Why you're the last to know a customer is leaving

Customer churn is not a single event. It begins when a familiar pattern breaks.

Most businesses discover they have lost a customer in a report: a subscription was not renewed, a contract was terminated, or a whole quarter passed without a purchase. By then, the customer made their decision long ago. The change began weeks earlier, beyond the view of the metrics most businesses routinely track.

Customers rarely leave overnight. First, the intervals between interactions grow longer. Track only the outcome and you can record churn; track behaviour and you can anticipate it early enough to act.

Reactive measurement captures an outcome, not a process

Traditional churn indicators—a cancellation, a non-renewal or no activity during a set period—have one thing in common: they describe something that has already happened. They are accurate, easy to measure and useful for reporting. For retention, they come too late.

Churn unfolds over time. It has a beginning, a progression and an end; termination is only the final step. Prediction is valuable when it catches that process while there is still time to reverse it.

Regularity is a stronger signal than volume

Businesses naturally monitor volume: revenue, order counts and monthly usage. These measures are easy to collect and present, but often fail to flag churn risk in time.

In our experience, widening gaps between interactions precede churn by weeks and provide an earlier warning than falling volumes. Erratic activity and longer pauses significantly increase churn risk. A customer with lower but steady activity may therefore be less likely to leave than one whose high activity comes in bursts.

Volume tells you how much a customer has done. Their rhythm tells you whether they are likely to keep doing it.

A discount does not equal retention

The usual response to a customer who may leave is a financial incentive: a discount, credit or gift. Its effect can be measured, but it is short-lived. Activity rises briefly, then returns to its earlier trajectory. A one-off incentive does not change long-term behaviour.

The business has bought time, not loyalty. Genuine retention means changing the customer's routine rather than the price. The aim is to restore regular engagement, not create a one-off spike in the data.

A score without an explanation is of little use

Many churn prediction solutions deliver only a number. A 78% probability that a customer will leave sounds precise, but gives marketing little to act on. It does not explain why the customer is at risk or what intervention might help.

Useful output answers three questions: who is at risk, why, and what should we do? That is why Codium's churn prediction provides risk factors and recommends specific measures alongside the probability of churn.

Customer value and churn risk also need to be assessed separately. They are two distinct dimensions, and the greatest potential benefit lies at their intersection: valuable customers whose interactions are becoming less regular. Broad campaigns overlook this group while consuming budget unnecessarily. Targeted, timely outreach is more efficient because it avoids offering incentives to customers who would have stayed anyway.

Defining churn is a business decision

What counts as churn varies widely between businesses. For one, it is a subscription that is not renewed; for another, 28 days without activity; for a third, a missed service cycle. Get that definition wrong, and a model may accurately predict an outcome the business does not actually care about.

The business, rather than the data team, should therefore define churn before any model is developed. Apply the same method with a different definition and you will get different risk factors and recommendations. That reflects a different business question, not an inconsistency.

Before you begin, we recommend answering these questions:

1. When, from our business's perspective, does someone cease to be an active customer?
2. What is our customers' usual pattern of interaction, and how much deviation should trigger a warning?
3. Which customer segments are most valuable to us, and whose departure would cost us most?
4. Who will use the results, and what actions can they take?

The limits of prediction

Churn prediction deals in probabilities, not certainties. A model can identify which customers are at risk, when that risk emerges and which behavioural patterns contribute to it. It cannot reveal personal reasons that leave no trace in the data.

Consistent customer interaction data is essential. The quality of that data sets the upper limit on what any model can achieve.

Where to begin

Your first steps require a change in how you look at customers, rather than advanced tools:

• Segment customers by current behaviour rather than historical volume. RFM segmentation is a useful starting point.
• Prioritise valuable customers whose interactions are becoming less regular. This is where an intervention can make the greatest difference.
• Replace broad campaigns with targeted outreach at the right time.

You cannot prevent every customer from leaving. You can, however, recognise the signs early. That is the difference between simply reacting to churn and managing it.

Want to stay up to date?

Data is changing how businesses understand their customers. On our blog, we write clearly and practically about AI, data, software development, business and everything we are working on at Codium.