Predict Churn and Drive Expansion with Health Scores

Building the Signal Architecture

A health score is only as predictive as the signals it aggregates. The most common design mistake is building a score on assumed signal weights rather than validating those weights against historical churn data. A signal that feels like it should predict churn may not actually correlate with churn in a specific customer base. The correct sequence is to start with historical account data: which accounts churned in the past two years, what did their product usage, support volume, NPS responses, and engagement patterns look like in the six months before renewal? The signals that consistently appeared in churned accounts and were absent or different in retained accounts are the signals that belong in the model, weighted according to their actual predictive correlation.

For most recurring revenue businesses, the signals that carry the highest predictive weight fall into four categories. Product engagement signals measure whether the customer is actually using the product in the way that delivers value: login frequency, feature adoption across the core workflow, active user percentage relative to licensed seats, and whether usage is growing or declining over time. Outcome signals measure whether the customer can point to a business result from the product: are they achieving the metric the purchase was justified by? Relationship signals measure the health of the human connection: when was the last executive sponsor interaction, what is the CSM’s qualitative assessment, and has the customer responded to recent outreach? Risk signals measure active friction: open high-priority support tickets, billing disputes, any expressed dissatisfaction in any channel.

Health Scores as an Expansion Signal, Not Just a Churn Warning

The expansion use case for health scores is underutilized relative to the churn use case. An account that scores in the top quartile on the health model is demonstrating the conditions that typically precede successful expansion: deep product engagement, positive outcome achievement, and an active relationship. These are not accounts to leave alone until their next renewal. They are accounts to engage proactively with an expansion conversation grounded in the usage data.

The data that makes a high-health account a churn predictor in reverse also makes it an expansion evidence base. If an account has 80 percent of its licensed seats active and is using five of eight available core features with high engagement, there are three features they are not using. A conversation about those three features is not a sales call framed as a success call. It is a success call that happens to identify potential product expansion. The expansion recommendation is grounded in the same usage data that the CSM reviews in every account meeting. The difference is that the health score model surfaces which accounts are positioned to receive that conversation successfully rather than requiring the CSM to develop that judgment independently for each of their accounts.

Operationalizing the Score

A health score that lives in a spreadsheet and is reviewed quarterly is not an operational tool. For the score to change CSM behavior, it needs to be visible in the system where CSMs manage their accounts, updated on a frequency that reflects how quickly conditions can change, and connected to a defined set of actions that each score tier triggers.

The action triggers are as important as the score itself. A red score that does not automatically create a task, route to a manager for review, or generate a stakeholder notification does not change what happens to the account. The score is the diagnosis. The action triggers are the treatment protocol. Defining those triggers explicitly, by score band and by the specific signals driving the score, converts the health score from a reporting metric into an operational system.

The refresh cadence should match the natural velocity of the signals being tracked. Product usage and support volume update daily; the score should reflect that. Relationship signals update on whatever cadence the CSM logs interactions; the score should incorporate new interaction data within hours of it being logged. An account that was amber on Monday and whose executive sponsor just expressed serious concerns in a call on Thursday should not still be showing amber on Friday. The score should reflect current reality, which means the data pipeline needs to support near-real-time updates rather than batch refreshes that lag the actual account state by days or weeks.