Revolut serves over 80 million customers in more than 40 markets, building a massive global dataset from billions of cross-border transactions to continuously improve its AI models and create a self-reinforcing intelligence loop for fraud detection, risk assessment, and user behavior forecasting.

The company plans to train native foundation models on its global operational data, reinforcing model performance as the dataset grows and enabling real-time financial insights.

PRAGMA is the underlying AI backbone, separate from AIR, with no data-sharing integration implied between PRAGMA and OpenAI’s ChatGPT bundle.

The initiative is contextualized by external partnerships and references, including links to Quantum Zeitgeist and NVIDIA, highlighting broader AI and hardware collaboration trends.

Revolut will centralize proprietary AI deployments and ML initiatives, moving toward an in-house, data-driven approach rather than relying on external software.

PRAGMA employs a three-branch encoder setup—Profile State Encoder, Event Encoder, and History Encoder—trained on a large banking corpus and fine-tuned with LoRA to minimize parameter updates.

Regulatory context matters: the EU AI Act became generally applicable in early August 2026, with high-risk obligations for credit scoring deferred, and GDPR-style explainability challenges for encoder embeddings raise trust questions.

Executives emphasize a unified, real-time foundation model built from scratch to understand financial behavior, leveraging collaboration with global tech leaders to differentiate from traditional banks.

Regulatory and product strategy includes a UK banking license secured and a US charter application, while PRAGMA’s open-source stance aims to build trust alongside performance.

Differentiation comes from owning research and infrastructure rather than simply integrating existing AI tools.

Early deployments on historical data show major performance gains, underscoring the efficiency of the unified foundation approach.

This move aligns with a broader European neobank trend to build proprietary AI models, echoing actions by peers like Starling and Bunq.