BIGGER PICTURE Public & everyday services
Treasury says AI helped recover $1 billion lost to cheque fraud
A payment leaves a record when it is issued and another when the banking system processes it. Treasury used machine learning to help analysts spot suspicious differences at enormous scale.

- Reported cheque-fraud recovery
- US$1 billion
- Reporting period
- Fiscal year 2024
- Fraud decision
- Human analysts
How the idea works
Match issued and returned records
Check information is reconciled across payment stages.
Flag suspicious differences
Machine learning helps prioritize items for review.
Investigate and recover
Analysts decide and pursue the recovery process.
Two versions of the same payment
The Bureau of the Fiscal Service handles a very large share of US federal payments. Cheque fraud creates a particular reconciliation problem: does the information returned through the banking system match the cheque the government originally issued? Comparing those records can expose suspicious changes without requiring a general-purpose model to guess someone’s intentions.
Sources: [2]
Turn mismatches into an analyst’s queue
Treasury’s April 2024 explanation to Congress says the machine-learning tools supplement existing cheque reconciliation and post-payment analysis. They flag suspicious items for human review; they do not decide that fraud occurred. The documented inputs concern the payment itself. This makes the intervention a focused detection aid within an established recovery process.
Sources: [2]
A billion dollars, with a precise boundary
For fiscal year 2024, Treasury attributed US$1 billion in cheque-fraud recoveries to faster identification using machine learning. Its wider announcement reported more than US$4 billion prevented or recovered through several different initiatives. Those larger totals must not all be assigned to AI. The directly supported claim is the cheque-fraud recovery component reported by the agency.
Sources: [1]
Useful detection must lead to useful action
The public material does not supply a randomized comparison or a detailed false-positive rate. Nevertheless, it describes an operational outcome beyond an accuracy benchmark: money recovered through a workflow involving analysts and financial institutions. The transferable lesson is to connect a model’s signal to a specific, reviewable action whose result can be accounted for.
The original sources
- US Department of the TreasuryEnhanced fraud detection processes in fiscal year 2024
- US Department of the Treasury / US CongressTreasury response to the House Judiciary Committee on AI use