How Revolut built a single AI model to power fraud detection, credit risk, and customer engagement — simultaneously
Based on a presentation by Viktor Smits, Senior Strategy & Operations Manager, Revolut
Pragma: Revolut’s Foundation Model
Overview
In a world where financial institutions typically build dozens of isolated AI models — each trained from scratch for a narrow task — Revolut took a fundamentally different approach. The company developed Pragma, a single foundation model trained on 24 billion banking events across 26 million customers in more than 100 countries.
Rather than instructing the model to detect fraud or assess credit risk, Revolut let the model learn the underlying patterns of financial behaviour on its own, then adapted it for specific tasks with minimal additional effort.
The results were striking: Pragma improved external fraud detection recall by 65%, boosted credit risk prediction efficiency by 130%, and outperformed existing models in targeted customer communications by 136% — all within months of development, and without hand-crafting a single feature.
The Problem with Traditional Model Building
Before Pragma, Revolut’s approach to AI mirrored the industry norm. Each team — fraud, credit, compliance — would independently assemble data engineers, spend weeks or months in manual feature engineering, and eventually train a bespoke model. The process was then repeated, in full, for every new use case.
This created three compounding problems:
- Domain bottlenecks: Each team owned its data and its model, with little
transfer of insight between them. - Duplication of effort: Feature engineering was reinvented repeatedly across teams.
- Slow iteration: Launching any new AI capability meant restarting the pipeline from scratch.
The Key Insight: Banking Events Behave Like Language
Pragma’s conceptual breakthrough was recognising that sequences of financial
transactions share structural similarities with natural language. Both are sequential and contextual. Both contain patterns — and anomalies. In language, an anomaly might be a typo; in banking, it might be fraud. This parallel meant that the architectural techniques behind large language models (LLMs) could be adapted to read and reason about financial event sequences.
However, classical LLMs struggle with financial data because they are not designed to handle numerical values like transaction amounts. Pragma addresses this through careful tokenisation — breaking each transaction into its type, direction (inbound or outbound), and a discretised amount bucket rather than a raw figure. This allows the model to generalise across transactions without being confused by specific monetary values.
How Pragma Was Built
The Training Data
Revolut trained Pragma on data from four categories of events, all drawn from 26 million anonymised customer accounts across more than 100 countries:
- Transactions — payments, transfers, card usage, and similar financial activity.
- In-app events — the flows and actions customers took within the Revolut application.
- Trading events — investment and trading activity on the platform.
- Communication events — whether customers engaged with emails and push notifications sent by Revolut.
In total, the model was trained on 24 billion distinct events, with training data spanning a two-year window.
The Architecture
Pragma uses a two-part architecture. A profile state encoder captures static
information about each customer — their age, location, account creation date, and subscription plan. This is combined with a dynamic event stream of recent transactions and interactions. Together, these inputs are used to construct an
individual mathematical profile for every user.
Training uses masked modelling — the same technique at the core of many LLMs. Roughly 15% of events in a sequence are randomly removed, and the model is trained to predict the missing entries. Run billions of times, this process forces the model to learn the deep statistical structure of financial behaviour without ever being told what to look for.
Fine-Tuning with LoRA
Once the foundation model is trained, adapting it to a specific task requires very little additional work. Using a technique called Low-Rank Adaptation (LoRA), Revolut freezes 98–99% of the model’s parameters and retrains only 1–2% for the target task. This makes fine-tuning fast and computationally efficient — meaning the same base model can be quickly adapted for fraud detection, credit risk, marketing, product recommendations, or any other downstream application.
Results: Performance Improvements
All results below reflect relative improvements over Revolut’s existing production models — the cumulative work of fraud, credit, and operations teams over several years.
- External fraud detection: Pragma improved recall by approximately 65%,meaning Revolut now catches two-thirds more fraudulent transactions than before.
- Credit risk: Revolut serves a relatively young, thin-file customer base — people with limited credit histories. Pragma improved credit default prediction efficiency by 130%.
- Customer reactivation communications: When customers drop out of a flow (such as a credit application), Pragma is 136% more effective at identifying the right channel and message to re-engage them.
- Additional use cases: The model also showed strong performance in product recommendations and predicting recurring transactions.
Where Pragma Falls Short
Pragma is not without limitations. The model underperforms on money laundering detection — and the reason is instructive. Anti-money-laundering work is inherently relational: it requires analysing networks of individuals and understanding how transactions flow between them. Pragma, by contrast, is designed around individual customer profiles. It excels at understanding a single person’s behaviour in depth, but cannot readily analyse connections across people.
There is also a challenge around explainability. For regulated activities such as credit decisioning and fraud detection, Revolut may be required to explain why a model reached a particular conclusion. Pragma, like most deep learning models, is not inherently interpretable. Revolut is exploring techniques to inspect the model’s reasoning — or to use Pragma’s outputs as inputs into more deterministic, explainable models.
The Operational Opportunity
Beyond the performance gains, Pragma fundamentally changes how Revolut develops and maintains AI capabilities. The key shifts are:
- No feature engineering: The weeks or months previously spent crafting features for each model are eliminated entirely.
- One model, many applications: A single foundation model serves as the basis for all downstream tasks, ending duplication across teams.
- Low maintenance: Instead of maintaining dozens of separate models, teams maintain and improve one.
- Rapid deployment: Fine-tuning for a new task is fast, since only 1–2% of model parameters need to be updated.
Current Status and the Road Ahead
As of the time of this presentation, Pragma is a research initiative rather than a fully deployed production system. Revolut trained the model on anonymised data in collaboration with NVIDIA — the computational demands of training at this scale are significant, running to millions of pounds per month without the benefit of proprietary infrastructure.
The team’s intention is to run Pragma in parallel with existing models for an extended period before committing to a full transition. Regulatory constraints also apply: under current EU AI Act frameworks, the model can be used for marketing and communications but requires further adaptation before it can be deployed for credit or fraud decisions.
The full technical paper has been published and is publicly available for those who wish to explore the methodology in depth.
Conclusion
Pragma represents a meaningful shift in how financial institutions can think about AI. Rather than treating each risk or product challenge as an isolated modelling problem, Revolut has demonstrated that a single, large-scale foundation model — trained without task-specific labels — can outperform purpose-built models across a range of high-stakes applications.
The approach is not without challenges — the compute costs are substantial,
regulatory paths require careful navigation, and money laundering remains a domain where the model struggles. But for a company operating at Revolut’s scale, the economics and performance gains are compelling. Pragma offers a glimpse of what AI infrastructure in financial services could look like: not a collection of siloed models, but a single intelligence that learns from the full breadth of customer behaviour.
The team’s intention is to run Pragma in parallel with existing models for an extended period before committing to a full transition. Regulatory constraints also apply: under current EU AI Act frameworks, the model can be used for marketing and communications but requires further adaptation before it can be deployed for credit or fraud decisions.
The full technical paper has been published and is publicly available for those who wish to explore the methodology in depth.
Presentation by Viktor Smits, Senior Strategy & Operations Manager, Revolut
