Problem
Reliance Jio, an Indian telecom operator, wanted a churn-prevention platform for hundreds of millions of customers. The usual pipeline does not scale: analysts hand-craft features from tribal knowledge, subject-matter experts (SMEs) define about 10 actions, and a campaign team builds and sends creatives. Scores rarely turn into actions, and nobody checks whether they worked. Churn was around 2%, and the goal was to halve it to 1%.
Example from the article: a customer with repeated IVR complaints and a gateway heartbeat drop is flagged as a silent risk, placed in a "bill-shock + service" cohort, and gets a fix and an offer after a human approves.
Solution in brief
An agentic, templated closed loop: data, prediction, action, execution, with four backward learning loops. Nine small agents, each with a fixed input, output and guardrail, do the analytics work. Humans approve each step in a Databricks App. The build is a template, so the next business line needs a new configuration, not a rebuild.
Questions this case answers
- How to score every customer daily for likelihood to leave and pick the next best action and channel for each at-risk customer among millions
- How to close the loop from a churn score to an executed retention action and feed the outcome back into features, model and routing
- How to predict churn for every customer daily and pick the next best action and channel for each at-risk customer
- How to let AI agents do the analytics work while humans approve every step before outreach
Tools and methods
- Churn Predictor with SHAP — daily scores and driver attribution
- Next best action (NBA) and routing matrix — action, timing and channel per cohort
- 5T agent design — Inputs, Actions, Outputs, Unknowns, Constraints
- Human-in-the-loop gates — approve, edit or reject in Databricks Apps
- Databricks stack — Unity Catalog, Feature Store, Lakebase, AI/BI Genie, Model Serving
- Claude Code — agents that explore data and write code
Google or a one-time email code. No password.