Earlier churn signals for Telstra retention teams.

The challenge.

For a telco the size of Telstra, churn is a constant and expensive problem — winning a customer back costs far more than keeping one. But the signals of a customer about to leave were scattered across billing, network performance, service tickets and interaction history, and no single view brought them together. By the time an account showed up on a report, the customer had often already made up their mind.

Retention teams wanted to act early, but they were working from lagging indicators and gut feel. There was no reliable way to rank which of millions of customers were most at risk this week, or why — so effort was spread thin and the customers most worth saving were easy to miss.

What RUBIX did.

RUBIX treated this as a data problem first and a modelling problem second. Working alongside the telco's data, network and customer teams, we:

• Built a governed data foundation for the models — agreed definitions of a customer, a service and a churn event, with lineage and quality rules so the features fed to the model could be trusted.
• Engineered 200+ features from billing, usage, network quality, service tickets and contact history, capturing the behavioural patterns that precede a customer leaving.
• Trained and validated machine learning churn models, testing several algorithms and tuning them against real retention outcomes rather than accuracy in isolation.
• Put the models into production — scored daily, monitored for drift, and delivered as a ranked, explainable list of at-risk customers straight into the retention team's workflow.

Throughout, delivery stayed fixed-scope and vendor-independent — no bloated team, no open-ended engagement, and model governance that fit the telco's existing risk and change processes.

The results.

With the models live and scoring daily, retention teams began seeing at-risk customers weeks before the old reports would have flagged them. Rather than chasing everyone, they focused on the highest-risk accounts first, with a clear reason for each — a billing shock, a run of dropped calls, an unresolved complaint — so the conversation could be relevant.

• At-risk customers surfaced roughly six weeks earlier than the previous reactive approach.
• Churn models in production, scoring the customer base daily with monitoring for drift.
• Retention effort prioritised on the top 10% of accounts by risk, with a per-customer reason to guide the offer.
• A repeatable, governed pipeline the telco's own teams can retrain and extend.

Representative example — figures illustrate the type of outcome RUBIX delivers for organisations of this kind and are not a precise account of a specific engagement.

Why it matters.

Machine learning only works on a governed data foundation, and a model only earns its keep once it reaches production. RUBIX's AI consulting starts with the definitions, lineage and quality that make features trustworthy, then engineers, validates and deploys models into the workflows where decisions actually happen. A churn score sitting in a notebook changes nothing — a governed model scoring daily into a retention team's queue changes revenue.