BrightkeelHow to build a customer health score model
A health score is only useful if it moves before the customer leaves. This guide shows how to choose signals, weight them and set thresholds, with a worked example you can copy.
Pick signals that predict churn
Start with four groups of signals and keep only the ones that moved before past churn events.
Usage
Weekly active users as a share of paid seats, adoption of the two or three features tied to the customer's goals, and the trend over the last 30 days.
Support
Open ticket volume, time to resolution and satisfaction scores. A spike in tickets during onboarding is normal; a spike after month four is a warning.
Weight the signals
| Signal group | Weight |
|---|---|
| Usage | 40 percent |
| Relationship | 25 percent |
| Support | 20 percent |
| Billing | 15 percent |
Check the weights against the last 12 months of renewals. If the score would not have flagged at least two thirds of churned accounts 60 days ahead, adjust it.
Set thresholds and a worked example
Use three bands: green at 70 and above, yellow from 40 to 69 and red below 40. Alert on a move into red, or on a drop of more than 15 points in 30 days.
Northwind Labs, a fictional account
Usage scores 80, support 55, relationship 70 and billing 100. The health score is 0.4 x 80 + 0.2 x 55 + 0.25 x 70 + 0.15 x 100 = 75.5, which is green. If usage falls to 50 the score drops to 63.5 and the account moves to yellow.
Relationship and billing
Relationship covers champion activity and executive sponsor meetings. Billing covers late invoices and downgrade requests.
Review the model every quarter
Products change and so do the signals that matter. Brightkeel Signals keeps score history so you can test a new model against past renewals before switching. See Brightkeel Signals.