FF News — The Fintech News Network

BBVA Leverages Transformer AI Models to Revolutionize Mobile Banking Personalization

By Lauren Towner · 19 November 2026

In brief: BBVA is using Transformer AI models to analyze 1.6 billion mobile app sessions, improving its ability to predict customer needs and product adoption by 2.7% for 8 million users.

Key facts
Customer Base8 million active users
Data Volume1.6 billion navigation sessions
Daily Processing4 million sessions per day
Performance Lift2.7% relative increase in AUC metric
Press Release: BBVA Leverages Transformer AI Models to Revolutionize Mobile Banking Personalization | Featured Image by FF News

BBVA has unveiled a new AI framework that applies Transformer architectures—the technology underpinning large language models—to analyze customer navigation within its mobile app. By treating digital interactions as sequential "sentences," the bank is improving its ability to predict customer needs and personalize services at a scale of millions of daily sessions.

What was announced

BBVA researchers presented a scientific paper at the ACM International Conference on AI in Finance (ICAIF) in Milan on November 15. The research details a technical shift from traditional Word2Vec embeddings to Transformer-based models to represent digital navigation. While the bank has utilized numerical representations (embeddings) to identify data patterns since 2018, this new approach treats each screen visited as a "token" and a full session as a "sentence," preserving the context and sequence of interactions that previous models often lost.

The initiative analyzed approximately 1.6 billion navigation sessions from eight million customers over a one-year period using anonymized data. From this pool, 90 million sessions were selected to train the new models. To handle the app's constant evolution, researchers implemented a hierarchical tokenization system that groups related screens, allowing the model to interpret new features based on existing categories without requiring a full system rebuild whenever the interface changes.

In real-world testing, the Transformer-based model was used to estimate the likelihood of a customer taking out a financial product. The results showed a 2.7% relative increase in the AUC metric—a measure of predictive accuracy—compared to the previous baseline model. Currently, the solution is integrated into a production environment serving eight million active users and processing four million sessions per day, meeting internal requirements for efficiency, governance, and reliability.

The companies involved

BBVA (Banco Bilbao Vizcaya Argentaria) is a global financial services group with a significant presence in Spain, Mexico, South America, and Turkey. The bank has positioned itself as a digital leader in the Eurozone, often pioneering the use of emerging technologies in retail and corporate banking. Unlike many traditional peers, BBVA has maintained a dedicated research arm for nearly a decade to explore advanced data science and artificial intelligence, moving beyond simple automation into deep learning and foundation models.

The bank’s strategy has increasingly focused on a "100% digital" model in expansion markets, while leveraging its massive existing data sets in established regions to drive operational efficiency. By moving toward foundation models, the institution is attempting to consolidate disparate data streams—including transactions, account activity, and human interactions from branch visits or phone calls—into a unified understanding of customer behavior. This technical infrastructure is designed to support a new generation of AI agents and proactive banking services, moving the institution beyond reactive transaction processing into a more predictive, data-driven relationship model.

What FF News has reported before

FF News has tracked BBVA’s aggressive digital expansion and its frequent collaborations with technology leaders. Recently, the bank achieved a significant milestone by reaching one million customers in Italy through its digital-only model. Its commitment to high-tech infrastructure was further evidenced by a partnership with NVIDIA to secure autonomous AI agents, a move that aligns with the research presented at ICAIF.

Beyond AI, the bank has been active in telecommunications and blockchain. It recently launched integrated mobile eSIM services powered by Gigs. Furthermore, BBVA participated in the creation of the world’s first end-to-end blockchain syndicated loan alongside Iberdrola and Santander, demonstrating its willingness to apply distributed ledger technology to complex corporate finance structures.

What this means

This announcement signals a shift in how tier-one banks view customer data. Moving from static segmentation to sequential, Transformer-based analysis puts pressure on mid-tier banks that lack the internal research capacity to build proprietary foundation models. While many institutions use AI for fraud detection, BBVA is applying it to the very logic of the user interface, turning the mobile app into a continuous feedback loop. The challenge for the industry remains the "black box" nature of these models; as banks move toward multi-source foundation models that ingest phone calls and branch visits, regulatory scrutiny regarding data privacy and "explainable AI" will likely intensify across the sector.

Frequently asked questions

How is BBVA using Transformer models in its mobile app?

BBVA uses Transformer models to treat sequences of app screens as 'sentences' and individual visits as 'tokens.' This allows the bank to analyze the context and sequence of customer interactions, creating anonymized vector representations that help predict user needs and improve service personalization for 8 million customers.

What are the benefits of BBVA's new AI research for customers?

The primary benefit is a more personalized and proactive banking experience. By accurately predicting the likelihood of a customer needing a specific financial product, BBVA can deliver better-tailored services and optimize app processes, resulting in a 2.7% improvement in model accuracy for real-world applications.

Can BBVA's AI system handle large-scale banking data?

Yes, the system is designed for scale, currently processing four million sessions per day from a pool of eight million active users. The researchers developed a hierarchical tokenization system that allows the AI to adapt to new app features without needing a full system rebuild.

More AI in Finance news →

About the companies

BBVA

Website · 53 stories and 3 video interviews on FF News since 2017

Previously on FF News:

Companies in this story: BBVA

More from News