Revolut Stopped Paying for AI and Started Winning
Revolut Stopped Paying for AI and Started Winning
Publish Date: 2026-04-21 13:54:00
Source Domain: www.pymnts.com
-
Neobank Revolution: Revolut’s Self-Built Modell: Revolut differs from most banks, which pay for AI access from companies like OpenAI or Anthropic, as it built its own AI model named PRAGMA.
-
Comprehensive Fraud and Risk Detection: PRAGMA is trained on 40 billion transactions and app interactions, assessing fraud risk, credit risk, and customer retention; claimed to outperform previous task-specific systems.
-
Mastercard’s Generative AI Effort: In contrast, Mastercard is building a generative AI model on anonymized payments for fraud, loyalty, and analytics—still focused on specific tasks rather than a general-purpose solution.
-
Limitations of General-Purpose AI in Finance: General-purpose AI like text-based models (e.g., ChatGPT) struggle here since financial data isn’t text; structured event sequences are misunderstood and real-time decision speeds aren’t achievable.
-
PRAGMA’s Unique Training Approach: PRAGMA reads financial events as structured data (what happened, specific details, when) to recognize complex patterns like a customer’s behavior leading to potential fraud.
-
Efficiency through One Model: PRAGMA consolidates multiple AI functions into one, replacing the need for individual models for fraud detection, credit scoring, and customer lifetime value into a single model.
-
Scalability and Speed: The model operates in real-time on thousands of GPUs, catering to different transaction speeds and decision accuracy levels with various parameter sizes.
-
Reduced Overhead with Model Adaptations: Adjusting the model’s parameters for new tasks is quicker than building a new model from scratch, allowing for faster deployments and minimized maintenance efforts.
