JPMorganChase says one AI model now handles fraud detection, payment authorisation predictions and payment retries, and is already scoring millions of transactions a day. In the bank’s reported tests, it improved detection of card fraud by up to 85.5% against task-specific models, though the comparison is with JPMorgan’s own systems, not a public benchmark.
Watch Desk analysis
What happened
The Large Payments Model learns from transaction records rather than ordinary text. JPMorgan researchers say its training data covers card payments where the bank is the merchant’s acquirer, plus ACH and wire transfers. The largest version was trained on up to 400 billion tokens.
The reported results vary by task. Using up to six months of transaction history raised the card-fraud detection improvement to 85.5%; a version scoring each transaction on its own improved the measure by 62.7%. The model also surfaced 27.4% more fraudulent ACH payments at a fixed alert-review level. For wire transfers, it caught 5.6% more fraudulent payments but surfaced no additional fraud by dollar value.
The bank’s researchers say the production model serves millions of transactions daily, completes 99% of requests within 60 milliseconds and costs less than $0.0001 per score. PYMNTS reports these details from the researchers’ paper and its account of the model.
Why it matters
The practical bet is that one model can learn patterns shared across payment jobs instead of teaching separate systems the same transaction behaviour repeatedly. JPMorgan’s reported serving figures suggest this is a deployed system, not just a lab result, while the differences between card, ACH and wire performance show why “one model” does not mean one uniform outcome.
The gains are relative to the bank’s own task-specific models and its own payment data. The paper, as described by PYMNTS, does not give absolute model scores, compare performance with a public benchmark, or report fraud losses, false declines or savings. Those are important missing measures when assessing what the improvements mean for customers and the bank.
Our read
This is a substantial example of AI being built around the structure of financial transactions rather than adapted from a text model. The most interesting result is not the headline card figure on its own, but the uneven performance across payment types and the use of transaction history. That makes the research useful, while leaving plenty for independent evaluation to settle. Impressive percentages are welcome; knowing what they translate into is better.
What to watch
- Whether JPMorgan publishes absolute scores, false-decline rates or evidence of changes in fraud losses.
- How the model performs on wire transfers as its training data and deployment develop.
- Whether other banks report comparable results from transaction-trained models.
Discussion spark: For a bank deploying AI to flag payments, should the headline measure be improved fraud detection, fewer false declines, or the balance between the two?
Sources and evidence
- Source update (17 November 2026, 20:54 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.