Imagine a bank that has observed billions of customer decisions over decades: How people save, borrow, invest, respond to nudges, and navigate financial shocks. That behavioral history is becoming more than a source of insight. Increasingly, it is becoming training data for a new generation of AI models.

Banks Are Beginning To Build Private Foundation Models

Banks are beginning to build private foundation models trained on proprietary financial data rather than relying only on general-purpose AI models. But the real opportunity is not simply building the model. It’s the proprietary behavioral and transactional data behind it, and the ability to connect customer behaviors to the actions they take and the outcomes that follow that competitors cannot easily replicate. Uncovering patterns and decision-making signals is a strong potential source of competitive advantage.

This reflects a broader trend that Forrester has described as the “private AI model explosion”. Across industries, organizations are increasingly training models on proprietary data and business knowledge. Healthcare providers are building models trained on clinical data, manufacturers on operational and sensor data, and retailers on purchasing behavior. Financial services is now following the same path.

This trend is accelerating. Revolut’s PRAGMA, developed with NVIDIA, is trained on 24 billion events across 111 countries and has demonstrated early results in areas like credit scoring, fraud detection, and product recommendations. Stripe introduced a foundation model for payment in 2025, Visa launched TransactionGPT, and NuBank’s nuFormer is live in Brazil. JPMorgan Chase, Mastercard, and PayPal are also investing heavily in proprietary AI capabilities and the data foundations required to support them.

As capable foundational models become increasingly commoditized, simply controlling the model provides little durable differentiation. Competitive advantage will shift to proprietary data, distribution, and outcome-based feedback loops that allow firms to learn, adapt, and improve those models faster than competitors.

From “Customer 360” To Behavioral Understanding

For more than a decade, financial institutions have invested heavily in creating a “customer 360” view: A unified profile of products, transactions, interactions, and demographics.

The next evolution is behavioral understanding. Instead of focusing primarily on who a customer is, behavioral intelligence focuses on how customers make decisions, why they make them, and what they are likely to do next. It captures spending habits, financial goals, responses to interventions, life events, and changing circumstances. This deeper understanding creates the foundation for a new generation of AI-driven financial experiences.

What Is A Private, Domain-Specific Model

A private, domain-specific model is trained on an organization’s proprietary data, business processes, and operational knowledge. Unlike general-purpose models designed to perform broad tasks across industries, these models develop deep understanding of a specific business context and the decision-making patterns that shape it.

Rather than relying exclusively on retrieval techniques to access enterprise knowledge at runtime, they can encode organizational knowledge directly within the model. The result is a differentiated source of intelligence that reflects an institution’s unique data, expertise, customer relationships, and operational history.

What Is A Transaction Foundation Model

A transaction foundation model is a specialized AI model trained on financial behavior rather than language, images, or code. Instead of learning to predict the next word, these models learn from transactions, customer interactions, product usage, financial events, and business outcomes to anticipate customer needs, identify risks, and uncover opportunities. By training on billions of financial events, Revolut’s PRAGMA creates a reusable representation of customer behavior that can support fraud detection, risk management, customer engagement, and product recommendations.

Transaction foundational models can be proprietary – developed by a financial institution using its own data – or provided by a third-party and trained on industry-wide datasets. PRAGMA provides an early example of the proprietary approach, while FICO’s transaction sequence model illustrates the vendor-led model.

Think of transaction foundation models as a behavioral intelligence layer for financial services. Historically, banks have relied of separate models for fraud detection, credit risk, marketing, customer analytics, and product recommendations. Transaction foundation models create the possibility of a shared intelligence layer that continuously learns from customer actions and outcomes across all of these domains.

Why It Matters

The greatest AI advantage in financial services may ultimately come from the intelligence behind the experience rather than the experience itself.

Many financial institutions are investing heavily in AI assistants, conversational banking, and AI agents. While these experiences may improve customer engagement, they are unlikely to deliver sustainable differentiation if competitors can access similar models, tools, and interfaces.

The stronger source of advantage may be proprietary behavioral intelligence: The unique data, connected between customer behavior, actions taken, and observed outcomes; and institutional knowledge that continuously improve decision-making over time. These feedback loops are particularly powerful in financial services because outcomes are already captured as part of everyday operations. Whether a customer repays a loan, defaults, accepts an offer, disputes a transaction, or changes a financial behavior, those outcomes create valuable training signals at relatively low cost. Over time, models trained on this data can become the intelligence layer powering financial assistants, advisory experiences, proactive support, and autonomous financial services – the engine of intelligent finance.

What Happens Next

Over the next several years, expect more banks, fintechs, payments providers, insurers, and wealth managers to explore behavioral foundation models trained on transactions, customer interactions, and financial events. Forrester data shows that 18% of financial services business and technology decision-makers plan to invest in foundational models over the next 12 months, indicating growing interest in this area.

For most institutions, however, the key strategic question will not be whether to build a model, but where to build and where to buy. Only a small number of organizations possess the combination of data scale, AI expertise, infrastructure, and capital needed to pre-train proprietary foundation models from scratch. Revolut can draw on years of transaction history across tens of millions of customers. Many institutions – whose customer data still arrives in overnight batch files – lack comparable data assets and will generate greater returns by fine-tuning existing models, enriching them with proprietary, or orchestrating multiple models. In practice, firms should build only where their data creates a meaningful advantage and acquire capabilities everywhere else.

That does not mean they cannot differentiate. As foundation models become more widely available, competitive advantage will shift from the models themselves to the proprietary feedback loops that make them smarter. The true asset is not the model weights, which can quickly depreciate as new base models emerge, but the record of customer behaviors, actions, and outcomes that no vendor can replicate. Defaults, repayments, chargebacks, offer acceptances, and other real-world outcomes create a continuous stream of learning signals that improve performance over time. The firms that pull ahead will be those that treat these feedback loops as a strategic asset, turning proprietary data, expertise, and customer trust into a unique intelligence layer that competitors cannot easily replicate.

Please feel free to schedule a guidance session or inquiry call with us.

Share