Freddy Lecue describes Wells Fargo’s transaction foundation model in this 18-minute AI Engineer talk.

Train on the data the bank actually has

  • Pretrain a transformer from scratch on customer transactions, then adapt the shared backbone to individual tasks.
  • Predicting the next transaction teaches representations that can support fraud classification and other downstream work.
  • Fraud is the demonstrated use case; credit-line changes, anti-money laundering and marketing are further investigations.

Tokenization carries much of the engineering

  • Start from roughly 100 raw features spanning who, what, how, when and where.
  • Encode amounts, time gaps, merchant categories and locations into a manageable transaction vocabulary.
  • Preserve transaction histories so the model can relate a new event to earlier behavior.

Judge the model on decisions and operating limits

  • Lecue reports more fraud detected and fewer false alarms than the production XGBoost comparison.
  • False declines cost transactions and create customer-support work, so both metrics matter.
  • The model must fit the real-time card-decision window; the talk cites a processing requirement below 250 milliseconds.
  • Architecture and pretraining scores matter only insofar as they improve downstream decisions.

Caveat: these are Wells Fargo’s reported results. The capture does not establish exact improvement percentages or independently verified savings.

“The most complex part was really the tokenization.” — Freddy Lecue, 12:16