Revolut Stopped Paying for AI and Started Winning 

The London-based neobank trained a proprietary model called PRAGMA on 40 billion transactions, app interactions and financial events drawn from 25 million users across 111 countries. It doesn’t answer questions or write emails. It makes decisions: who’s likely to commit fraud, who’s a credit risk, which customers are about to leave. According to Revolut, PRAGMA outperforms every task-specific system the company previously ran.

Mastercard announced in March that it’s building a generative AI model on hundreds of billions of anonymized payment transactions, with applications spanning fraud, loyalty and portfolio analytics. Plaid introduced a transaction model built to standardize how merchant data is read across institutions. As PYMNTS has reported, the companies that process the world’s payments are now training AI on the structured transaction data they’ve spent decades collecting.

PRAGMA sits apart from that group in two ways. Mastercard and Visa operate as networks: they see transactions, but not the full arc of a customer’s financial life. Plaid sees account data, but its model targets a specific problem, merchant name standardization, rather than a broad decisioning layer.

Revolut, as a neobank, holds the complete picture. It sees where customers spend, what they trade, how they move money internationally and how they navigate the app before and after each transaction. PRAGMA is trained on all of it, fused into a single representation of each user.

Why General-Purpose AI Falls Short in Finance

AI models like ChatGPT learn by reading text. Feed them enough of the internet and they get good at predicting what word comes next. That makes them useful for drafting, summarizing and answering questions in plain language. It doesn’t make them useful for fraud detection.

The reason is that financial data isn’t text. A bank’s record of a customer’s behavior is a sequence of structured events: a $14.99 subscription charge on a Tuesday, a $150 transfer on Christmas Day, a salary deposit the following Friday. Each event has a type, an amount, a currency, a timestamp and a relationship to every other event in the customer’s history. When you feed that into a text-based AI, it reads the dollar signs and digits the same way it reads letters. A $1,000 transfer and a $10 transfer look nearly identical. The patterns that signal fraud or credit risk get lost.

Speed is the other problem. A fraud check has to complete in milliseconds, before a transaction clears. General-purpose AI models are built for depth, not for that kind of real-time decisioning at scale.

One Model That Does What Dozens Used To

PRAGMA is built differently. It reads each financial event as three things: what happened, the specific details and when. It learns patterns across a customer’s entire history. A customer who topped up their account, made a small card payment, then sent an international transfer, then went quiet represents a pattern. PRAGMA learns to recognize what that pattern typically precedes.

According to Revolut, the model was trained on 207 billion data points spanning 25 months of real customer activity. It comes in three sizes, from a lean 10 million parameter version built for real-time fraud checks to a 1 billion parameter version for decisions where accuracy matters more than speed. All three run on more than 200 NVIDIA H100 GPUs, as detailed by Nebius, Revolut’s cloud infrastructure partner.

The operational payoff is consolidation. Most financial institutions run a separate AI model for each problem: one for fraud, one for credit scoring, one for customer lifetime value, one for product recommendations. Each requires its own training data, its own maintenance, and its own engineering team. PRAGMA replaces that entire stack with a single shared foundation. Revolut adapts it to a new task by adjusting a small fraction of the model’s parameters rather than building from scratch, cutting retraining cycles and accelerating deployment across the business.