The $40 Billion Fraud Window

The Federal Reserve Bank of Atlanta's 2024 fraud report put global payment fraud losses on track to exceed $40 billion annually by 2027. That number is not a failure of fraud detection algorithms. It is a failure of architecture. Specifically, it is the failure of batch processing, the decades-old operational pattern that gives criminals hours to move stolen funds while banks wait for the next scheduled detection run.

The fraud window is not a technology gap. It is a time gap. Batch processing is what keeps it open.

The companion piece in this issue argues that fraud operations is the highest-ROI first agentic deployment in retail banking, and walks through the math. This piece is about the layer underneath the agent. Without a streaming data layer, the agent reasons fast and acts late. With one, the agent reasons and acts inside the window where fraud loss is still recoverable.

The batch detection problem, restated

Here is how fraud detection still works at most institutions in 2026. Transactions are collected throughout the day. At predetermined intervals (typically every few hours, sometimes overnight) they are loaded into a data warehouse and run through detection models. Suspicious transactions are flagged. Alerts are generated. An analyst reviews them the next morning.

The problem is obvious once you state it plainly. By the time a fraudulent transaction is flagged, the money is gone.

Modern fraud operates in minutes, not days. Authorized push payment fraud, where victims are socially engineered into initiating transfers, accounts for over $4 billion annually in the UK alone, per UK Finance's 2024 Annual Fraud Report. The average authorized push payment fraud takes less than fifteen minutes from initiation to fund movement. A batch detection system running every four hours flags the transaction approximately three hours and forty-five minutes too late.