B2B Churn Happens Quietly
In B2C subscriptions, churn is recorded the moment a customer clicks "cancel." B2B accounts don't work that way. There is no cancellation notice. Weekly orders turn biweekly, the number of items per order shrinks, and one day you realize nothing has come in for three months. By then, the volume has already moved to a competitor.
That is why a B2C-style "cancelled or not" classifier breaks down immediately: you can't even define the label. B2B has no contract end date, only gradual decline. Churn prediction therefore starts with deciding what churn actually means.
Define Churn First, or the Model Has Nothing to Stand On
In practice, two criteria are used together:
Applying one rule to every account floods you with false alarms. Seasonal customers, such as those in agriculture or construction, should be compared with the same period last year, not the preceding months. Project-based customers with fewer than four transactions a year go quiet as a matter of course, so put them in a separate group and either exclude them from prediction or manage them on their own track.
The Signals That Predict Churn
Churn usually shows up in the numbers before anyone says a word:
These signals live in different systems: ERP (orders, revenue, payments), CRM (sales activity, contact changes), and customer support (claims, inquiries). The core work is unifying the customer master by business registration number and building monthly snapshots that contain only what was knowable at that point in time. Let future information leak in, and validation scores will look great while real-world performance collapses.
More Important Than the Model: Turning Scores Into Action
A risk score that sits on a dashboard changes nothing. Each score band needs a playbook that spells out who does what, and by when:
Salespeople won't trust a score without a reason. Explanation techniques such as SHAP let you show the top drivers for each account, for example "order interval up 2.4x, items purchased down from 12 to 7, one quality claim last month." That tells the rep what to ask about when they walk in.
Measuring Impact and Running It Over Time
"The accounts we contacted stayed" proves nothing on its own. They might have stayed anyway. Set aside a random slice of the high-risk group (say 20%) as a no-contact control group and compare churn rates. If the control group churns at 30% and the contacted group at 18%, the intervention prevented 12 percentage points of churn.
Buying patterns keep shifting, so we recommend retraining the model every quarter and reviewing the churn definition once a year (the gap multiple and decline threshold) against the accounts that actually churned.
How POLYGLOTSOFT Builds Churn Prediction
POLYGLOTSOFT starts with the ERP transaction data you already have, before asking you to collect anything new. Phase 1 unifies the customer master and settles the churn definition. Phase 2 builds the prediction model along with per-account explanations. Phase 3 connects risk scores and playbooks to your CRM and measures the impact of each response. If you want to act when orders start to slip rather than after they stop, get in touch with POLYGLOTSOFT.