Financial institutions are abandoning fragmented AI architectures in favor of unified transaction foundation models, and early results suggest the shift is paying off. 

Chip designer Nvidia launched its AI-driven Build Your Own Transaction Foundation Model tool today, Pahal Patangia, head of global industry business development and payments at Nvidia, told FinAi News. The tool allows FIs to train AI models on billions of proprietary transactions rather than relying on purpose-built models for each use case such as geographic transaction detection and unique transaction attempts. 

“Transaction foundation models serve as the semantic layer that agentic systems need to act intelligently, not just react,” Patangia said. “Instead of flagging individual signals, they understand full transactional context — which lets agentic systems make decisions grounded in behavioral history rather than rules.” 

From silos to single horizontal layer 

Nvidia’s new AI tool works across the fraud detection system horizontally, rather than functioning in vertical silos, he said, adding that the AI tool connects multiple data sets within the FI to evaluate threat at a granular level. 

“Traditional models operate in silos — a fraud model sees fraud signals, a credit model sees credit signals, and never the two shall meet,” Patangia said. “Transaction foundation models unify previously siloed data including payments, transfers, product interactions, behavioral signals, authorization history, fraud patterns, chargebacks, merchant location, and loyalty data.” 

Contextual signals like timing, device, location and prior activity are interpreted together, he said, adding that a midnight payment in an unfamiliar city reads very differently depending on the customer’s full behavioral history. 

Live deployments 

Digital bank Revolut, Mastercard, Stripe and Adyen are early adopters and are posting high efficiency gains from Build Your Own Transaction Foundation Model, according to Nvidia’s release today. 

Revolut, for example, built AI-driven tool PRAGMA with Nvidia and trained the model on 24 billion proprietary data points from 26 million users across 100 countries, Patangia said. The tool helped the London-based digital bank increase its fraud recall by 64.7%. 

“Recall in fraud detection is the percentage of actual fraud cases your model catches,” he said. “If a bank previously caught 40% of fraud and now catches 65%, the incremental 25 percentage points translate directly into recovered losses.” 

Industry benchmarks suggest raising fraud recall by even 10 points can prevent more than $2 million a year for every $1 billion in transaction volume, he said, adding that Revolut eliminated weeks, possibly months of feature engineering time for new use case deployed with the tool. 

Stripe has also seen success with the tool, with the payment company blocking close to $112 billion in fraud last year, achieving a 38% average reduction in fraud rates, Patangia said. 

Register here for the FinAi Lending Summit, set for Oct. 7-8 in Las Vegas.