Financial Institutions Embrace Transaction Foundation Models for Unified AI Intelligence

Financial institutions are increasingly adopting transaction foundation models to consolidate their AI efforts and achieve a more unified understanding of customer financial behavior. This shift addresses the limitations of previously siloed, task-specific AI models, enabling a more comprehensive approach to intelligence gathering.

The current landscape sees a significant portion of financial institutions utilizing AI, with a strong trend towards deploying or assessing these technologies and maintaining or increasing investment. However, the complexity of scaling AI has brought fragmented model architectures to the forefront as a limiting factor. To counter this, leading firms are moving from specialized models to transformer-based transaction foundation models. These models are trained on vast amounts of proprietary financial events, transforming raw data into actionable intelligence.

A Unified View of Consumer Behavior

Unlike traditional fraud models that examine isolated signals, foundation models interpret financial events within their broader context. Factors such as timing, device, location, and prior activity are crucial in shaping the meaning of a transaction. This contextual depth, powered by transformer architectures applied to tabular data, allows for the extraction of signals previously undetectable by older algorithms.

Revolut, in collaboration with NVIDIA, has developed PRAGMA, a series of transformer-based foundation models. Trained on billions of events across millions of user records globally, PRAGMA leverages NVIDIA’s AI stack, including Hopper GPUs and the Nemotron open models, running on Nebius cloud. A single PRAGMA model has demonstrated superior performance to task-specific models in credit scoring, fraud detection, and product recommendations, significantly reducing the need for manual feature engineering.

Industry Leaders Adopt Foundation Models

Mastercard is developing its own proprietary large tabular foundation model for payments. This model, trained on billions of anonymized transactions and designed to scale, integrates data from various sources including fraud, authorization, and loyalty information. Built with support from NVIDIA, AWS, and Databricks, the model aims to reduce reliance on numerous individual AI models and has shown promising early results in outperforming standard machine learning techniques across applications like cybersecurity and personalization.

Adyen has already deployed transaction foundation models to process trillions of dollars in payments. By employing reinforcement learning, Adyen optimizes conversion rates and minimizes risk for merchants, noting that even minor improvements in authorization rates can lead to substantial financial gains.

The Rise of Agentic AI and Transactional Understanding

With 42% of financial firms already exploring or using agentic AI, systems capable of executing transactions are becoming more prevalent. This evolution necessitates AI that understands the full context of transactional behavior. Stripe, utilizing NVIDIA and AWS platforms, is building foundation models to achieve this, having blocked significant amounts of fraud and achieved substantial reductions in fraud rates by moving beyond reactive, signal-based detection.

The core advantage of transaction foundation models lies in their utilization of unique, proprietary transaction data that competitors cannot easily replicate. The necessary data already exists within institutions, the architecture is proven, and the infrastructure is readily available.

Implementation and Support

NVIDIA offers a “Build Your Own Transaction Foundation Model” developer example, enabling teams to begin building transformer embeddings on tabular transaction data without requiring a complete overhaul of existing systems. This example is accessible on Amazon Web Services (AWS), deployed via Amazon SageMaker HyperPod, and on Nebius AI Cloud.

Beyond self-service, financial services firms can engage with partners like EXL, Infosys, GFT IT Consulting, and Thoughtworks. EXL is integrating these models into its EXLerate.ai platform to create a unified intelligence layer from siloed data. Thoughtworks is assisting institutions in operationalizing these models within complex banking environments, focusing on integration into payment, servicing, and risk management, alongside establishing governance frameworks.

Datos clave
| Aspect | Description |
| :———————- | :———————————————————————– |
| Core Technology | Transformer-based Transaction Foundation Models |
| Key Benefit | Unified understanding of consumer financial behavior, breaking data silos |
| Performance Improvement | Enhanced accuracy in fraud detection, credit scoring, and recommendations |
| Industry Adoption | Featured companies: Revolut, Mastercard, Adyen, Stripe |
| NVIDIA Contribution | AI stack, GPUs, NeMo framework, developer examples |

This development is significant for ReviewArticle’s readership as it signifies a major architectural shift in how financial institutions leverage AI. By moving towards unified foundation models, firms can achieve deeper insights, improve risk management, and enhance customer experiences by understanding financial behavior in a more holistic and contextualized manner. This trend indicates a maturing AI landscape within finance, moving beyond isolated solutions to integrated intelligence systems.

Fuente: NVIDIA AI Blog, https://blogs.nvidia.com/blog/financial-institutions-transaction-foundation-models/

Datos clave