In Short – Key Takeaways
- Revolut has created Revolut Research, a dedicated division inside its broader AI department, announced August 25, 2026.
- The unit centers on PRAGMA, Revolut’s proprietary foundation model built with Nvidia, designed to unify fraud detection, credit risk and product recommendations into a single system.
- Early internal testing shows PRAGMA outperforming Revolut’s legacy models, including roughly 2.3x higher accuracy in flagging credit default risk and up to 65% more fraud cases caught.
- The announcement follows Revolut’s recent French banking licence, reinforcing its push to deepen roots across the EU while betting that proprietary AI, not off-the-shelf tools, will be its long-term edge.
Revolut has set up a new internal division, Revolut Research, dedicated to building artificial intelligence for its banking operations from the ground up. Announced on August 25, the unit will serve as the institutional home for PRAGMA, a foundation model the London-based neobank developed in partnership with Nvidia. The move signals a broader industry shift: rather than licensing generic AI tools, Revolut is betting that owning the entire model stack will give it a durable edge over both traditional banks and other fintech rivals.
Why Revolut is building its own AI stack
Most financial institutions still stitch together separate machine learning models for different tasks — one for fraud, another for credit scoring, a third for customer service. Revolut argues this approach is inefficient and increasingly outdated. PRAGMA instead trains a single foundation model on the bank’s own operational data, aiming to capture a more complete picture of customer behavior across products and markets.
Revolut operates across more than 80 million customers in over 40 markets, generating billions of transactions each month. PRAGMA itself, however, was trained on a narrower slice of that activity: the model’s pre-training corpus covers roughly 25–26 million user histories, spanning tens of billions of behavioral events. Revolut says that even this subset — sourced entirely from its own operations rather than third-party data — gives PRAGMA an advantage that’s difficult for incumbent banks to replicate, with the rest of its customer base available to expand training further over time. Pavel Nesterov, Revolut’s head of AI, framed the strategy as a deliberate rejection of “third-party blueprints,” describing the approach as “build, don’t bolt on.”
Early performance claims
According to figures released alongside the announcement, testing PRAGMA against historical data produced measurable gains over Revolut’s previous production systems:
| Metric | Reported improvement |
|---|---|
| Credit default risk accuracy | ~2.3x higher |
| Fraud cases caught | Up to 65% more |
| Fraud alert precision | +17% |
| Product recommendation relevance | +41% |
These figures come from the company itself and have not yet been independently verified through third-party audits or regulatory review, a point worth noting given that credit and fraud decisions carry real consequences for consumers.
Where the research division fits strategically
Beyond internal deployment, Revolut says the new division intends to publish research findings and open-source some technical frameworks, positioning itself within the broader AI research community rather than treating its work purely as a trade secret. Anton Repushko, head of Revolut Research, has already presented related findings at international conferences, and the team is expected to appear at further industry events later this year, alongside plans for regular internal meet-ups with academic partners.
The timing is notable. Earlier in August, Revolut secured a full banking licence in France, giving it a second EU regulatory hub alongside its existing operations and reinforcing its ambitions to operate as a licensed bank across the bloc rather than solely as an e-money institution. Building proprietary AI infrastructure in parallel suggests Revolut is trying to differentiate on technology just as licensing barriers to entry fall for it in more EU markets.
Revolut is not alone in this race. Payment giants including Stripe, Mastercard and Visa have all pursued their own foundation models for financial data over the past year, reflecting a broader consensus among large financial platforms that proprietary, large-scale behavioral data is becoming as strategically valuable as the algorithms built on top of it.
What comes next
For competitors, Revolut’s move raises the bar on what “AI-native” banking is expected to look like, and may pressure other digital banks and even traditional lenders to accelerate their own foundation-model efforts rather than relying on vendor tools. For regulators, particularly under the EU’s AI Act and existing financial supervision frameworks, a single model influencing credit decisions, fraud flags and customer-facing recommendations across tens of millions of accounts is likely to draw scrutiny over explainability and bias, especially as Revolut expands its banking licence footprint across the bloc. For customers, the near-term impact should be subtle — faster fraud detection and more tailored product suggestions — but the longer-term question is how much oversight such models will face as they take on a larger role in decisions that directly affect people’s access to credit and financial services.
Sources
- Revolut launches AI unit as proclaims virtues of proprietary AI
- Revolut launches dedicated AI research lab built on its own foundation model
- banksandbankers.com/category/fintech-and-innovation/
This article is for informational purposes only and does not constitute financial advice.
Add Comment