Revolut’s AI Strategy Banking Bet: One Model, 64.7% Less Fraud

Revolut has built something that most banks are still sketching on whiteboards. The fintech giant’s proprietary AI foundation model, PRAGMA, doesn’t just automate isolated tasks — it understands the full arc of a customer’s financial life, from their morning coffee transaction to their investment portfolio, and uses that unified picture to make smarter decisions across every function simultaneously. That’s a fundamentally different approach to AI strategy in banking, and the numbers behind it are hard to ignore.

Key takeaways

• PRAGMA is Revolut’s proprietary AI foundation model, trained on 40 billion events and interactions with 25 million users, now serving over 70 million customers.
• The model improved fraud detection by 64.7%, credit risk prediction by 16%, and product recommendations by 41%.
• Powered by 200 NVIDIA H100 GPUs, PRAGMA runs as a single shared model where gains in one function improve performance across all others.
• Revolut’s AI assistant handles 75% of customer support requests without human involvement.
• AIR, Revolut’s agentic AI system, is now available to UK customers for managing subscriptions, budgeting, and travel arrangements.

Revolut Unveils PRAGMA: A Unified AI Model for Banking

Most financial institutions have approached AI the same way they approached software for decades — bolt on a tool for fraud, another for credit scoring, another for customer service, and hope the pieces eventually talk to each other. Revolut went in the opposite direction. PRAGMA is a single foundation model trained on 40 billion events and interactions from 25 million users, designed to understand financial behavior holistically rather than in isolated slices.

The scope of what PRAGMA ingests is what makes it structurally different. It pulls together transactions, app and website usage, trading and investing activity, bill payments, subscription behavior, and customer support interactions — all into one connected system. That breadth of data gives the model a level of behavioral context that narrower, task-specific tools simply cannot replicate.

Technical Infrastructure and Scalability

Running a model at that scale requires serious hardware. PRAGMA is powered by 200 NVIDIA H100 GPUs, and the infrastructure has allowed Revolut to grow from 38 million users in 2023 to more than 70 million today without fragmenting its AI stack. That’s not just a technical achievement — it’s a strategic one. Keeping the model unified as the user base nearly doubles means the intelligence compounds rather than dilutes.

Performance Improvements Driven by PRAGMA

The results Revolut reports are specific enough to deserve attention. PRAGMA delivered a 64.7% improvement in fraud detection — a figure that matters enormously in an industry where financial crime costs run into the billions globally. On the credit side, risk prediction performance improved by 16%, which directly affects how accurately Revolut can price and extend lending products. Product recommendations became 41% more effective, meaning customers are more likely to see offers that are actually relevant to their financial behavior.

What’s analytically significant here is not just the magnitude of each improvement, but the mechanism behind them. These gains all flow from a single algorithmic and data pipeline. When the model improves its understanding of fraud patterns, that learning feeds into how it evaluates credit risk. When it gets better at recognizing behavioral signals for product recommendations, those same signals sharpen its fraud detection. The model learns in tandem across functions, adapting as economic and behavioral trends shift. That’s compounding intelligence — and it’s the structural advantage that fragmented AI stacks cannot replicate.

Agentic AI Tools and Customer Support Automation

AI Handling of Customer Support Requests

PRAGMA is also the engine behind Revolut’s customer service operation. Its AI assistant now handles 75% of support requests without any human intervention. For a platform serving more than 70 million users across multiple markets, that’s a substantial operational shift — and a signal of how far automated financial support has come from the clunky chatbots of five years ago.

Introduction of AIR for UK Customers

The more forward-looking development is AIR — AI by Revolut — the company’s first customer-facing agentic AI system. Currently available to UK customers, AIR moves beyond answering questions into taking action. It can manage subscriptions, cancel lost cards, assist with budgeting, and even arrange travel. That’s a meaningful step: rather than surfacing information, the AI acts on a customer’s behalf with real-world consequences.

Agentic AI in consumer finance is still relatively new territory, and Revolut’s decision to launch AIR in the UK first suggests a measured rollout — testing autonomous financial actions in one regulated market before broader expansion.

Strategic Impact of Revolut’s Unified AI Architecture

Single Shared Model Enabling Cross-Functional Learning

The strategic bet Revolut has made is that a single shared model will outperform a collection of best-in-class specialized tools over time. The logic holds: because all improvements feed back into the same model, the system gets smarter across every function simultaneously. A fraud detection refinement doesn’t stay siloed in the fraud team — it ripples through credit, recommendations, and customer experience.

Competitive Advantages for Agility and Personalization

For larger incumbent banks still running legacy systems stitched together with point solutions, this architecture gap is genuinely difficult to close. Revolut’s unified stack allows it to respond to shifting customer behavior or emerging fraud patterns faster than institutions that need to coordinate updates across multiple disconnected models and data pipelines.

The broader implication for the industry is that AI architecture is becoming a strategic moat, not just an operational efficiency tool. Banks and fintechs investing in unified data and model infrastructure are better positioned to deliver personalized services at scale — and to build the kind of agentic capabilities that could redefine what a financial app actually does for its users. Revolut’s PRAGMA offers a concrete proof point that the unified model approach works, and its performance metrics give competitors a benchmark they’ll find uncomfortable to ignore.

FAQ

What is PRAGMA and how does it function?

PRAGMA is Revolut’s proprietary AI foundation model trained on 40 billion events and interactions from 25 million users. It holistically integrates data from transactions, app usage, investments, and customer support to understand financial behavior as a whole, rather than processing each function separately.

How has PRAGMA improved Revolut’s banking services?

PRAGMA improved fraud detection by 64.7%, credit risk prediction by 16%, and product recommendations by 41%. Because all gains flow from a single shared model, improvements in one area strengthen performance across all other functions simultaneously.

What is AIR and what functionalities does it offer?

AIR — AI by Revolut — is the company’s first customer-facing agentic AI system, currently available to UK customers. It autonomously manages subscriptions, assists with budgeting, cancels lost cards, and can make travel arrangements on behalf of users.

How does Revolut’s AI strategy differ from traditional banking AI approaches?

Traditional banks typically run fragmented AI stacks with separate models and data pipelines for each task. Revolut uses a single unified foundation model, which means intelligence compounds across all functions — learnings from fraud detection improve credit risk prediction, and vice versa — giving it a structural advantage that isolated tools cannot replicate.

Article produced with the assistance of artificial intelligence and reviewed by the editorial team.