JPMorganChase has built a single AI model, the Large Payments Model (LPM), to handle fraud detection, authorization prediction and payment retries, PYMNTS reported, citing a paper the bank's researchers published November 11. The model learns from structured payment records rather than text, and its 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 researchers wrote that LPM is in production, "serving millions of transactions per day," completing 99% of requests within 60 milliseconds at under $0.0001 per transaction. A version that reads up to six months of prior transactions improved card fraud detection by 85.5% in average precision over a task-specific model; a version without history improved it by 62.7%. The model was also 80% better at predicting which of about 100 card network responses would come back and 22.8% better at predicting responses to retried payments.
For ACH, with alert volume held fixed, LPM caught 27.4% more fraudulent payments and 63.5% more fraud by dollar value. Wires lagged at 5.6% more fraud caught and no added dollar value, which the researchers tied to wires being a smaller share of training data. PYMNTS noted that all gains are measured against JPMorgan's own unnamed models, with no absolute scores, public benchmark, or figures on fraud losses, false declines or cost savings.
PYMNTS placed the work alongside a Stripe payments model introduced in May 2025 and transaction models from Mastercard and Plaid. A PYMNTS Intelligence survey of 90 US financial executives found 71% plan to build new fraud systems in-house, up from 44% a year earlier.
