Revolut PRAGMA now has a division behind it. The bank launched Revolut Research on August 25 from London. It is a dedicated unit inside the AI department, built to develop proprietary machine learning models rather than buy them.

Revolut PRAGMA figures are specific enough to check later, which is more than most bank AI announcements offer. Revolut reports 2.3 times higher accuracy identifying credit default risk, 65 percent more fraud cases caught with 17 percent greater alert precision, and 41 percent more relevant product recommendations.

Read the qualifier carefully, because Revolut supplies it rather than hiding it. These are early deployments on historical data measured against legacy baselines. Backtesting flatters a new model, since it validates against outcomes already known. No figures exist yet for fresh transactions in production.

The Revolut PRAGMA Paper Is Already Published

One framing needs updating. Revolut is not promising to publish Revolut PRAGMA findings at some future point.

The paper is out. PRAGMA: Revolut Foundation Model sits on arXiv under identifier 2604.08649, carrying fourteen authors including Repushko. It describes a family of foundation models for banking event sequences. A transformer architecture gets pre-trained with masked modelling on a heterogeneous banking event corpus.

Technically the approach is interesting. Revolut PRAGMA uses a self-supervised objective tailored to the discrete, variable-length nature of financial records. The authors report strong downstream performance from a simple linear model trained on the extracted embeddings. Lightweight fine-tuning improves it further. That means the expensive part happens once, and individual tasks like credit scoring or fraud detection ride cheaply on top.

Revolut PRAGMA Trained on Less Data Than the Headline Suggests

Here the Revolut PRAGMA paper and the press release tell different stories, and the difference matters.

The announcement leads with 80 million customers across more than 40 markets. The research describes pre-training on 24 billion banking events. Those came from 26 million user records across 111 countries. So the training corpus covers roughly a third of the customer base rather than all of it.

Neither figure is wrong. They measure different things, and 111 countries reflects where transactions occurred rather than where Revolut holds licences. Even so, a reader taking 80 million as the training set would overstate the scale by a factor of three.

The metrics reconcile more cleanly. A 2.3 times accuracy gain equals the 130 percent increase in credit-scoring precision-recall reported in the paper. Likewise, 65 percent more fraud cases matches a 64.7 percent improvement in fraud recall. The press release chose the framing that sounds larger.

Revolut PRAGMA Is Not the First Bank Foundation Model

Data advantage is the load-bearing claim, and it needs qualifying. Revolut is not alone in publishing a banking foundation model, nor does it hold the largest dataset.

Nubank published nuFormer, trained on more than 100 billion transactions across 100 million customers. That is roughly four times the transactions behind Revolut PRAGMA, and nearly four times the users. Mastercard released LTM, oriented toward cyber-risk identification from card transaction data.

What distinguishes the Revolut approach is breadth rather than depth. PRAGMA ingests app navigation and support interactions alongside transactions. That is a multi-source architecture rather than a transaction-only one. That is a narrower and more defensible claim than having data nobody else can match, and it is the version worth making.

Scale still helps. Revolut security models review nearly a billion transactions monthly, feeding the same shared layer that powers AIR, its in-app assistant.

What to Watch on Revolut PRAGMA

Live performance is the gap. Whether outside auditors or regulators get access to validate the fraud and credit figures on production traffic, rather than backtests, is the question the announcement leaves open.

Explainability deserves attention too. An encoder model producing embeddings for a downstream classifier is harder to explain than a traditional scorecard. European rules on automated credit decisions require reasoning a customer can understand. How Revolut squares Revolut PRAGMA with those obligations is a genuine open question rather than a rhetorical one.

Finally, watch the publishing cadence. Repushko presented at ICML in Seoul this year. NVIDIA GTC Berlin follows in October and ICAIF in November, alongside quarterly meetups at Revolut offices. A research division that keeps publishing invites scrutiny, which is the point. Whether customers notice fewer false positives is a slower and more important test than any conference slot.

For related reading, our guide to AI in fintech tracks adoption across financial services. Our analysis of AI-driven fraud threats covers the detection problem PRAGMA targets, while our guide to challenger banking innovators maps the competitive field. Revolut published the division announcement on its newsroom. The PRAGMA paper is available on arXiv, and Tech Times compared it against other published banking models.

Fintechbits covers AI in banking, machine learning research and financial infrastructure. Nothing here constitutes financial or investment advice. All analysis represents the editorial views of Fintechbits.