JPMorgan Builds One AI Model to Catch 85% More Fraud

Highlights

JPMorgan built a single AI model for fraud detection, payment authorization predictions and retries, trained on up to 400 billion tokens of card, ACH and wire transaction data.

The model improved card fraud detection by up to 85.5%, ACH fraud detection by 27.4% and payment retry predictions by 22.8% compared with the bank’s task-specific models.

Seventy-one percent of financial institutions plan to build new fraud systems themselves, up from 44% a year earlier, PYMNTS Intelligence found.

JPMorganChase has built one artificial intelligence model to handle fraud detection, authorization prediction and payment retries, three jobs that draw on much of the same transaction data. Building a separate model for each one means teaching every model the same payment patterns again, the bank’s researchers wrote in a paper published Nov. 11.

    Get the Full Story

    Complete the form to unlock this article and enjoy unlimited free access to all PYMNTS content — no additional logins required.

    Subscribe to our daily newsletter, PYMNTS Today.

    By completing this form, you agree to receive marketing communications from PYMNTS and to the sharing of your information with our sponsor, if applicable, in accordance with our Privacy Policy and Terms and Conditions.

    The Large Payments Model, or LPM, learns from payment records instead of text. Each transaction enters as a fixed set of fields that hold numbers, categories or short text, such as dates, postal codes and descriptions. The training data covers card payments where JPMorgan is the merchant’s acquirer, plus ACH and wire transfers. The largest version was trained on up to 400 billion tokens, the units of data a model reads.

    The model is already in production. It is “serving millions of transactions per day” and completes 99% of requests within 60 milliseconds, the researchers wrote. Each transaction costs less than $0.0001 to score.

    Transaction History Pushes Card Fraud Gains to 85%

    JPMorgan tested two versions of LPM on card payments, the researchers wrote. The first scores each transaction on its own and improved fraud detection by 62.7% over a model built only for that task, the paper’s results table showed. The measure, average precision, tracks how well a model ranks rare fraud cases above legitimate payments, the researchers said.

    The second version reads up to six months of prior transactions tied to the same party. That history pushed the fraud detection gain to 85.5%.

    The same model also predicts what happens to a payment after it is sent. It was 80% better at predicting which of about 100 responses a card network would return and 22.8% better at predicting how a card network would respond to a retried payment.

    LPM also handled a job it was not trained for. Asked to predict whether a retried payment would eventually be approved, the model needed only a small add-on to score 0.5% above a model built for that job, the researchers wrote.

    Every result is a gain over JPMorgan’s own single-task models, measured on the bank’s own card and treasury payments, the researchers said. The paper does not name those models, give their absolute scores or test LPM against a public benchmark. It includes no figures on fraud losses, false declines or cost savings.

    ACH Gains Reach 27% While Wires Lag

    Treasury payments, meaning ACH and wire transfers, were tested differently. Fraud teams can review only so many alerts a day, so the researchers held that number fixed and measured how much more fraud the model surfaced. LPM caught 27.4% more fraudulent ACH payments and 63.5% more fraud by dollar value, the paper showed.

    Wire transfers did not see the same lift. The model caught 5.6% more fraudulent wires and no additional dollar value, the paper found, a gap the researchers tied to wires making up a smaller share of the training data.

    The design has limits, the researchers said. Adding new data fields may require more training, and neither version removes the need to summarize long payment histories.

    Stripe, Mastercard and Plaid Also Train AI on Transactions

    JPMorgan is the latest payments heavyweight to train a model on its own transaction data. Stripe introduced a payments model in May 2025 that lifted its detection rate for card-testing attacks on large businesses to 97% from 59%, PYMNTS reported.

    Mastercard said in March that it is building a model on billions of anonymized transactions and plans to scale to hundreds of billions. Plaid followed two weeks later with a transaction model for merchant recognition, categorization and risk signals, PYMNTS reported.

    Banks are moving the same work in-house, too. Seventy-one percent of financial institutions plan to develop new fraud systems themselves, up from 44% a year earlier, the PYMNTS Intelligence report “Scams Get Personal: Why Banks Pay More to Fight Fraud” found in a July survey of 90 executives at U.S. banks, credit unions and FinTechs.

    Those institutions spend $2.27 on prevention for every $1 lost to fraud, up from $0.94 in 2024. Seventy-eight percent named the complexity of new technology systems as their top barrier to innovation, up from 48% a year earlier, according to the same report.

    For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.