The client, a leading U.S. consumer fintech, wanted to reduce churn among its direct-deposit members, whose retention best reflects full relationship health across behavior, product usage, and support interactions. Keeping these members engaged is central to the business, but the signals that predict when one is about to leave were scattered and hard to act on.
Retention signals lived in two very different worlds: a large body of structured behavioral and account features on one side, and unstructured chat and support transcripts on the other. Static, rules-based segments couldn’t detect emerging risk early enough, and they offered little explanation of why a member was drifting away. There was also no scalable way to take a known high-risk customer and quickly find others who looked just like them.
Compounding the challenge, analysts needed to explore these insights directly, without writing complex code against tangled data. To that end, any solution had to keep personally identifiable information and model outputs within a governed, compliant boundary.
IBM Consulting partnered with the client and Snowflake to build a retention intelligence layer natively on the Snowflake AI Data Cloud that fuses structured and unstructured signals into weekly, explainable churn scores.
At the core is a fusion model that unites the client’s structured behavioral features with chat and support transcripts embedded using Snowflake Cortex, so that qualitative context and quantitative behavior inform a single view of risk. Each member receives a refreshed churn score on a weekly cadence, paired with explainability that surfaces the true drivers behind the score rather than leaving teams to engage in constant guesswork.
To make the intelligence actionable, the team added vector similarity search that would allowing the client to take any high-risk member and instantly find the members most like them. They also developed a Snowflake Cortex Agent that answers plain-language questions, so business users can investigate retention trends conversationally without writing code. Together these capabilities power dynamic segmentation that drives targeted, timely retention interventions, all while sensitive data and model outputs remain inside the client’s governed, fully-compliant compliant environment.
