Predicting Churn Without a Data Science Team

You do not need machine learning to see churn coming. Three behavioural signals catch most of it, and you can track them in a spreadsheet.

Churn is decided long before the cancellation email. By the time someone asks to leave, the decision is weeks old and the conversation is a formality. The value is entirely in the period before that, and finding it rarely requires a model.

The three signals that catch most of it

1. Usage decay. Not zero usage — the slope. A customer at 60% of their own three-month average is a stronger signal than one who was always light.
2. Support silence after a bad experience. A complaint followed by nothing is worse than a complaint followed by five more; the second customer is still engaged.
3. Champion departure. The person who bought you leaves the company and nobody inherits the relationship. This is the most predictable churn there is, and the easiest to miss.

What to do with the signal

A prediction nobody acts on is an expensive report. Define the intervention in advance and make it human: a real person, a specific offer of help, no automated 'we noticed you have not logged in' email, which reads as surveillance rather than service. The intervention rate matters more than the prediction accuracy.

When a model earns its place

Once you have several hundred churn events and the obvious signals are already handled, a model can find the combinations you cannot see by eye. Below that volume it will confidently learn your noise. The honest sequence is: measure, intervene manually, then automate what worked.

• 60% of own average usage as a warning line
• 3 signals that catch most churn
• 1 human intervention, defined in advance