The Short Answer
Revolut launched a dedicated AI research lab on 25 August 2026 built around PRAGMA, a family of foundation models pre-trained on its own banking event sequences, with reported gains of 2.3x in credit default risk identification and 65% more fraud caught. Almost no other institution has the corpus or the lab to copy that. On ibl.ai you own all the code and the data, which delivers the same independence without pre-training a model from scratch.
The strategic lesson is not "build a foundation model." It is that Revolut refused to rent the layer where its advantage lives — and that refusal is available to institutions that will never train one.
What is Revolut's PRAGMA model?
PRAGMA is a family of foundation models for banking event sequences, published by Revolut and now the centrepiece of its research division.
Architecturally it is encoder-style Transformer models trained with masked modelling — self-supervised learning over sequences of banking events rather than over natural language.
The paper reports that it supports downstream tasks including credit scoring, fraud detection, and lifetime value prediction.
One detail in the abstract is more instructive than the headline results: strong downstream performance was achievable by training a simple linear model on top of the extracted embeddings. The heavy lifting is in learning a good representation of the institution's own event data, not in elaborate task-specific modelling.
Revolut Research launched on 25 August 2026 around this model, with reported early results on historical data of a 2.3x uplift in identifying credit default risk, 65% more fraud cases caught, and 41% more relevant product recommendations, and with NVIDIA as a partner on continued development.
Why can't most financial institutions copy this?
Most institutions cannot copy Revolut's approach because pre-training a foundation model requires a proprietary corpus and a research organization, and the corpus is the harder of the two to acquire.
Revolut trained on banking event sequences drawn from roughly 25 million users across 111 countries. That volume of first-party behavioural data is the actual asset; the architecture is comparatively ordinary.
A regional bank, a credit union, or a mid-size insurer does not have a corpus of that scale and cannot buy one, because the whole point is that it is proprietary.
The second requirement compounds it. Standing up a research lab, retaining the people, and running the compute is a sustained operating commitment, not a project — and one that competes with every other technology investment on the roadmap.
So "become an AI company" is a correct description of what Revolut did and an unhelpful instruction for almost everyone reading about it.
What is the wrong conclusion to draw from Revolut's AI lab?
The wrong conclusion is that institutions unable to train a foundation model should therefore rent one through a per-seat assistant.
That reasoning inverts the lesson. Revolut's move was a refusal to depend on an outside vendor for the capability closest to its competitive advantage.
Responding to it by handing your transaction data to a vendor-hosted assistant adopts precisely the dependency Revolut spent a research lab avoiding.
The cost shape makes it worse at scale.
Per-seat licensing at roughly $30–60 per user per month bills every employee regardless of usage, so a 5,000-person institution pays $150,000–300,000 a month for a system whose model it cannot inspect, cannot host, and cannot move when the contract renews.
For regulated financial workloads there is a further problem: the data that makes the model useful is the data least permitted to leave the perimeter.
What is the third path between training a model and renting one?
The third path is to own the platform rather than the frontier model — keeping the data and the stack, while treating the model itself as a swappable component.
Own the platform layer. The agents, memory, retrieval, identity, audit logging, and governance run on infrastructure you operate, under a licence you hold. This is the layer where institutional knowledge accumulates, and it is the layer per-seat products never transfer.
Stay model-agnostic. Run open weights self-hosted, or a commercial API, and change your mind without rebuilding. Model leadership has traded hands repeatedly in the last year; a platform that treats the model as configuration converts that churn from a migration risk into a routing decision.
Let your proprietary data differentiate through retrieval and fine-tuning. You do not need to pre-train from scratch to get value from data nobody else has. Retrieval over your own systems and targeted fine-tuning capture most of the advantage at a fraction of the cost — and PRAGMA's own finding, that a linear model over good embeddings of your data goes a long way, is the argument for it.
Keep the data inside the perimeter. For financial services this is usually the binding constraint rather than a preference, and self-hosting resolves it structurally instead of contractually.
On ibl.ai you own all the code and the data, run it model-agnostic across any LLM, and pay with no per-seat pricing — so you can deploy anywhere, from your own cloud to a fully air-gapped network. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.
ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on. For the operating-model view of this shift, see banks building AI workforces on infrastructure they own.
Frequently asked questions
Is PRAGMA an LLM?
Not in the usual sense. It is an encoder-style Transformer trained on banking event sequences rather than natural language text, which is why its downstream tasks are credit scoring, fraud detection, and lifetime value prediction rather than chat.
Does an institution need its own model to get differentiated results?
No. Most of the differentiation comes from proprietary data being reachable at inference time. Retrieval over your own systems and targeted fine-tuning capture a large share of that advantage without a pre-training program.
What is the minimum requirement for keeping financial data inside the perimeter?
That the model runs on infrastructure you operate. Once inference happens inside your network, the transit question that dominates vendor security review disappears rather than being mitigated.
The bottom line
Revolut's lab is a real signal, and the reported gains are large enough to take seriously. But the transferable part is not the foundation model.
It is that the institution decided the layer nearest its advantage was not something to rent. That decision is available to organizations that will never train a model of their own — it just gets implemented as owning the platform rather than owning the weights.