How to build Transaction Foundation Models in Banking and Payments

2026-06-19 · Source: Thoughtworks Insights · Field: Finance & Economics — FinTech & Digital Financial Services, Artificial Intelligence & Machine Learning, Banking & Financial Services · Depth: Advanced, long

Summary

Transaction Foundation Models (TFMs) are an emerging intelligence layer for banking and payments, designed to learn from transaction histories and generate reusable embeddings. Published on June 19, 2026, this guide outlines an end-to-end architecture for building TFMs, combining data engineering, transaction tokenization, transformer training, and integration. Instead of separate models for fraud, credit, or authorization, TFMs create a shared representation of transaction behavior, improving multiple downstream use cases. The architecture is not a one-off model but a reusable intelligence platform. Key steps include defining a measurable business problem, preparing and sequencing data, tokenizing transaction data, training a transformer backbone (e.g., using NeMo AutoModel), generating reusable embeddings (typically 128–768 dimensions), and evaluating against baselines. The NVIDIA Transaction Foundation Model developer example is referenced for implementation.

Key takeaway

For AI Architects or Directors of AI/ML evaluating unified intelligence layers in banking, Transaction Foundation Models offer a strategic shift. You should prioritize fraud detection as an initial use case to prove value, then extend to credit risk or authorization. Implement a staged approach: a 3-day concept phase, a 3-week prototype, and a 3-month MLP. This mitigates risk, accelerates deployment, and builds a proprietary intelligence layer.

Key insights

Transaction Foundation Models unify financial intelligence by learning from transaction sequences to generate reusable embeddings.

Principles

• Start with a measurable business problem.
• Design TFMs as reusable platforms.
• Tokenization is a critical design choice.

Method

Build TFMs by defining a use case, preparing and tokenizing data, training a transformer backbone, generating embeddings, and integrating into workflows.

In practice

• Use fraud prevention as a first TFM use case.
• Bucket amount fields using log-spaced thresholds.
• Store embeddings in a vector database.

Topics

• Transaction Foundation Models
• Banking and Payments
• Fraud Detection
• Transformer Models
• NVIDIA AI Platform
• Embeddings

Code references

• NVIDIA-AI-Blueprints/transaction-foundation-model

Best for: Machine Learning Engineer, AI Architect, Director of AI/ML

Related on AIssential

• The power of Transaction Foundation Models: Building the unified intelligence layer for payments — Transaction Foundation Models learn the "language of money movement" from large-scale financial data to create reusable intelligence across banking and payments.
• Relational Foundation Models for Enterprise Data [Jure Leskovec] - 768 — Relational Foundation Models leverage graph neural networks for in-context learning on raw multi-tabular data, enabling accurate predictions without explicit training.
• The Silicon Soil — AI and real-time data can transform banking from land-dependent to a dynamic wealth orchestrator, mitigating systemic risk.
• AI Hub — Domain-specific AI models are crucial for production systems in specialized industries like banking.
• Every Infrastructure Revolution Has a Missing Layer. AI Agents Just Found Theirs. — AI agents require a purpose-built, non-human payment infrastructure to overcome identity, currency, and scope limitations of traditional finance.
• Unlocking the relational power of graph neural networks in fraud prevention — Graph Neural Networks enhance fraud detection by modeling relational data, uncovering networked fraud invisible to traditional tabular methods.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Thoughtworks Insights.