A churn prediction is not a decision: from risk score to retention action

Where the score stops

A typical churn dashboard ranks customers by risk and sends the top segment the same offer, usually a discount. That loses money in two places.

• Customers who would have stayed anyway get the discount too. Where the offer does not change behaviour, it is pure cost.
• Low-value customers cost more to save than they are worth. High risk does not make every customer worth saving.

Same 78%, two different decisions

Two customers, both at 78% churn risk. One pays ₺400 a month, the other ₺20. Three options: a reminder email (₺10), personal outreach (₺150) and 40% off for three months (1.2× the monthly value). Expected value = risk × effect × 12-month value − cost.

Effects and costs are representative assumptions: the email lowers risk by 4%, outreach by 25%, the discount by 12%. The real effect is only known from a controlled measurement.

For customer A the discount is not the best option even though it is the most expensive: cheaper outreach protects more value. For customer B no intervention covers its cost; spending money is worse than losing the customer.

The four inputs a decision needs

1. Calibrated risk: if the model says 78%, about 78% of those customers should actually leave.
2. Customer value: protected revenue, ideally margin, discounted by the cost of capital (WACC).
3. Intervention cost: direct cost plus forgone revenue such as a discount.
4. Effect (uplift): how much this intervention lowers risk for this kind of customer.

Why calibration matters

Expected value is proportional to risk. If only 50% of the customers a model scores at 78% actually leave, every intervention's expected benefit is overstated by 56%, and unprofitable actions look profitable.

Where to start

One customer-level table is enough: a customer key, the churn outcome and an observation window. Past campaign records let you model intervention effects. No direct identifiers are needed. Decisions are then compared with the current approach over the same period, with a control group.

Vinctum Aegis runs this end to end: calibrated risk, explanations, expected value and a decision log. You can try the same calculation with your own values in the simulator on the Aegis page.