The signals were in four systems: usage, tickets, invoices, calls. Each team saw its part. The account manager called after the notice, and the discount cost money with no effect.
The account manager calls after the notice. By then the decision is made, and the discount offered is a cost with no effect.
Usage in the product, tickets in support, invoices in billing, calls in the CRM. Each team sees its part. Nobody sees the customer.
The same email to the loyal and the leaving. The budget is spread thin, and the customers who needed a call got a newsletter.
We join the signals, train a model on who left before and why, and give each account a churn risk with the reasons behind it. Retention works from the list, in order, and the result is measured against a control group.
Three parts: join the data, train the model, put the list in front of the team.
One record per customer, over time.
Trained on who left, tested on who left later.
A list the team works every week.
Four phases, each with a number at the end.
Define churn, find the data, set the target.
One customer record, updated on a schedule.
Train, test, explain.
Score, act, measure, retrain.
Retention calls the right accounts, earlier, and knows what it achieved.
The score moves before the customer does. The call happens while there is still something to fix.
A ranked list instead of a mailing. The team's hours go to the accounts where a call changes the outcome.
Churn in the treated group against the control group, in customers and in revenue. The model's worth is measured, every quarter.
| Side by side | Today | With churn scores |
|---|---|---|
| When you know | At the cancellation | Weeks before, from the signals |
| Who gets contacted | Whoever complains, or everyone | The ranked list, top down |
| What the call is about | A generic offer | The factors behind that account's score |
| Where the budget goes | Spread across all accounts | To the accounts a call can keep |
| Proof it worked | Churn went up or down | Treated against control, in euros |
It depends on your data and how churn is defined. We report precision and recall at the list size your team can work, tested on periods the model never saw, before anything is deployed. The number that matters afterwards is the difference between the treated group and the control group.
A history of who left and when, and the signals before it: product usage, support tickets, invoices and payments, CRM contacts. A year or more of history is a good start. We inventory what you have in discovery and say what the model can and cannot learn from it.
Discovery is two weeks at a fixed price. The first scored list typically lands within weeks after the data is connected, and the control group needs a full churn cycle to show a result. We fix the timeline after discovery, when the data is known.
Customers, products and prices change, and the patterns before churn change with them. We monitor the score distribution and the precision of each list, and retrain when they drift. Every action and outcome recorded by the team becomes training data for the next version.
We are a small team of senior specialists. We pick the right model and the right layer, and we build the least machinery that does the job. You get a call with an engineer, not a sales deck.
