Revolut Launches Dedicated AI Research Division to Build Its Own Financial Intelligence
Revolut has launched Revolut Research, a dedicated AI research division focused on developing foundation models and machine learning systems for financial services.
Revolut is taking a bigger step into artificial intelligence.
The fintech company has launched Revolut Research, a dedicated research unit within its broader AI department. The new division will focus on machine learning architecture for financial services and will work with academic and technology partners on new AI systems.
At the center of the effort is PRAGMA, Revolut’s proprietary foundation model for financial data. The company says the model is being developed to support areas such as credit risk, fraud detection, product recommendations and other financial operations.
The announcement comes as banks and fintech companies race to move beyond basic AI assistants and task-specific machine learning systems. Revolut’s approach is different. Rather than adding separate AI tools to individual parts of its business, the company says it wants to build a shared intelligence layer across its financial platform.
What is Revolut Research?
Revolut Research is being established as the research arm behind the company’s broader AI strategy.
The division will develop machine learning systems and work with researchers and technology companies. It will also provide a central structure for Revolut’s proprietary AI projects.
That matters because financial institutions typically operate many separate models.
One system may handle fraud detection. Another may assess credit risk. A third may recommend products to customers.
Each system can require its own data pipelines, features and engineering work.
Revolut wants to reduce that fragmentation.
Its plan is to use a shared foundation model that can learn patterns from financial behavior and then adapt those representations to different tasks.
The company’s PRAGMA research supports that idea. The model is designed to learn from sequences of banking events rather than treating financial activity as ordinary text.
PRAGMA is at the center of the strategy
PRAGMA is a family of Transformer-based foundation models designed specifically for banking event sequences.
That distinction is important.
Large language models learn from text. Financial activity looks very different. A customer’s history can contain payments, transfers, card activity, account events and other interactions that occur at different points in time.
PRAGMA is designed to capture those patterns as a sequence.
The research paper behind the model describes a self-supervised approach that learns representations from heterogeneous banking event data. Those representations can then be adapted for downstream applications such as credit scoring, fraud detection and customer lifetime-value prediction.
In other words, Revolut is not positioning PRAGMA simply as a chatbot.
It is building a financial intelligence layer that can sit underneath multiple applications.
How Revolut says PRAGMA performs
Revolut says early deployments on historical data have produced significant improvements over its previous machine learning baselines.
The company reports:
- 2.3 times higher accuracy in identifying credit default risk.
- 65% more fraud cases detected, with 17% greater precision in fraud alerts.
- 41% more relevant product recommendations across retail and business accounts.
NVIDIA has published similar benchmark results from its work with Revolut. Its case study reports a 2.3x improvement in credit-default risk accuracy, a 64.7% improvement in fraud detection and a 16.7% increase in fraud precision, alongside a 41% improvement in product-recommendation relevance.
The small differences between the figures in Revolut’s announcement and NVIDIA’s published material appear to reflect how the results are presented and rounded.
These figures should also be viewed as reported benchmark results, rather than proof that PRAGMA will produce the same gains in every live banking environment.
The data behind Revolut’s AI strategy
The biggest advantage for Revolut may not be the model architecture alone.
It is the amount and variety of financial activity available to train and evaluate its systems.
Revolut says it now serves more than 80 million retail customers across more than 40 markets.
NVIDIA’s technical case study provides a more specific picture of the dataset used for PRAGMA. It says the model was trained using records from about 26 million users, covering more than 24 billion events across more than 100 countries.
That distinction is useful.
The company’s total customer base is not the same thing as the dataset used to train a particular model.
The PRAGMA research instead focuses on a large collection of historical financial events. The model learns relationships between those events and uses them to build representations of customer behavior.
Why financial event data is different
Financial behavior has a strong time dimension.
A payment made today can mean something different when it follows several transactions from the previous hour. Likewise, a customer’s behavior over months can reveal patterns that are not visible in a single transaction.
PRAGMA is designed to capture those relationships.
Its architecture processes user profiles alongside sequences of financial events. The research also uses temporal information to help the model understand the time between events.
That approach gives Revolut a potential advantage over systems that depend heavily on manually engineered features.
Instead of creating a new collection of features for every individual use case, the company can train a common representation and adapt it for different applications.
NVIDIA is a key technology partner
Revolut developed PRAGMA with technology support from NVIDIA, while NVIDIA’s published case study identifies Nebius as the cloud partner.
The training work used NVIDIA H100 GPUs on Nebius AI Cloud.
NVIDIA says Revolut achieved 2–5 times higher training throughput through techniques such as sequence packing and dynamic batching. It also reports that model-development cycles became roughly 3–5 times faster in relevant workflows.
This is an important part of the story because foundation models are expensive to train.
For a financial company, the challenge is not only building a model. The company also needs to update it as customers, products and markets change.
Faster training can therefore translate into faster experimentation and deployment.
Revolut wants to move away from fragmented AI systems
The company’s strategy reflects a broader change taking place across financial technology.
Traditional machine learning systems are often built around individual business problems.
Fraud teams build fraud models. Credit teams build credit models. Marketing teams build recommendation systems.
That approach can work well, but it can also create duplicated infrastructure.
Different teams may use different versions of the same customer information. Engineers may also spend large amounts of time preparing similar features for separate models.
Revolut says PRAGMA is intended to reduce that fragmentation.
The goal is to create a shared behavioral representation that can support multiple areas of the business.
NVIDIA’s case study says this approach can reduce feature-engineering work and allow updates to the shared model to flow into several business applications.
Fraud detection is one of the biggest use cases
Fraud prevention is likely to be one of the most important practical applications.
Revolut says its advanced security models review nearly one billion transactions every month to identify suspicious activity.
PRAGMA adds another layer to that broader AI strategy.
Instead of looking at transactions in isolation, a foundation model can learn patterns across a customer’s broader behavioral history.
That could help identify unusual combinations of activity that may be difficult for narrowly focused systems to recognize.
Still, higher benchmark performance does not automatically mean that every fraud case will be detected in production.
Fraud systems must also balance detection with false positives. An overly aggressive system can incorrectly flag legitimate customers and create a poor user experience.
That is why the reported increase in fraud precision is particularly relevant alongside the increase in detection.
AI could also change how Revolut recommends products
Fraud and credit risk are not the only applications.
Revolut says PRAGMA can also support personalized product recommendations.
The company reports a 41% improvement in recommendation relevance across retail and business accounts.
The underlying idea is relatively simple.
A financial app knows much more about a customer’s financial activity than a typical online retailer knows about a shopper.
It can see patterns involving spending, saving, transfers, travel and other financial behavior.
If an AI system can interpret those patterns responsibly, it could potentially recommend products that are more relevant to a customer’s circumstances.
That could make personalization a much bigger part of digital banking.
Revolut is also building AI directly into the customer experience
Revolut’s AI strategy extends beyond research.
The company already uses AI across its platform. Its in-app assistant, AIR, is designed to execute more complex financial tasks in a single step.
Revolut says AIR is currently available only in the UK.
That creates a link between the company’s research work and its consumer products.
The research division develops the underlying intelligence. Product teams can then potentially use those capabilities to build new customer-facing features.
Revolut Research plans to publish its work
The new research division is also positioning itself as a participant in the wider scientific community.
Revolut says it plans to publish research findings and open-source technical frameworks.
Its researchers are expected to participate in major industry and academic events. The company specifically mentioned NVIDIA GTC Berlin in October and ICAIF in November.
Revolut also plans to host quarterly scientific meetups at its offices.
That is notable because the company is not keeping PRAGMA entirely behind closed doors.
The underlying research has already been published publicly. The paper, titled “PRAGMA: Revolut Foundation Model,” is available through arXiv and lists researchers from Revolut among its authors.
What this means for the banking industry
Revolut’s move signals a broader shift in financial AI.
Banks have spent years using machine learning for individual tasks. The next stage may involve foundation models that can learn a common representation of financial behavior and support many applications.
That could change how financial institutions build AI teams.
Instead of maintaining dozens of independent models, companies may increasingly develop shared intelligence layers.
However, the model itself is only one piece of the puzzle.
Financial institutions also need strong data governance, privacy controls, security, regulatory oversight and human supervision.
Those requirements become even more important when AI systems are used for credit decisions, fraud prevention or other areas that can directly affect customers.
What comes next for Revolut Research?
Revolut says its research team plans to expand PRAGMA beyond its current capabilities.
The company’s technical work points toward areas such as additional data types, continuous pre-training and new applications including customer lifetime-value prediction, churn forecasting and anomaly detection.
The company is also exploring ways to make its AI infrastructure more adaptable as new markets and products are added.
That could become increasingly important as Revolut expands internationally.
The company now operates across more than 40 markets. A common AI foundation could allow it to reuse some of its technology while adapting models to different products and customer behaviors.
The bigger picture
Revolut’s launch of Revolut Research is more than the creation of another internal AI team.
It represents a decision to make financial intelligence a core part of the company’s technology stack.
PRAGMA is central to that strategy. Its purpose is to understand financial event sequences and turn those patterns into reusable representations for different applications.
The early benchmark results are striking, particularly in credit risk, fraud detection and recommendations. Yet the more significant development may be architectural.
Revolut is trying to replace a collection of narrow machine learning systems with a shared foundation.
That approach could make AI development faster and reduce duplicated engineering work.
It could also give fintech companies a new way to compete with traditional banks, provided they can manage the challenges that come with using large-scale behavioral data.
For now, Revolut Research is still at an early stage.
But the direction is clear: Revolut wants AI to become part of the financial infrastructure itself, rather than simply another feature inside a banking app.
Conclusion
Revolut’s launch of Revolut Research marks a significant step in its push to build proprietary AI for banking.
The division will sit behind PRAGMA, a foundation-model family designed around financial event data. The technology is being developed with NVIDIA and has already shown strong reported results across credit risk, fraud detection and product recommendations.
The real test, however, will come outside the benchmark environment.
Revolut will need to show that the technology can deliver reliable improvements in live financial products while maintaining privacy, security and responsible decision-making.
If it succeeds, the impact could extend beyond Revolut.
The company is effectively betting that the next generation of digital banking will be built around shared financial intelligence, not a collection of disconnected AI tools.
For an industry still working through the transition from traditional banking software to modern AI infrastructure, that is a significant bet.
Source note: This article is based on Revolut’s August 25, 2026 announcement, NVIDIA’s technical case study and the published PRAGMA research paper. Reported performance figures are attributed to those sources and should not be interpreted as independently verified guarantees of future production performance.
Revolut Announces Launch of Dedicated AI Research Division: Revolut Research


Comments
No comments yet. Be the first to share your thoughts.
Leave a comment