JPMorgan Builds AI Model Trained on Payment Records

The Large Payments Model detects more card, ACH and wire fraud than single-task systems, according to the company's figures.

Updated on Nov. 17, 2026 in Artificial Intelligence

JPMorgan Builds AI Model Trained on Payment Records

Live Poll

Do you trust your bank's AI systems to protect your money from fraud?

JPMorgan has built the Large Payments Model, an AI system that learns from payment records rather than text. In the reported results, reading six months of transaction history improved card fraud detection by 85.5% over a model built only for that task.

Why it matters

Fraud losses and payment-processing costs fall on banks, merchants and cardholders alike, so better detection could reduce losses across the U.S. payments system. The model is already in production, serving millions of transactions per day.

The model scores each transaction in under 60 milliseconds for 99% of requests at a cost below $0.0001, and its largest version was trained on up to 400 billion tokens. Independent verification of these figures and their performance on other payment networks has not been reported.

The players

JPMorgan

JPMorgan is a major U.S. bank that operates large payments and merchant acquiring businesses and has invested heavily in artificial intelligence.

The details

Rather than processing text, the model converts each transaction into a fixed set of fields such as dates, postal codes and descriptions. Its training data covers card payments where JPMorgan is the merchant's acquirer, along with ACH and wire transfers.

Timeline

  1. Six months of prior transaction history is the window the stronger model version reads.

  2. A PYMNTS Intelligence survey of 90 bank and FinTech executives was conducted in July.

The Tech Race

The build-in-house approach reflected in the PYMNTS survey, where 71% of financial institutions plan their own fraud systems, is the path JPMorgan has taken with its payments model. The move positions the bank against vendors selling off-the-shelf fraud tools and against rival banks developing similar internal systems.

Cardholders may see fewer fraudulent charges go through if the model's detection gains hold up in practice. Merchants and consumers could also benefit from faster payment decisions, since 99% of requests are completed within 60 milliseconds.

The takeaway

Banks that build fraud models on their own transaction data can catch more fraud than generic tools, but the reported gains are the company's own figures and should be checked as independent results emerge.

Further reading

For more on how banks are deploying machine learning, see Artificial Intelligence coverage on Wisevoter.

Source note: This article includes information reported by PYMNTS.

Live Poll

Do you trust your bank's AI systems to protect your money from fraud?