Financial Firms Move to Transformer-Based Transaction Foundation Models to Centralize Intelligence

Banks, payments firms and fintechs are replacing siloed, task-specific models with transformer — based transaction foundation models trained on billions of events.

Major financial institutions are shifting from many fragmented, task-specific systems to transformer — based transaction foundation models to centralize intelligence and cut duplication, according to a 2026 industry report. The report found 65% of institutions now use AI, nearly 90% are deploying or assessing it, and almost all are maintaining or increasing AI budgets — trends that amplify the operational and cost pressures of siloed model architectures as enterprise datasets grow.

Transaction foundation models are transformer systems trained on billions of financial events — payments, transfers, product interactions and behavioral signals — that produce unified embeddings of consumer behavior. By modeling sequences rather than isolated events, these systems interpret context such as timing, device and location so that the same signal can mean different things when seen in sequence (for example, many quick payments from a new device versus the same payments from an established device).

Revolut’s PRAGMA family provides a production example: transformer — based models trained on 24 billion events across 26 million user records from more than 100 countries. Running on a full accelerated AI stack and cloud infrastructure, a single foundation model reportedly outperformed strong task-specific systems across credit scoring, fraud detection and recommendations, demonstrating cross — domain gains from shared behavioral representations.

Vendors are packaging developer tooling and turnkey examples to accelerate adoption. One vendor released a Build Your Own Transaction Foundation Model developer example to help teams produce transformer embeddings from tabular transaction data and integrate them into existing pipelines without rebuilding whole systems. Another large payments firm is developing a proprietary large tabular foundation model trained on billions of anonymized transactions and designed to scale toward hundreds of billions, built with major cloud and platform partners and using automated model tooling.

Industry users say the approach addresses the cost of fragmentation: each new use case or market previously required bespoke models and frequent retraining. Early tests indicate foundation models can reduce reliance on handcrafted features and dozens of task-specific systems, delivering better cross — domain performance and enabling reuse of shared embeddings across fraud, loyalty, personalization, portfolio optimization and analytics.

For engineering teams the practical implications are concrete: feature engineering that used to take weeks or months can be shortened by reusing transformer — derived embeddings; teams can integrate transformer workflows into existing data pipelines and cloud environments; and deployments can rely on standard accelerated compute and open-model components. Demonstrations from established firms suggest these benefits are already producing measurable gains while lowering the maintenance burden of many narrowly focused models.