A transaction foundation model is an AI model trained on large volumes of transaction data so one model can serve many tasks, such as fraud detection, credit risk and product recommendations. Until recently, banks built a separate model for each task. Now some are building one that all of them share.
For a founder selling into banks, that changes the first question. It is no longer only “is my model better?” It is “does this bank need a model from me at all?”
What a transaction foundation model does
Revolut and Starling both announced one in recent weeks.
- Revolut’s PRAGMA reads each customer’s history as a sequence of events, much as a language model reads text. According to NVIDIA’s case study, it was pre-trained on about 26 million user records across 111 countries, covering tens of billions of events, and it replaces separate pipelines for fraud, credit, marketing and recommendations.
- Starling’s model is being trained on more than 15 billion anonymised transactions from its own data. Starling says it will form a single layer, “Starling Intelligence”, and plans to put it into production in early 2027, first for fraud detection and then for its AI assistant.
The attraction is reuse. One shared layer replaces many separate builds, and a new use case starts from the same foundation instead of from scratch.
Why banks are building their own
Four reasons come through in the announcements.
- Control of data. Starling trains on its own data, on a UK sovereign cloud, so the data stays in UK jurisdiction.
- Supplier concentration. Starling says its choice follows the UK Government’s Financial Services AI Adoption Plan (July 2026), which flags reliance on a small number of global AI and cloud providers.
- Speed. Revolut reports model development cycles three to five times faster. That is Revolut’s own figure, published by NVIDIA.
- Cost. NVIDIA reports that bunq’s assistant Finn, running on open-weight models, matches leading closed models at 60% of the cost per answer.
Treat the performance figures as claims until you see how they were measured. Revolut’s reported gains, such as a 64.7% improvement in fraud recall, are comparisons against its own previous models.
The Own, Adopt or Sit Beside test for a transaction foundation model
Banks are taking three routes to a transaction foundation model. Each leaves a different space for a startup.
Own
The bank trains a model on its own data. Revolut, Starling and Lloyds (with its ATUM model, per NVIDIA) are examples. A startup selling “a better model” on the same data faces a hard pitch, because the bank holds the data and the team to use it.
What can still work: tooling around the model, such as monitoring, validation, review interfaces and audit trails.
Adopt
The bank takes an open-weight model and tunes it. NVIDIA names bunq, Santander (an early adopter of its Kumo models for structured financial data) and Modulr, which is building specialised models on an open Nemotron model.
What can still work: specialised models and data for a narrow task, plus the integration work that makes an open model usable inside a bank.
Sit beside
The startup does not compete on the model. It builds what surrounds it: workflow, evidence, review and accountability. These are the problems we covered in AI financial crime investigation, where the question was who signs off, not which model is faster.
What can still work: most things a small team can realistically ship.
How to use the test
Ask four questions before you build.
- Does my buyer build or adopt? The answer decides your pitch. Ask early.
- Is my edge the model, or what surrounds it? If the model is the edge, check whether the bank could train its own on data you will never see.
- Where does the data have to run? If a buyer insists on a particular jurisdiction, find out before you design your hosting.
- Can a reviewer see why? Whatever the model, an analyst or auditor must be able to follow the reasoning. Products that make this easy are easier to sell.
Where to start
You do not need your own transaction foundation model to build in this space. Narrower starting points are easier to test:
- Review and explanation tools that show why a model-driven alert fired.
- Validation and monitoring for banks tuning models on their own data.
- Workflow tools that connect a model’s output to the action a person takes.
Pick one and speak to five people who work in it. Our guide to customer discovery for an AI startup covers how, and identifying the right buyer shows how to find who signs. For wider context on where the large players are heading, see our analysis of JPMorgan vs Revolut.
Frequently Asked Questions
It is an AI model pre-trained on large volumes of transaction data so one model can support many tasks, such as fraud detection, credit risk and recommendations.
Revolut (PRAGMA) and Starling are building their own, and NVIDIA says Lloyds has built ATUM. Starling’s model is not yet in production; it is planned for early 2027.
No. It means a startup selling a model on the bank’s own data faces harder competition. Workflow, evidence, review and accountability tools sit around the model and remain open.
They are company-reported. Revolut’s figures are internal comparisons against its previous models, and the cost figures for bunq are NVIDIA’s. Ask how any figure was measured.
Some institutions, such as Starling, want data and model to stay in their own jurisdiction. If your product handles bank data, find out where it must run before you design it.
