Financial institutions have spent years building Artificial Intelligence across fraud, credit, recommendations and risk, but siloed systems have limited how much those models can learn from broader customer behavior. As enterprise datasets grow, the gap is widening between what institutions know and what their Artificial Intelligence can reason over. That is pushing firms toward transaction foundation models, which use transformer-based architectures to learn a unified representation of consumer behavior from proprietary transaction data.
NVIDIA’s 2026 State of Artificial Intelligence in Financial Services report shows 65% of institutions now use Artificial Intelligence, with nearly 90% deploying or assessing it and almost all maintaining or increasing spend. Transaction foundation models are described as large-scale Artificial Intelligence systems trained on billions of financial events such as payments, transfers, product interactions and behavioral signals. Instead of evaluating isolated inputs, these models interpret behavior in context, where timing, device, location and prior activity shape meaning. That contextual approach is intended to improve performance across multiple tasks while reducing reliance on handcrafted features and separate model stacks.
In collaboration with NVIDIA, Revolut built PRAGMA, a family of transformer-based foundation models trained on 24 billion events across 26 million user records spanning over 100 countries. Running on Nebius cloud with NVIDIA infrastructure and software, the company says a single foundation model outperforms strong task-specific models across areas including credit scoring, fraud detection and product recommendations. Mastercard is also developing a proprietary tabular foundation model for payments, trained on billions of anonymized transactions today and designed to scale to hundreds of billions across additional datasets including fraud, authorization, chargeback, merchant location and loyalty data. Early testing shows it outperforming standard machine learning techniques across several financial applications.
The strategy is also extending into agentic commerce and payment operations. Forty-two percent of financial firms are already using or assessing agentic Artificial Intelligence. As these systems begin to execute transactions such as managing subscriptions, routing payments and making purchases, transaction data is being positioned as the semantic layer that can guide automated decision-making. However, several cited performance figures are incomplete in the source material and cannot be stated reliably. Even so, the broader message is clear: firms are treating proprietary transaction histories as a durable competitive asset for building domain-specific intelligence.
Deployment is increasingly tied to ecosystem support. NVIDIA’s Build Your Own Transaction Foundation Model developer example is available on Amazon Web Services with Amazon SageMaker HyperPod and on Nebius AI Cloud. Service partners including EXL, Infosys, GFT IT Consulting and Thoughtworks are adapting the approach for banking, payments, compliance, credit risk and governance. GFT IT Consulting says its Wynxx platform is used by over 100 financial institutions, while its Smaragd compliance engine reduces false positives by up to 75% for major banks. Together, these efforts reflect a broader move to replace fragmented model architectures with shared transaction intelligence layers.
