Revolut has launched Revolut Research, a dedicated AI research division, with the Revolut Research AI division’s centrepiece being PRAGMA, a proprietary foundation model built with Nvidia that the company says is designed to understand complex financial behaviour. The launch positions Revolut among a growing group of European digital banks that are moving beyond off-the-shelf AI tools and into the more demanding territory of training their own models from scratch.
PRAGMA is built to power fraud detection, credit risk assessment, platform operations and personalised product recommendations. Revolut says early deployments have delivered 2.3x better credit risk accuracy, 65% more fraud cases detected and 41% more relevant product recommendations. Those are the kinds of numbers that are easy to headline but harder to contextualise without seeing the baseline, which Revolut has not disclosed.
What Revolut Research AI division is actually building
The new unit sits within Revolut’s broader AI department and will work alongside academic and technology partners. According to Fintech Brain Food, PRAGMA’s 1B model was trained on 32 H100 GPUs in roughly two weeks, which gives some indication of the compute involved in getting a foundation model to a deployable state. The model is fed by data from 80 million customers, and Revolut describes this as a self-reinforcing loop: as the dataset grows, the models are intended to become more capable.
Pavel Nesterov, head of AI at Revolut, framed the unit’s founding philosophy in direct terms. ‘To lead the future of intelligent banking, you cannot rely on third-party blueprints,’ he said. ‘We have launched Revolut Research to institutionalise our “build, don’t bolt on” philosophy. By training native foundation models on our global operational data, we are giving our engineering teams an unprecedented engine to deploy smarter features faster, eliminate systemic friction, and give our customers a safer, radically better financial experience.’
Revolut Research also plans to publish its scientific findings regularly, open-source technical frameworks, and engage with the wider academic community through international conferences and meetups hosted at Revolut’s offices. Whether that openness extends to the data and methods behind headline performance figures like the fraud detection improvement is a different question.
The neobank AI arms race, and where Revolut fits
The launch arrives at a moment when AI has become a clear competitive front for Europe’s digital banks. The race arguably started last June, when Starling Bank introduced a spending intelligence tool that let customers ask questions about their finances in plain English. Dutch rival Bunq upgraded its AI customer support assistant in December. By March, Starling had gone further, launching what it described as the UK’s first ‘agentic’ AI assistant. Revolut followed with its own assistant in April.
Starling also announced last week the launch of smart tools drawn from crowdsourced customer feedback, with a weekly rollout planned through to the end of the year. The cadence was partly inspired, Starling said, by the continuous release cycles of frontier large language models (LLMs).
The difference Revolut is now staking out is less about the user-facing feature and more about what sits underneath it. Training a foundation model on proprietary financial data is a longer and more expensive bet than wrapping a third-party LLM in a banking interface. The 32-H100, two-week training run reported by Fintech Brain Food is modest by frontier AI standards, but it signals genuine infrastructure commitment rather than a chatbot layer dressed up as research.
The question the Revolut Research announcement leaves open is whether a purpose-built financial foundation model produces durable advantages over general-purpose models that are themselves improving rapidly. Revolut’s bet is that domain-specific training on 80 million customers’ data creates something those models cannot easily replicate. The early performance figures suggest something is working. Whether it holds as the broader LLM field catches up is what the new research unit will ultimately have to answer.
