A bank-scale AI system now profiles tens of millions of customers by studying spending patterns instead of scanning each customer's full transaction history.
Researchers built a three-stage pipeline called Resolve, Profile, and Tag. Resolve turns cryptic item names into readable descriptions, using web lookups when needed. Profile runs one large-language-model call per frequent spending pattern, rather than per person, and outputs category labels, free-text attributes, and prevalence estimates. Tag then clusters those free-text attributes into a searchable database. On the public Open e-commerce dataset, the resulting profiles matched the accuracy of a model that reads each user's raw history directly.
The trick is arithmetic: transaction volume grows with users, but the number of distinct spending patterns doesn't grow nearly as fast. That is why the same accuracy came at close to a thousandth of the inference cost, letting a Japanese bank run this at the scale of tens of millions of accounts. It is a template for any company sitting on transaction data too large to profile one LLM call at a time.
The paper does not say what the bank does with these profiles: underwriting, marketing, or fraud screening. That gap is worth watching as 'efficient categorization of millions of people' moves from research paper to deployed product.