BBVA

BBVA, one of the world's largest banks operating across multiple countries, relied on expert-designed commercial campaigns built from limited datasets and proxy signals rather than true customer-level data. A campaign targeting parents for a minor's savings account, for example, required manual filtering by indirect indicators such as school direct debits — a slow, imprecise, and unscalable approach. These static campaigns could not adapt in real time, could not account for individual customer context, and required full redesigns for each market. As digital commerce accelerated and customer expectations for relevant, timely offers rose, the gap between what BBVA could deliver and what customers expected grew into a measurable conversion problem.

BBVA's internal AI Factory built a production-grade multiclass machine learning pipeline using LightGBM — a gradient-boosted decision tree algorithm well-suited to high-dimensional financial data — to predict which commercial product each customer is most likely to need at any given moment. Feature engineering converts raw behavioral and transactional data into meaningful signals (e.g., average supermarket spend, app activity patterns), and standardized global data sources allow a single pipeline to serve all operating regions without country-specific rebuilds. CleanLab handles dataset imbalance by removing noisy observations from over-represented product classes, while Halving Search optimizes hyperparameters iteratively. The system integrates directly into BBVA's mobile app, delivering real-time, personalized offers at the individual customer level. No external vendor is identified; the system was developed entirely in-house by the AI Factory team.

79% of BBVA's total sales are now completed through digital channels, with the recommendation engine operating across all of the bank's active regions. The shift from static, expert-designed campaigns to a continuously learning ML system produced measurable improvements in offer relevance and conversion precision. Key outcomes include:

• 79% digital sales share — the headline metric reflecting the scale of digital channel adoption supported by personalized recommendations
• Real-time personalization deployed at production scale across multiple countries from a single pipeline
• Continuous model improvement via online A/B testing, where new model versions are only promoted to production when improvements are statistically significant
• Fairness and explainability embedded into the evaluation cycle, reducing risk of biased recommendations across protected population groups

• Standardizing feature engineering across geographies allows a single ML pipeline to serve a global institution — avoiding the cost and drift of country-specific model rebuilds.
• Dataset imbalance is a material risk in product recommendation: tools like CleanLab should be applied proactively to prevent models from over-predicting high-volume products at the expense of strategic ones.
• Online A/B testing should gate model promotions — deploying only statistically significant improvements prevents regression and builds organizational trust in the system.
• Fairness metrics (PPV, FPR, FNR) belong in the standard model evaluation checklist, not as an afterthought — especially when recommendations affect financial access across diverse customer segments.