Why Financial Institutions Are Converging on Transaction Foundation Models to Build Their Own Intelligence

The financial services industry stands at an inflection point in how it deploys artificial intelligence. Major institutions—from retail banks to payment processors—are abandoning fragmented, task-specific AI architectures in favour of unified transaction foundation models trained on proprietary transaction data. This structural shift represents a fundamental rethinking of how financial organisations can extract intelligence from their most valuable asset: billions of daily financial events. Revolut's recent development of PRAGMA, a family of transformer-based foundation models trained on 24 billion financial events across 26 million user records spanning over 100 countries, exemplifies this transformation. Built in collaboration with NVIDIA and powered by the company's full technology stack including Hopper GPUs and the cuDF library, PRAGMA demonstrates that a single foundation model can now outperform purpose-built task-specific models across multiple domains simultaneously—from credit scoring to fraud detection to product recommendations.

The context for this architectural revolution lies in the limitations of the incumbent model-building paradigm. Over the past fifteen years, financial institutions have constructed elaborate portfolios of statistical and machine learning models, each optimised for a specific business problem. A credit risk team would build its own models using tailored feature sets; a fraud prevention group would develop separate systems with different data pipelines; recommendation engines would operate in yet another silo. While individually effective, this approach created profound inefficiencies at scale. As enterprise transaction datasets have expanded exponentially—encompassing trillions of daily payments, transfers, interactions and behavioural signals—the gap between available data and what existing AI systems could actually process has widened dramatically. NVIDIA's 2026 State of AI in Financial Services report reveals that 65 percent of financial institutions now deploy AI in some form, with nearly 90 percent either actively deploying or formally assessing it. Yet beneath this apparent adoption lies a critical limitation: fragmented model architectures prevent institutions from developing truly unified understanding of customer financial behaviour. The industry recognises that this fragmentation represents the binding constraint on AI effectiveness, not computational power or data scarcity.

Transaction foundation models operate fundamentally differently from traditional financial AI systems. Rather than hand-engineering features for each specific prediction task, these models learn unified representations of customer behaviour by processing billions of raw financial events through transformer architectures originally developed for language processing. The PRAGMA case study provides concrete evidence of this approach's efficacy. The system was trained on 24 billion discrete financial events—payments, transfers, product interactions, and behavioural signals—across 26 million user accounts in more than 100 countries. Crucially, the single foundation model derived from this training data outperforms task-specific predecessors not just marginally but across multiple domains simultaneously. Tadas Kriščiūnas, head of group credit data science at Revolut, articulated a second critical metric: the collapse of feature engineering timelines. Where traditional machine learning pipelines required weeks or even months of manual feature engineering—the process of identifying, extracting and transforming raw data into inputs suitable for model training—the foundation model approach requires essentially no manual feature engineering at all. This distinction moves beyond mere productivity savings; it represents a fundamental shift in how financial institutions can translate data into actionable intelligence.

The practical implications for financial services organisations are substantial and immediate. Traditional fraud detection systems evaluate isolated signals: a high-value transaction, an unusual location, a new device. These models operate with limited context because adding more contextual variables exponentially complicates feature engineering and model training. Transaction foundation models invert this constraint by treating context as foundational. A payment at midnight carries different meaning when it represents the fourth transaction in ten minutes, originates from an unfamiliar device, occurs in a city where the customer has never previously transacted, and follows an extended period of inactivity. The model learns these contextual relationships directly from data rather than requiring analysts to hypothesise and encode them manually. This contextual reasoning improves performance across every downstream task—not through domain-specific optimisation but through deeper understanding of underlying customer behaviour patterns. For large institutions processing millions of daily transactions, even marginal improvements in fraud detection accuracy or credit assessment precision translate directly into measurable financial impact. A one-percentage-point improvement in fraud detection can prevent hundreds of millions in losses annually at systemically significant institutions. Similarly, more accurate credit scoring reduces both false positives that turn away qualified borrowers and false negatives that extend credit to risky ones.

The broader significance of this architectural shift extends far beyond individual institution efficiency gains. The convergence toward transaction foundation models reveals a maturing recognition that AI's bottleneck in financial services has never been algorithmic cleverness or processing power—it has been extracting maximum signal from proprietary data. This insight carries profound implications for competitive dynamics within financial services. Historically, advantages accrued to institutions that could hire the most skilled data scientists and machine learning engineers to hand-craft superior features and models. The foundation model approach partially democratises capability while simultaneously rewarding institutions with superior data. A bank with 100 years of transaction history and more customers now possesses a structural advantage more fundamental than engineering talent alone. The shift also reveals broader trends in enterprise AI deployment. The industry is moving away from the premise that different business problems require fundamentally different AI approaches—the assumption underlying the proliferation of siloed models. Instead, it is embracing the principle that unified representations of underlying phenomena enable superior performance across multiple applications. This principle, validated through transformer-based language models in the past five years, is now being proven in financial services specifically. The pattern suggests that mature AI applications across other industries—healthcare, logistics, manufacturing—may follow similar trajectories, moving from task-specific systems toward foundation models trained on comprehensive operational data.

Financial institutions and technology providers should monitor several critical developments in coming months and years. NVIDIA's new Build Your Own Transaction Foundation Model developer example, now available to enterprise teams, represents the first publicly accessible toolkit for institutions to begin constructing their own transaction foundation models. This democratisation of the technology will likely accelerate adoption across smaller and mid-sized institutions that previously lacked the internal capability to develop such systems independently. Second, the competitive response from other major financial services providers will prove instructive—whether other global payment processors and banks develop analogous foundation models, when they deploy them, and which institutions achieve meaningful competitive advantages through earlier adoption will shape the industry trajectory through 2027 and beyond. Third, regulatory clarity around foundation models in financial services remains incomplete. Supervisory bodies including the Federal Reserve, the Financial Conduct Authority and the European Central Bank are still formulating expectations for how financial institutions should validate, govern and explain decisions made by foundation models trained on transaction data. The resolution of these regulatory questions will either accelerate or constrain the pace of adoption. The convergence on transaction foundation models represents not a speculative future architecture but an emerging present, with leading institutions already demonstrating measurable performance gains. The question for the industry is no longer whether this architectural transition will occur but how quickly institutions can execute it and what competitive advantages accrue to early movers.