Close panel

Close panel

Close panel

Close panel

AI Solutions 18 Nov 2026

BBVA uses AI to better understand its customers and takes a step towards developing a foundation model

BBVA's AI researchers have developed a model that interprets how customers interact with the bank's mobile app, providing a deeper understanding of their needs and paving the way for more personalized and proactive banking. The research, presented at ICAIF, a leading conference on AI applied to finance, marks another milestone in BBVA's efforts to lay the groundwork for an AI foundation model capable of integrating multiple data sources.

AI Solutions - BBVA aplica la IA para comprender mejor al cliente y avanza hacia un modelo fundacional de IA

Every time customers use BBVA's mobile app, they perform a sequence of actions, whether checking their transactions, reviewing their cards, exploring loan terms or making a transfer, among many other possibilities. When analyzed together, these interactions provide valuable insights that can help optimize the bank's processes, improve the customer experience and deliver services better tailored to individual needs and interests.

Harnessing this information through AI is the goal of a research initiative that BBVA has been pursuing for almost a decade. Its latest milestone is a scientific paper developed by the bank's researchers and presented on November 15 in Milan at the ACM International Conference on AI in Finance (ICAIF).

Asier González-Rodríguez, one of the paper's authors, during its presentation at ICAIF

Specifically, the paper was presented at a pioneering event held as part of the conference, dedicated to research into foundation and multimodal models for the banking sector.

A new approach to representing digital navigation

BBVA began exploring the use of embeddings in 2018. These numerical representations enable AI models to identify patterns and relationships within large volumes of data. In 2023, this research took another step forward by focusing on navigation data from the bank's app, transforming sequences of screens visited by customers into anonymized vector representations that can be reused by different analytical models.

This technology has already demonstrated its value in real-world applications at BBVA, such as customer segmentation and estimating the likelihood that a customer will take out a financial product. However, the approach used until now, based on Word2Vec, had an important limitation: although it could represent the screens visited, combining this information resulted in the loss of some of the context and sequence of interactions.

To overcome this limitation, researchers explored the use of Transformers, a deep learning architecture that makes it possible to analyze sequences and identify relationships between their elements, and which underpins large language models (LLMs).

The approach draws inspiration from natural language processing. Just as the meaning of a word depends on the words surrounding it, an action within the app can convey different information depending on the interactions that come before or after it.

In this case, each screen visited in the app is treated as a token, the basic unit of data that AI models read and generate, while the sequence of screens viewed during a session functions much like a sentence. Through self-supervised learning, the model learns to recognize navigation patterns directly from activity logs, without the need to manually label each sequence.

For this research, approximately 1.6 billion navigation sessions from eight million customers, covering one year of activity, were analyzed using anonymized data. Of these, 90 million sessions were selected to train the new models.

One of the main technical challenges was ensuring that this approach could operate at scale and adapt to an app that is constantly evolving. To address this, researchers developed a hierarchical tokenization system that groups related screens and makes it possible to interpret new features based on existing categories, avoiding the need to rebuild the system whenever the app changes.

Building on this foundation, different Transformer architectures were compared, with a decoder architecture ultimately selected. This architecture learns to predict the next screens in a sequence and combines contextual information from the entire navigation journey to generate a compact, reusable representation. As a result, each navigation session is converted into a vector that summarizes a user's behavior within the app. This vector can be used directly or combined with other sessions to obtain an overall representation of a customer's interactions.

From research to real-world banking applications

The results show that Transformer-based models outperform the previous technology by capturing navigation sessions more accurately and generating useful representations for analyzing customer interactions.

One of the most significant evaluations involved a model that estimates the likelihood of a customer taking out a financial product, a real-world use case already deployed at the bank. By incorporating the new embeddings, the AUC metric—which measures a model's ability to distinguish between customers who will take out a product and those who will not—increased by 2.7% in relative terms compared with the baseline model, also outperforming the result achieved with Word2Vec. This improvement is particularly significant given that the existing model already demonstrated strong predictive performance.

This improvement shows that using sequential models to interpret digital interactions provides additional information that can enhance the predictive capabilities of analytical models. Furthermore, the resulting representations can be reused across different models without requiring each one to process the customer's entire navigation history.

The solution is integrated into a process that serves approximately eight million active users and processes four million sessions per day, demonstrating its viability at scale while meeting requirements for efficiency, governance, reliability and maintainability.

Towards a foundation model that can learn from multiple data sources

The models developed so far focus on one specific dimension of customer behavior: digital navigation. However, much of a person's relationship with their bank is also reflected in other sources of information, ranging from transactions and account activity to interactions with relationship managers, phone calls and branch visits.

BBVA is already taking the first steps towards developing a multi-source foundation model capable of integrating these different dimensions and providing a foundation for building multiple AI applications.

Unlike models specialized in a single data source, this approach makes it possible to learn a comprehensive representation of BBVA customers' financial behavior from heterogeneous information. The goal is to move towards an intelligent, comprehensive and unified understanding of customers, providing a foundation for deploying AI models and agents across different use cases.

The research presented at ICAIF represents another step towards realizing this ambition. By combining scientific experimentation, accumulated knowledge and real-world deployment, BBVA is building a foundation for a new generation of AI models that can contribute to a deeper understanding of customers' needs and expectations, and to the development of banking that is increasingly personalized, proactive and people-centered.