Unified Transaction Models with AI: The End of Banking Silos

What happened?

According to the NVIDIA blog, financial institutions are moving away from task-specific AI models (fraud, credit, recommendation, risk) in favor of transaction foundation models. These models — based on transformer architectures — are trained on trillions of financial events (payments, transfers, product interactions, behavioral signals) using exclusively proprietary data. The result is a unified representation of consumer behavior that can be applied across multiple domains.

This shift is not sudden. For years, banks built separate models for each function, creating siloed systems that prevented a holistic view of the customer. As NVIDIA notes in its blog, "siloed systems prevent institutions from developing a unified understanding of consumer financial behavior." With the growth of enterprise datasets, the gap between what institutions know and what their AI can reason about has widened, creating an opportunity to build intelligence using proprietary data.

A landmark case is PRAGMA, a model developed by Revolut in collaboration with NVIDIA, trained on 24 billion events from 26 million users across over 100 countries. PRAGMA uses the full NVIDIA stack (Hopper GPUs, cuDF, Nemotron) on the Nebius cloud, and already outperforms specific models in tasks such as credit scoring. This model represents a milestone: it is the first large-scale transaction foundation model trained on real, heterogeneous financial data.

Why is it important?

The 2026 State of AI in Financial Services report by NVIDIA indicates that 65% of institutions already use AI, and nearly 90% are implementing or evaluating it. Moreover, almost all maintain or increase their AI spending. However, model fragmentation limits the ability to obtain a holistic view of the customer. A unified foundation model allows interpreting the full context of a transaction (time, device, location, history) rather than isolated signals, improving performance across all tasks.

For example, a midnight payment takes on different meaning if it is the fourth transaction in 10 minutes, from an unknown device, and in a new city. This contextual depth, previously inaccessible, is now possible thanks to transformers applied to tabular data. NVIDIA highlights that "a traditional fraud model evaluates isolated signals; a foundation model interprets behavior in context where time, device, location, and prior activity shape meaning."

This advancement is comparable to the leap from statistical models to decision trees in the 1990s, or from simple neural networks to deep learning in the 2010s. But here the difference is that it applies to tabular data, which constitutes the majority of financial data, not just text or images.

Consequences for the sector

• Operational efficiency: A single model replaces multiple specific models, reducing maintenance costs and complexity. According to NVIDIA estimates, institutions adopting foundation models can reduce model development time by 40% and infrastructure costs by 30%.
• Improved fraud detection: By analyzing complete behavior, false positives are reduced and complex patterns are identified. For example, PRAGMA has demonstrated a 20% reduction in false positives in fraud detection compared to previous models.
• Personalization: Institutions can offer products and services tailored to the unified customer profile, increasing retention and customer lifetime value.
• Competitive advantage: Those who adopt these models can innovate faster and with greater precision. Revolut, for instance, is already using PRAGMA to improve its product recommendation engine, which has increased conversion rates by 15%.
• Privacy challenges: The massive use of proprietary data requires robust governance frameworks and regulatory compliance, such as GDPR in Europe or CCPA in California. Institutions must implement techniques like differential privacy and federated learning to mitigate risks.

What readers should know

Transaction foundation models represent a paradigm shift: we move from siloed AI to integrated intelligence. Institutions already investing in this technology (like Revolut) will gain a significant advantage. For startups and fintechs, the barrier to entry is high (requires large data volumes and compute capacity), but collaborations with cloud and GPU providers can democratize access. NVIDIA offers its platform and tools like cuDF and Nemotron to accelerate development.

The market should prepare for accelerated adoption, especially in banking, insurance, and fintech. The key will be the quality and exclusivity of proprietary data, and the ability to integrate these models into production workflows. According to the NVIDIA report, it is expected that by 2027 more than 40% of financial institutions will have implemented transaction foundation models.

Moreover, this movement is not isolated: large tech companies like Google and Amazon are already developing similar models for their financial services. Competition will intensify, and regulators will need to adapt to ensure fairness and transparency.

“A transaction foundation model not only improves fraud or credit; it changes how institutions understand their customers.” — TheVortiq

In summary, we are facing a transformation that will redefine financial intelligence in the coming years. Institutions that act now can capitalize on this advantage, while those that lag will face a growing competitive disadvantage.