Revolut’s new in-house AI research division has produced a banking foundation model that outperforms task-specific systems across fraud detection, credit scoring, and personalization — but the architecture that drives those gains creates a structural tension with regulations requiring explainable decisions in AI-assisted credit, obligations that are now coming into force for European deployers.

The neobank announced on August 25, 2026 the formal launch of Revolut Research, a dedicated division within its AI department built around PRAGMA — a family of encoder-style transformer models pre-trained on 24 billion banking events drawn from 26 million user records across 111 countries. According to the company, PRAGMA has already delivered a 64.7 percent improvement in fraud recall, a 130 percent increase in credit-scoring precision-recall performance, and a 40 percent gain in the relevance of product recommendations — all measured against Revolut’s own internal specialist models on historical data.

One Model, Every Task

For years, the dominant approach to AI in financial services has been fragmentation: a separate model for fraud, another for credit, another for customer recommendations, each with its own training pipeline and feature engineering cycle. The cost of maintaining dozens of isolated systems is compounding at every institution that has followed that path.

PRAGMA is designed to retire that model. Rather than training task-specific systems from scratch, Revolut pre-trained a single backbone on the full breadth of a customer’s financial life — card payments, currency exchanges, in-app navigation, support requests — and then adapts that backbone to specific tasks through Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning technique that updates only 2 to 4 percent of the model’s weights. A new fraud typology that might previously have required weeks of feature engineering can now be addressed with a LoRA adapter that trains in as little as 12 hours.

The PRAGMA paper, submitted to arXiv in April 2026 (2604.08649) by a joint team of Revolut Research and Nvidia engineers, describes a three-branch encoder architecture. A Profile State Encoder ingests static user attributes — account tenure, subscription plan, geographic region. An Event Encoder processes each individual banking event independently, using a key-value-time tokenization scheme that handles numerical, categorical, and text fields with type-appropriate methods (percentile bucketing for amounts, single-token encoding for merchant categories, subword tokenization for free-text descriptions). A History Encoder then combines those outputs across the full event timeline to produce a record-level embedding. Pre-training follows a masked-modeling objective borrowed from BERT: a fraction of input tokens are masked, and the model learns to reconstruct them using bidirectional context across the user’s entire history.

How the Training Numbers Work

The corpus used to pre-train PRAGMA spans a 25-month temporal range from 2023 to 2025, capturing 24 billion events and approximately 207 billion tokens. The smallest model variant (PRAGMA-S, 10 million parameters) trained in approximately two days on 16 Nvidia H100 GPUs. The largest (PRAGMA-L, one billion parameters) required 32 H100 GPUs running for roughly two weeks — compute supplied through Nebius AI Cloud, the Amsterdam-registered, Nasdaq-listed AI infrastructure company formed from the international assets of Yandex after that company divested its Russian operations in 2024.

To reach those training speeds, the team designed a storage and batching system specifically around the irregular nature of financial event data. Banking histories are long-tailed: some users generate a handful of transactions per month, others tens of thousands. Naïve padding-based batching — the default approach in most transformer training pipelines — would waste the majority of GPU compute on blank padding tokens. Revolut’s engineers instead partitioned records by event count into Parquet shards, then applied sequence packing with variable-length attention kernels, eliminating padding overhead along both the event and token dimensions. The result, per the paper, was two to five times higher training throughput compared to a conventional padded baseline.

The model is a family of three size variants — 10 million, 100 million, and one billion parameters — allowing Revolut to choose the operational point that balances performance against inference cost for each downstream task. The paper reports that scaling gains are task-dependent: credit scoring shows the most pronounced benefit from moving to the large variant (+35.2 percent PR-AUC over the small variant with LoRA fine-tuning), while simpler tasks like recurrent-transaction detection see minimal improvement beyond the smallest model.

Why Build Instead of Buy

Pavel Nesterov, Revolut’s Head of AI and a co-author of the PRAGMA paper, framed the division’s launch in terms that leave little ambiguity about the competitive intent. “To lead the future of intelligent banking, you cannot rely on third-party blueprints,” Nesterov said in the company’s official announcement. “We have launched Revolut Research to institutionalize our ‘build, don’t bolt on’ philosophy.”

The economic logic of that philosophy runs through data compounding. Every new Revolut product and market generates more labeled behavioral data — the specific kind of labeled behavioral data no competitor can purchase from a general-purpose AI vendor, because it belongs to Revolut’s customers and encodes Revolut-specific product structures. A general-purpose large language model, however capable, cannot be pre-trained on 26 million neobank customer histories spanning 111 countries. Revolut can.

Anton Repushko, who leads Revolut Research and presented the division’s findings at the International Conference on Machine Learning (ICML) in Seoul earlier in 2026, put the architectural thesis plainly. “Revolut Research has been established to responsibly build financial intelligence at its deepest layer, rather than patching together narrow, specialized models,” Repushko said at the launch. “In PRAGMA, we are developing a single, unified foundation model capable of understanding the true nuance of financial behavior in real time.”

Revolut is not the first neobank to publish a banking foundation model. Nubank published nuFormer, trained on more than 100 billion transactions across 100 million customers, and Mastercard released LTM, a foundation model oriented toward cyber-risk identification from card transaction data. What distinguishes Revolut’s approach in the published research is the multi-source architecture — PRAGMA ingests not only transactions but app navigation, support communications, trading activity, and static profile attributes simultaneously, rather than relying on a single event source. An independent analysis comparing Revolut to Nubank’s model describes PRAGMA as the more architecturally ambitious of the two.

The Documented Limitation

The PRAGMA paper is unusually candid about where the architecture falls short. Section 3.4.5, titled “Limitations in Highly Relational Tasks: Anti-Money Laundering,” reports that PRAGMA drops 47.1 percent on the F-0.5 metric for the AML downstream task evaluated in the study. The detailed results in the paper place this constraint in context alongside tasks where the model excels.

The reason is structural. PRAGMA is a per-user encoder: it processes each customer’s history in isolation and produces a representation for that individual. Anti-money laundering, by contrast, is a relational problem — detecting money-laundering networks requires identifying patterns that span multiple users and multiple accounts simultaneously. That is the domain of graph neural networks, not per-user encoders. The authors explicitly describe PRAGMA as a complement to graph-aware components for AML use cases, not a replacement for them.

Independent researcher Philipp Dubach, who published a detailed technical breakdown of the PRAGMA paper in June 2026, noted the same constraint: PRAGMA’s per-user encoding architecture is a property of the research backbone, not a statement about Revolut’s actual production AML stack, which is not described in the paper. For enterprise readers evaluating the architecture, the distinction matters: PRAGMA does not offer a single-model solution to AML, and Revolut has not claimed that it does.

What the Regulatory Clock Means

The Revolut Research announcement lands at a specific regulatory inflection point. The EU AI Act became generally applicable on August 2, 2026, meaning general transparency and registration obligations are now in force. However, the core high-risk obligations for credit-scoring AI — the explainability, human oversight, and conformity assessment requirements under Chapter III of the Act — were deferred to December 2, 2027 by Regulation (EU) 2026/1744, the Digital Omnibus on AI, adopted in July 2026. For Revolut, this means the compliance runway is real but not immediate: the architecture choices made now will need to meet those standards within roughly 16 months.

The Act requires, among other obligations, that high-risk AI systems deployed in credit decisions provide explanations sufficient for human review and for the affected individual to contest an automated outcome. GDPR Article 22 independently requires meaningful explanations for automated decisions affecting individuals. A compliance analysis published by Openlayer in July 2026 concluded that a system satisfying GDPR’s explainability requirement by providing a generic score rationale may still fail the EU AI Act’s human oversight standard under Article 14, if no qualified reviewer can actually intervene in the automated output before it propagates downstream.

Encoder-only models like PRAGMA produce dense vector embeddings — compressed numerical representations of a user’s financial history. Those embeddings are computationally powerful, but they do not naturally yield interpretable, per-feature explanations of the kind regulators demand. Techniques like SHAP values and attention attribution can be applied post-hoc, but the reliability of such attributions for deep encoder representations is an open research question. In the United States, the CFPB’s Circular 2026-03 separately requires that complex models used in lending decisions produce specific, reviewable adverse-action reasons.

Revolut Research’s commitment to publishing its methods and open-sourcing technical frameworks suggests the company understands that scientific credibility and regulatory trust will be as important as benchmark performance over time. Whether a billion-parameter encoder can satisfy regulators demanding legible decision trails is one of the harder engineering and policy questions in the space — and it is one the PRAGMA paper, by design, does not answer.

Revolut’s Expanding Data Surface

The strategic significance of PRAGMA scales with Revolut’s geographic footprint. The company received its full UK banking license from the Prudential Regulation Authority on March 11, 2026, unlocking consumer credit, mortgages, and FSCS deposit protection in its largest home market. Revolut filed for a US national bank charter with the Office of the Comptroller of the Currency and the FDIC on March 5, 2026, an application that — if approved, typically a 12 to 18 month process — would open all 50 US states to full banking services.

More credit products in more jurisdictions means more labeled behavioral data for PRAGMA — and a training signal that grows more valuable as it becomes more diverse. Revolut’s 2025 annual results, published in March 2026, reported overall revenue of £4.5 billion (approximately $6.1 billion), up 46 percent year-on-year, with profit before tax rising to £1.7 billion (approximately $2.3 billion). Subscription revenue — the segment most directly tied to AI-driven personalization and product recommendations — reached £708 million (approximately $966 million), a 67 percent increase year-on-year and the company’s fastest-growing major revenue line.

Revolut Research is scheduled to present further findings at Nvidia GTC Berlin in October 2026 and at the International Conference on AI in Finance (ICAIF) in November 2026, with quarterly community meetups planned at Revolut offices. The division also stated plans to publish research and open-source technical frameworks on an ongoing basis.

Exchange rate as of August 26, 2026; conversions are approximate. All pound-sterling figures converted at £1 = $1.3635 USD.

Frequently Asked Questions

What does PRAGMA actually do that existing fraud detection models cannot?

PRAGMA is a shared behavioral representation layer — a single pre-trained model that understands the full range of a customer’s financial activity rather than a narrow slice of it. Traditional fraud models are trained on transaction features alone; PRAGMA ingests transactions, app navigation, support interactions, trading activity, and static profile attributes simultaneously. That broader context helps it distinguish between, for example, a customer who is genuinely traveling internationally and one whose account is being accessed from abroad fraudulently, because it understands what that customer’s normal pattern looks like across every product dimension at once. The paper reports a 64.7 percent fraud recall improvement and a 16.7 percent improvement in fraud precision compared to Revolut’s internal task-specific baselines — meaning it catches significantly more fraud without generating proportionally more false alerts. Note that all figures are relative to internal baselines; absolute performance metrics are not publicly disclosed.

If PRAGMA is so effective at credit scoring, can Revolut actually use it to make credit decisions under EU law?

This is the open question the Revolut Research launch makes urgent. The EU AI Act classifies AI systems used for credit scoring in retail lending as high-risk under Annex III, requiring documented explainability, human oversight protocols, and post-market monitoring logs. The core compliance obligations under Chapter III of the Act — including these explainability and oversight requirements — were deferred to December 2, 2027 by the Digital Omnibus on AI regulation, giving deployers additional runway. However, GDPR Article 22 explainability duties remain in effect now, and an EU AI Act compliance analysis by Openlayer in July 2026 found that satisfying GDPR’s generic explanation standard may still fall short of the full human oversight requirement. The challenge with encoder-only models is that their outputs are dense numerical embeddings, not decision trees with readable split conditions. Post-hoc explanation techniques can be applied, but their reliability is not settled. Revolut’s stated commitment to open-sourcing technical frameworks and publishing research suggests the company is aware that regulatory trust, not just benchmark performance, determines whether PRAGMA can operate legally in EU credit markets at scale.

Does PRAGMA replace Revolut’s other AI tools, including AIR and its ChatGPT partnership?

No. PRAGMA is a backend infrastructure layer — an encoder backbone that produces behavioral representations used for risk, fraud, and personalization decisions. AIR (AI by Revolut), the in-app conversational assistant that launched in April 2026, is a different product built on a conversational architecture for user-facing tasks. Revolut’s partnership with OpenAI, which bundles ChatGPT Go access across Revolut subscription tiers, is a separate commercial arrangement with no data-sharing integration — users receive a redemption code for ChatGPT, not a connection between the two products. PRAGMA, AIR, and the ChatGPT bundle occupy different positions in Revolut’s AI stack and are not in competition with each other.

What is LoRA fine-tuning, and why does it matter for how quickly Revolut can deploy new AI features?

LoRA (Low-Rank Adaptation) is a technique that allows a large pre-trained model to be adapted to a new task by updating only a tiny fraction of its parameters — typically 2 to 4 percent — rather than retraining the entire model from scratch. The PRAGMA paper demonstrates that LoRA fine-tuning consistently matches or outperforms full-parameter training from scratch on a range of downstream tasks, suggesting that the pre-trained representations are general enough to transfer effectively across diverse banking use cases. For Revolut, this means that when a new fraud typology emerges, a new credit product launches in a new market, or a new personalization objective is defined, engineers can produce a task-specific adapter in as little as 12 hours — compared to the weeks or months that feature engineering for a dedicated specialist model would previously require.