Revolut unveils RevolutOS

Revolut is moving more of its banking decisions onto one in-house artificial intelligence system instead of separate models for lending, fraud and product recommendations. (arxiv.org) The company’s research team published a paper on April 9, 2026 describing PRAGMA, a Transformer model trained on raw banking event sequences rather than hand-built features for each task. The paper says the same base model can be adapted for credit scoring, fraud detection and customer lifetime value prediction. (arxiv.org) A banking event sequence is the running history of card payments, transfers, balances and other account actions over time — more like a movie of customer behavior than a single transaction snapshot. Revolut said that approach lets a simple layer on top of the model handle downstream tasks that banks usually split across separate systems. (arxiv.org) That design lines up with how regulators now describe artificial intelligence use in banking. European Central Bank supervisors said in November 2025 that banks are using AI more often for credit scoring and fraud detection, with decision-tree models common in both areas and neural networks used mainly in fraud work. (bankingsupervision.europa.eu) Revolut’s timing also reflects the scale it has reached. The company said on March 24, 2026 that it ended 2025 with 68.3 million retail customers, 767,000 business customers and $67.5 billion in customer balances, giving it a much larger stream of financial behavior data to train on than smaller fintechs. (revolut.com) The company has been adding AI features in public for more than a year. Revolut launched an AI scam-detection tool in February 2024 and said initial testing showed a 30% reduction in fraud losses in certain card scam cases, then rolled out an in-app assistant called AIR in the UK on April 9, 2026. (revolut.com, revolut.com) The harder part is not just technical performance but compliance. Under the European Union’s Artificial Intelligence Act, systems used to evaluate a person’s creditworthiness are listed as high-risk, while systems used to detect financial fraud are carved out from that specific category. (eur-lex.europa.eu, ai-act-service-desk.ec.europa.eu) That means a single model spanning both functions could leave Revolut with a tougher governance job than a bank running isolated tools, because one shared system would touch underwriting, monitoring and personalization at once. The company’s paper focuses on model architecture and task performance, not on how product-level controls, audits or explainability will be separated in production. (arxiv.org, ai-act-service-desk.ec.europa.eu) Revolut has spent the past two years arguing that software, not branch networks, is its operating advantage. A single model for money decisions pushes that logic deeper into the bank: one representation layer underneath lending, fraud checks and the app experience. (revolut.com, arxiv.org)