Closing the context gap is what banking AI needs

Every bank at Sibos 2026 could point to an AI pilot. Fewer could point to one running in production. That gap between the pilots and production, rather than model quality, was the story that played out in Miami Beach, where Swift built its 2026 conference program around the theme "Digital finance for AI-driven economies." Thoughtworks, AWS and NVIDIA hosted an exclusive executive breakfast on the sidelines of Sibos, bringing together senior leaders across transaction banking, payments and technology.

Banking and payments leaders were on the fence about whether AI is delivering on its promise. Many had run experiments, but few had anything load-bearing in production. The instinct is to blame the model. The more useful diagnosis, offered at that breakfast, pointed to a context gap instead: AI usually isn't failing because it's wrong. It's failing because the context it needs is locked behind a legacy system, disconnected across the systems that hold it, or, most often, available and connected but still not understood.

General-purpose AI models don’t understand transactions

A general-purpose AI model is fluent in language. It has no native grasp of what a transaction, a merchant relationship or a fraud ring means. "It understands English," as Brian Blanchard, Thoughtworks' global VP of cloud services, put it at the breakfast. "It doesn't understand fraud or credit card transactions and the AI model needs context."

Transaction foundation models (TFMs) are built to close that gap and add the needed context. They are AI models trained specifically to reason across transaction events (such as timing, velocity, merchant category, and device history) the way a language model reasons across words. Mastercard's Large Transaction Model, announced at GTC 2026 with NVIDIA and Databricks, is already framed as an insights engine spanning fraud, cybersecurity and personalization. Research on Revolut’s PRAGMA model reports a 130.2% improvement in credit scoring and a 64.7% improvement in external fraud recall using a single pre-trained backbone, trained on 26 million user records and 24 billion events. Meanwhile, Stripe Radar, which draws on more than $1.9 trillion in annual payment volume, cuts fraud by 32% on average. 

Rebuilding context for every question drives up AI cost

Asking an AI model a question is the easy part, according to Pahal Patangia, global payments industry leader at NVIDIA. "The most expensive part of AI is every time you ask a question, you [have to] get the data to answer the question."

Picture 20 teams inside one bank, each running its own pilot, each re-uploading the same customer and transaction data, each burning budget to reconstruct the context the team next door already built. This is an architecture problem, and it drives up cost and inconsistency across the industry right now, long before anyone gets to a conversation about accuracy.

The fix several institutions are converging on is: one shared intelligence layer, trained once on an institution's own transaction history. Fraud, credit risk, payment routing and customer experience teams all draw from, instead of each rebuilding context from scratch.

A hybrid model strategy balances speed, cost and quality

The panel was candid that institutions rarely get speed, cost and quality all at once from a single model choice. Several panelists described converging on a hybrid approach. Closed, frontier-grade models for orchestrating complex reasoning, paired with smaller, open, fine-tuned models for narrow, repeatable tasks kept close to where the sensitive data lives. As Pahal Patangia summarized it, "the best of the best enterprises" treat that mix as a deliberate strategy.

Institutions that build one shared intelligence layer across fraud, credit risk and payments need governance in place from day one.

That means three things: a running record of what a model was asked and what it answered, so drift gets caught early; clear limits on what data any model or agent can touch; and the ability to reconstruct, for a regulator, exactly how a decision was reached.

The tipping point is a “number” on a loss report

None of this happens on the strength of a good demo. As Paul Chang, head of payment networks at AWS described it at the breakfast, the moment a use case moves from proof of concept to enterprise deployment is usually driven by a financial loss. The bank realises it "lost this much from this one use case" and concludes, belatedly, "we should have caught it." The number on the loss report finally justifies the investment.

The path forward

The institutions pulling ahead on AI in 2026 have stopped treating every fraud check, credit decision and routing choice as its own isolated model. They treat  them as different questions asked of the same shared understanding of their business.

Teams ready to test a transaction foundation model can start with NVIDIA's Build Your Own Transaction Foundation Model developer blueprint. For the architecture behind it, read The power of transaction foundation models, where Thoughtworks and NVIDIA for deeper.