Scaling Micro-Offers With Hybrid AI

Andre Victor Ruiz Pedroso, Jessyca Oliveira17 Jun 2026

Problem

A bank wants to send investment offers to hundreds of thousands of customers. A generic campaign converts at 3–5%, and blasting everyone trains customers to disable notifications, after which the bank cannot reach them at all. A propensity model alone does not fix this: a high score says the customer may accept, not that the offer fits his situation. The bank is regulated and serves vulnerable customers, so decisions need to be explainable to stakeholders and auditors.

Example from the article: Caixa Econômica Federal, a Brazilian public bank, started from about 400,000 users of its My Wallet app and about 200,000 of them scored as propensity-positive. It wanted to contact far fewer.

Solution in brief

The pipeline has three layers: K-means behavior segments, a Random Forest propensity model, and an LLM that plays a "digital twin" of each customer. The LLM says whether the offer makes sense for him now and why. A balance rule runs before the LLM.

Questions this case answers
  • How to narrow a propensity model's positive list to the few targets for whom an offer is also affordable and suitable
  • How to keep an explainable and auditable selection pipeline for targeting vulnerable people at large volume under regulation
  • How to scale personalized financial offers without spamming customers
  • How to filter a propensity model's output for customer suitability

Tools and methods

  • K-means — 11 behavior segments from balance, income, cash flow, investment history
  • Random Forest — propensity model, chosen because its trees can be shown to stakeholders
  • LLM digital twin — batch yes/no suitability check with a justification per customer
  • MLflow, Unity Catalog — model comparison, registry, and a log of every prompt and answer for auditors
  • Databricks (SQL, AI Playground) — data pipelines and cheap model selection
Continue reading — it's free

Google or a one-time email code. No password.