Scaling Micro-Offers with Hybrid AI

This work proposes a pipeline for responsible investment offer personalization in Brazilian Open Finance, integrating data governance, segmentation (K-Means), propensity modeling (Random Forest), and an ethical filtering layer via LLM-based digital twins. Applied to 427,929 Caixa Econômica Federal customers, the model achieved an AUC-ROC of 0.57. Digital twin validation reduced the high-propensity pool from 213,764 to 6,543 eligibles, applying suitability rules and generating auditable justifications to mitigate overcontact.Distributive analysis indicated a concentration of offers among higher-income profiles. In a pilot within the My Wallet app, the approach yielded a 21.0% positive intention rate, surpassing the 3–5% generic average. Results suggest that combining Open Finance, ML, and GenAI enhances assertiveness and algorithmic governance, though highlighting the need for formal fairness metrics and large-scale causal experiments.