Revolut, the UK-based neobank serving 45 million users globally, published research in April 2026 introducing PRAGMA — a Transformer-based foundation model trained on 24 billion real banking events drawn from 26 million users. It is the first publicly documented banking foundation model at this scale, and it outperforms hand-engineered models on fraud detection, churn prediction, and product personalization tasks.
This is not an incremental scoring improvement. It is a structural shift in how AI-native financial platforms can model behavior at population scale.
What PRAGMA Actually Is
Banking AI has historically meant a collection of narrow, isolated classifiers — one for fraud, one for churn, one for credit. Each is hand-tuned, feature-engineered, and siloed. PRAGMA replaces that collection with a single Transformer foundation model that ingests raw behavioral data and produces high-quality embeddings any downstream task can consume.
The output is not a fraud score or a churn probability. It is a dense vector representation of a user’s full financial behavior — a compact encoding of transaction history, app engagement, account state, and spending context. Downstream models then operate on these embeddings rather than hand-crafted feature matrices built by analysts who can only capture what they already know to look for.
The analogy is BERT’s arrival in NLP in 2018. Pre-trained representations replaced task-specific feature engineering across nearly every language task within two years. PRAGMA attempts the same substitution for financial behavior modeling.
The Training Data: 24 Billion Events, 26 Million Users
PRAGMA’s training corpus is, by any published standard, exceptional. Revolut trained the model on 24 billion discrete events — transactions, app interactions, profile state changes — from 26 million users. No traditional bank has published a behavioral model trained at this scope.
The input modalities span the full customer lifecycle: purchase transactions with merchant identity, amount, category, and timestamp; internal transfers; foreign exchange conversions; card usage patterns; app engagement signals; and account lifecycle events. Revolut describes the training data as raw — minimal preprocessing, no hand-engineered feature extraction at ingestion.
This breadth matters because financial behavior is deeply contextual. A £200 transaction at 2am carries different signal than a £200 subscription renewal at the same hour from the same user. PRAGMA learns these contextual dependencies from data rather than relying on analysts to encode them as explicit rules that bad actors eventually learn to route around.
Why the Transformer Architecture
The Transformer’s multi-head attention mechanism is naturally suited to sequential financial data. A customer’s transaction history is a temporal sequence, and the significance of any single event depends on the full context of what preceded it — exactly the problem attention was designed to solve.
Revolut trained PRAGMA as a sequence model over temporally ordered behavioral events, treating each event as a token the way a language model treats words. Meaning emerges from learned relationships across the full sequence rather than from hand-specified lookback windows.
Earlier approaches used recurrent networks — LSTMs and GRUs — for the same sequential structure. Transformers offer two concrete advantages: they parallelize over sequence length during training, substantially reducing compute time on 24 billion events; and their attention weights are interpretable, making it easier to satisfy regulators who require explanations for automated credit and fraud decisions.
Downstream Performance: Fraud, Churn, and Cold-Start Recommendations
Revolut benchmarks PRAGMA-derived embeddings against hand-engineered feature baselines across three production tasks: fraud detection, customer churn prediction, and product recommendation.
Fraud detection saw the largest reported lift. The foundation model captures cross-merchant behavioral patterns and subtle timing anomalies that explicit feature engineering systematically misses, because no analyst foresaw them as features worth encoding. Churn models using PRAGMA embeddings improved accuracy on early-stage disengagement — the hardest case, because early churn signals are sparse and non-obvious to rule-based systems. Recommendation quality improved most sharply on cold-start users, where collaborative filtering fails by definition because there is no prior interaction history to filter from.
The paper’s central argument is that a single well-trained foundation model producing general-purpose behavioral representations consistently outperforms the collection of narrow models it replaces — with lower ongoing engineering overhead after the initial training investment.
Why Traditional Banks Are Structurally Disadvantaged
Replicating PRAGMA requires two things most incumbent banks structurally lack: unified behavioral data at scale and ML infrastructure velocity.
Unified data is the harder problem. Revolut processes all transactions for its entire user base on a single, coherent platform — no product silos, no legacy subsidiaries with incompatible schemas. A UK high-street bank may have comparable or larger customer volumes, but its data fragments across mortgage origination systems, current account platforms, credit card processors, and wealth management suites built across decades of mergers and acquisitions. Training a Transformer foundation model requires a clean, unified behavioral graph. Most incumbents cannot produce one without a multi-year data engineering programme that itself requires cultural and organizational changes most institutions are not structured to execute.
Infrastructure velocity is the second barrier. Training at 24 billion events requires compute infrastructure and ML engineering culture that most banks are still assembling. Infrastructure investment at the scale of Nebius’s planned €10 billion European AI data center illustrates where the compute arms race is heading — fintech-native firms start far closer to that operational model than their legacy competitors.
The compliance argument against advanced ML in banking is real but frequently overstated as a blocker. EU and UK regulators have consistently approved sophisticated ML in financial risk management provided models meet documentation and explainability requirements. The actual structural blocker at most incumbent banks is data unification, not regulatory caution.
What Revolut Released — and What It Kept
Revolut published the PRAGMA research paper publicly, documenting the architecture, training methodology, data schema design, and downstream evaluation results. This is a meaningful contribution: no other major neobank has published a foundation model paper at this scope.
What the release does not include: the trained model weights or the proprietary 24-billion-event training dataset. The open contribution is methodological — a blueprint that institutions with sufficient data could replicate, but Revolut’s specific trained model remains internal.
This is the standard competitive disclosure playbook for AI research. Anthropic’s accidental source code release earlier this year illustrated how wide the gap between documented methodology and a runnable, deployed system can be. The PRAGMA paper describes what the model does. The weights encode what it learned from 24 billion real financial events — and those stay at Revolut.
What PRAGMA Changes for the Industry
PRAGMA establishes a concrete benchmark for what an AI-native bank can build when it controls behavioral data at scale. It makes the foundation model approach to financial modeling credible in the industry — not a research proposal but a deployed system with measured improvements on production tasks.
For the broader AI ecosystem, this is another data point in the shift from general-purpose pre-training toward domain-specific foundation models. MegaOne AI tracks 139+ AI tools across 17 categories, and the pattern is consistent: models pre-trained on domain-specific corpora consistently outperform general models fine-tuned on small in-domain datasets. PRAGMA applies this principle to the most data-dense domain in consumer finance.
For bank employees in fraud operations and manual risk review, the trajectory documented in papers like this is unambiguous. The displacement of human judgment by automated AI systems in financial services is no longer a hypothetical — it is a roadmap being published in peer-reviewed research.
For Revolut, PRAGMA is foundational infrastructure. Every high-stakes product decision — who receives a credit offer, which transaction flags for human review, who receives a personalized product nudge — now routes through a model that has observed more banking behavior than any publicly documented system.
Banks that cannot build an equivalent will eventually source behavioral intelligence from platforms that can. That acquisition market is forming now — and Revolut just published its price of entry.
