Revolut's PRAGMA AI Trains on 207 Billion Data Points to Consolidate Financial Models

Revolut's new PRAGMA AI model consolidates multiple financial decision-making systems into one, trained on 207 billion data points from customer activity. This system enhances real-time capabilities for functions like fraud detection and credit risk assessment.

Financial institutions are increasingly leveraging artificial intelligence, but financial data presents unique challenges for AI systems. Unlike generic text-based models that analyze human language, financial AI must interpret structured sequences of events, such as deposits, transfers, and purchases, to identify nuanced patterns. A $1,000 transfer and a $10 transfer, for instance, appear similar to text-based AI. Furthermore, critical financial operations like fraud checks demand processing speeds in milliseconds, a requirement not met by slower, general-purpose AI models. As a neobank, Revolut maintains a comprehensive view of its customers' financial lives, encompassing spending habits, trading activities, and international money movements across its platform.

Revolut has introduced PRAGMA, a proprietary AI model designed to replace its diverse array of task-specific systems. The company states that PRAGMA outperforms all its previous specialized solutions for financial decision-making. PRAGMA was trained on 207 billion data points, derived from 25 months of real customer activity from its global user base. This model interprets each financial event by analyzing "what happened," its specific details, and the timing, allowing it to discern complex user behavior patterns indicative of fraud or credit risk. Operationally, PRAGMA is available in three distinct sizes, ranging from a lean 10 million parameters (learnable values within an AI model) for rapid processing to a 1 billion-parameter version for higher accuracy demands. All three iterations run on over 200 NVIDIA H100 GPUs, powerful graphics processing units.

The deployment of PRAGMA signals a strategic shift from a fragmented, multi-model AI architecture to a single, consolidated foundation. Historically, financial institutions have managed separate AI systems for distinct functions, such as fraud detection, credit scoring, and product recommendations. Each individual system necessitated its own training data, maintenance protocols, and dedicated engineering teams. PRAGMA unifies this approach, enabling Revolut to adapt the model to new tasks by adjusting a small fraction of its parameters rather than building entirely new systems from scratch. This consolidation accelerates development cycles and the deployment of new functionalities, providing increased operational efficiency. The system's ability to process a complete picture of customer financial data offers enhanced accuracy and speed for decisions critical to risk management, fraud prevention, and customer engagement.

The financial sector will observe whether this consolidated AI model approach, leveraging extensive proprietary transaction data, establishes a new standard among other digital banks and payment processors aiming for greater efficiency and more sophisticated decision-making capabilities.