Before building a model, you need to understand which behavioural signals are relevant for your business. This varies significantly by sector:

• SaaS: declining usage, fewer active users per account, fewer login days, no trial conversion
• E-commerce: longer time between purchases, declining average order value, more returns, less response to emails
• Telecom/utilities: more customer service contact moments, complaints, comparing on price comparison sites (if measurable)
• Subscriptions: cancellation attempts, plan downgrades, not using premium features

Those signals are the input variables for your model. They must be available in your data sources. If they are not there, the best model will not help you.

Churn modelling sounds like an AI problem. In practice it is 70-80% a data problem. You need:

• Historical data of customers who left (labelled as "churn")
• Historical data of customers who stayed (labelled as "no churn")
• Behavioural data over a relevant period before the churn event
• Demographic or contract data per customer

That data needs to be clean, combined from multiple systems (CRM, usage data, billing), and at the right grain (per customer, per period).

Several model types are suitable for churn:

• Logistic regression: simple, interpretable, good as a baseline
• Random forest / gradient boosting: better predictive power on complex data, less interpretable
• Neural networks: strong with very large datasets and complex non-linear relationships
• Survival analysis: model not just whether a customer churns, but when

For most organisations just starting with churn modelling, gradient boosting (XGBoost, LightGBM) is a good choice: strong performance, reasonably interpretable.

A churn score without action is useless. The value lies in the workflow that follows:

1. Segment by risk: high risk, medium risk, low risk. Not every customer with elevated risk deserves the same intervention.
2. Link to customer value: a customer with high churn risk and high revenue gets priority for personal outreach. A customer with low risk and low revenue might receive an automated email.
3. Test interventions: which action (discount, personal conversation, product training) most effectively reduces churn? Measure and iterate.
4. Close the loop: record which interventions worked. This improves both the action strategy and the model.

Churn prediction has real limitations:

• Unobservable behaviour: if a customer has internally decided to leave but has not yet changed their usage patterns, the model will not pick that up.
• External factors: a competitor offering better prices, a market shift, an economic downturn: these are not visible in historical customer data.
• Self-fulfilling prophecy: if you only actively approach customers with high churn risk, you do not know whether the low-risk group would also have churned without intervention.
• Small datasets: churn models need sufficient examples of both classes. With hundreds of customers, machine learning is less reliable than with thousands.

Mach8 helps organisations set up churn prediction systems: from data extraction and model training to integration with CRM systems and action workflows. We always start with a data inventory: which data is available, what is missing, what is the quality?

Based on that inventory we determine whether machine learning makes sense or whether simpler rule-based systems are a better starting point.

AI-based churn prediction works, but quality depends heavily on the available data and the workflow that follows. A model without an action strategy has no business impact.

Want to address customer churn in your organisation? Get in touch with Mach8 for an analysis of your data situation and the possibilities.