Which payments count as improper
An improper payment went to the wrong recipient, in the wrong amount or without grounds. A fraudulent payment is a special case of an improper one: there is intent behind it. For control, the difference matters later, at the investigation stage, while the system flags both kinds the same way, by signs in the data.
The main types:
1. Duplicate payment: the same invoice is paid twice, sometimes with small differences in the number or the date.
2. Payment without grounds: there is no contract, delivery certificate or payment decision.
3. Payment to the wrong recipient: the bank details do not match the counterparty’s details in the contract.
4. Payment to a person with no right to it: the recipient has died, does not meet the payment terms or exists only on paper. Such cases are covered on the page about ghost beneficiaries.
An estimate by the US Government Accountability Office (GAO) shows the scale of the problem: the federal government loses an estimated USD 233–521 billion a year to fraud, or 3–7% of average federal obligations based on fiscal years 2018–2022. GAO itself warns that these percentages cannot be applied to individual agencies or programs.
US Treasury experience
The US Treasury makes about 1.4 billion payments a year worth more than USD 6.9 trillion to more than 100 million people. In October 2024 it reported that in fiscal year 2024 it prevented and recovered more than USD 4 billion of fraudulent and improper payments. A year earlier the figure was USD 652.7 million.
What the total consisted of:
1. Identifying and prioritizing high-risk transactions stopped USD 2.5 billion.
2. Machine learning sped up the search for Treasury check fraud and brought USD 1 billion in recoveries.
3. Expanded risk-based screening prevented USD 500 million.
4. Changes to the payment processing schedule prevented USD 180 million.
The key point in these figures is not the model but where the check happens. Most of the total is prevention: the money never left, and nobody had to recover it.
Machine learning here makes no payment decisions. It sorts billions of transactions so that staff look at the risky ones first. More on the federal government’s control systems is on the page about US experience.
Check before payment, not after
Getting money back after payment means finding the recipient, claims and court, and part of the money never returns. A check before payment takes seconds of machine time per payment.
The opposite example is also from the US. The Pandemic Response Accountability Committee (PRAC) estimated that checking applications against data before payment could have prevented about USD 79 billion of potentially fraudulent payments. In a sample of 662,000 records out of 67.5 million applications, the committee counted about 1.4 million questionable Social Security numbers.
What a check before payment looks like in a treasury or accounting team:
1. A payment order enters the queue, and the system matches it with the invoice, the contract, the delivery certificate and the payment history.
2. A payment with no flags goes out in the usual way and on the usual date.
3. A flagged payment waits for a decision by an employee, who sees the reason: a repeat invoice, new bank details, a recipient listed in death records.
For the check to work, it needs a data stream at the moment a payment is created, not a quarterly export. How this works is covered on the page on real-time monitoring.
Repeat invoices and changed bank details
Duplicate payments rarely look like an exact copy. More often a repeat differs in one detail, and a simple check for a matching invoice number misses it.
Signs of a repeat invoice that the system flags:
1. The supplier, the amount and the subject match, but the invoice number or date differs by one character or a few days.
2. The same delivery certificate or waybill is attached to two payments.
3. The total of payments under a contract exceeds the contract price.
Signs of changed bank details:
1. The recipient’s bank account changed shortly before a large payment.
2. The new account already appears in the records of other suppliers or individuals.
3. The letter about the change of bank details came from a new email address or without the director’s signature.
For each sign, the employee has a simple next step: match the accounts in the accounting system, call the supplier at the phone number in the contract, not in the letter, and ask the bank for confirmation.
A made-up example: Supplier B has two accounts in its record, and the second was added a day before payment of a large delivery certificate. The system holds the payment, the accountant calls the number in the contract and learns that the supplier did not change its details. The money stays in the budget, and the materials go to the security team.
If the payment for a certificate with no delivery behind it went to the right recipient, the problem is not the payment but the contract. Such cases are covered on the page about fake contracts.
Priority by risk and amount
If you flag everything suspicious, the review queue grows faster than people can clear it, and payments stall. So the system does not just set a flag: it computes a risk score and sorts the payments.
What the priority consists of:
1. Strength of the sign: an exact repeat invoice weighs more than a new supplier.
2. Number of signs: changed bank details together with a large amount and a short deadline matter more than any of them alone.
3. Amount: at equal risk, the payment with more money at stake is checked first.
4. History: if similar flags were confirmed before, the weight of the sign grows; if they were cleared, it falls.
The US Treasury describes the same principle. Identifying and prioritizing high-risk transactions gave it USD 2.5 billion in prevented losses in fiscal year 2024, more than any other area.
The analyst sees a short list at the top of the queue, not thousands of payments. A flag means risk, not proof: a payment with a high score may turn out to be lawful, and a person decides whether to hold it, recover it or pass the materials to an investigator. How such a score works is covered on the page on risk-based control.
What the treasury and accounting teams need to provide
You need no new accounting system to start. You need access to data you already have and people who will review the flags.
1. A register of past payments: the system is tuned on it and shows how many repeats and anomalies there were before.
2. The stream of new payment orders before they go to the bank.
3. A counterparty directory with the history of changes to bank details.
4. Contracts, invoices and delivery certificates in electronic form or as scans the system can read.
5. External registers: legal entities, death records, sanctions and restriction lists.
6. A responsible employee who reviews the flags and records the outcome: confirmed or cleared.
It makes sense to start with one check, for example a search for repeat invoices in last year’s archive. The result will show how much money passed by control and will give grounds for the next step.
The US Treasury went the same way: it first applied machine learning to one type of fraud, Treasury check fraud, and got USD 1 billion in recoveries in a year.