How banks and insurers create their own intelligence with transaction foundation models

Banks and insurance companies are converging on a new artificial intelligence paradigm: the transaction foundation models (TFMs). Instead of maintaining dozens of isolated models for fraud detection, credit analysis, and risk assessment, these institutions are adopting a single central model trained in all transactions — creating their own unified intelligence. NVIDIA, in an analysis published on its blog, details how this transformation is redefining the financial sector.

What are transaction foundation models?

A transaction foundation model is a large pre-trained AI model specifically for transactional financial data. Unlike generic models such as GPT or BERT, which understand language, TFMs learn patterns, sequences, and anomalies inherent to the universe of banking, insurance, and investment transactions. They function as a single base that can be adjusted (fine-tuning) for distinct tasks — from credit approval to money laundering prevention.

How do foundation models replace fragmented systems?

Traditionally, each area of a financial institution builds its own models: one for credit card fraud, another for default risk, another for compliance. This generates data silos, rework and maintenance difficulties. With the TFMs, the whole organization starts to share the same understanding about transactional behavior. NVIDIA points out that this approach reduces the development time of new models by up to 80% and improves accuracy in tasks such as real-time fraud detection.

Why do foundation models matter to Brazilian institutions?

The Brazilian financial market is one of the most digitized in the world, with high adoption of instant payments such as Pix and accelerated growth of fintechs. This generates a huge volume of transactions — and also opportunities for fraud and risks. Adopting transaction foundation models allows local banks and insurers to create smarter defenses without relying solely on imported solutions. As Microsoft highlighted in recent analysis, only AI will not transform your business; the system that executes it will. FMTs are precisely this systemic basis.

How does NVIDIA drive the foundation models?

NVIDIA offers hardware infrastructure (GPUs) and software (platforms such as NVIDIA AI Enterprise) that allow you to train and deploy TFMs on a scale. In the blog, the company details cases of use with customers who already use these models to unify credit risk analysis and anomaly detection — with reduced operating costs and increased efficiency. The promise is that soon any financial institution can have its own foundation model, as specific as its client portfolio.

Frequently Asked Questions about Foundation Templates

What is the difference between a transaction foundation model and a traditional machine learning model?

Traditional models are trained for a single task with specific data; the foundation models are pre-trained in large volumes of transactional data and can be adapted for multiple tasks with much less data and effort.

Do transaction foundation models require many computational resources?

Yes, initial training requires powerful clusters of GPUs, such as those offered by NVIDIA. However, the fine-tuning phase is much lighter and can run on smaller servers, making the total cost feasible for mediums and large institutions.

Can these models be used by fintechs or only by large banks?

Any institution with a significant volume of transactions can benefit. NVIDIA states that technology is increasingly accessible, including through cloud services that offer pre-trained models ready for customization.

Read the full review on the official NVIDIA blog: Finance Institutions Convergence on Transaction Foundation Models.