TL;DR
Revolut has set up an internal division to build its own AI rather than licence someone else’s. The first output is PRAGMA, a model developed with Nvidia and trained on transaction records spanning 80 million customers across 40-plus markets. The unit intends to publish research and open-source parts of its framework.
Build, not bolt on
The phrase belongs to Pavel Nesterov, who runs AI at the bank: the new division exists to institutionalise a “build, don’t bolt on” philosophy. In practice that means an in-house framework acting as the backbone for the bank’s machine learning work, in place of third-party software.
Anton Repushko will lead it. A senior engineer on the software development side, he arrived in 2024 from i2x, and before that from AWS.
PRAGMA is pitched at reading financial trends, assessing risk as it happens, running platform operations and generating product recommendations — and, as the dataset grows, improving on all of them. Separate security models already screen close to a billion transactions monthly for fraud, alongside work on AIR, the in-app assistant currently confined to British users.
Why the data claim matters
The strategic asset here is not the model. It is 80 million customers’ worth of transaction history across 40-plus markets, which is not something a competitor can licence, and which a general-purpose foundation model has never seen.
That is the honest argument for building rather than buying in financial services: the differentiator sits in proprietary data, and a bought model applies the same capability to everyone. The counter-argument is cost, and Revolut is large enough to absorb it where most UK fintechs are not.
Looking forward
Timing is deliberate. Full banking authorisation has landed in France, Australia and Britain over recent months, and a licensed bank carries model governance obligations that a payments firm does not.
For the wider UK sector this lands alongside HSBC, Lloyds and NatWest backing a sovereign British model earlier in the summer, and Lloyds recruiting for agentic AI roles. British banks are increasingly deciding that the intelligence layer is something to own. Whether the returns justify the expense is a question none of them can answer yet.