How ALICE works and how auditors handle its signals is covered on the page about Brazil’s experience.
Prevented versus recovered
The United Kingdom counts recovered money: the government reported 480 million pounds recovered from fraudsters from April 2024 to April 2025. Recoveries are easy to count, but they need an investigation, a court and time.
Prevented spending is harder to count, but it needs no collection. Kazakhstan’s anti-corruption service, by monitoring inflated prices in the Single Procurement Window, prevented wasteful spending of more than 67 billion tenge in two years: 29 billion in 2022 and 38.4 billion in 10 months of 2023.
The UK Fraud Risk Assessment Accelerator works even earlier: it checks draft new programs for weak points before they launch. Such an effect is almost invisible in reports, because the fraud simply did not happen.
The US federal Chief Financial Officers Council, which GAO cites, admits that the result of prevention is hard to measure but can give a large return. So a report needs both lines, and the prevented sum is counted by a rule described in advance.
The rule for the prevented sum is written down before the pilot. For example, count the price of a purchase canceled after a check, or the difference between the starting price and the price after review.
How to count the effect for your budget
The effect is counted from a pilot on one scheme, not from other people’s percentages. The calculation has four lines, and each can be checked against documents.
1. Prevented: the sum of purchases and payments stopped or changed before payment after a signal and a human check.
2. Recovered: sums collected or returned in cases that started with a system signal.
3. Time: how many inspector hours went into one purchase before and after, multiplied by the number of checks a year.
4. Costs: development, support and staff time spent on reviewing signals, including false ones.
5. Net effect: the sum of the first three lines minus the fourth. If it is negative after the pilot, that is a reason to change the scheme, not to add notional millions to the report.
Only sums on which a person made the decision count toward the effect: an inspector, a manager or a court. A signal nobody confirmed is worth nothing, and a report that counts it will make people distrust the system.
The data needed and the pilot plan are described on the page about how to deploy AI control.
What counts as a result
The result of the first year is often not billions recovered but the share of the flow that gets checked. ALICE checks every procurement notice, and auditors see the whole flow, not a random part.
A set of metrics for the report: the share of purchases and payments checked, the number of signals and the share confirmed, the sum stopped before payment, and the time per check.
Trends by quarter are also useful to a manager: whether the share of signals closed with documents grows, and whether the time between publishing a purchase and checking it gets shorter.
The US Treasury’s experience shows that growth can be fast when signals are sorted by risk: it was the prioritizing of transactions that stopped $2.5 billion.
A zero result is a result too. If a pilot on one scheme gave no confirmed signals, the report says so, and there is no need to scale that scheme.
What a report should never contain is a count of “corrupt officials found”. The system flags signs of risk, and an inspector or a court finds a violation.