Build a health score you can trust

A health score is how you turn a base too large to hold in your head into a queue you can work in thirty minutes. It is a single number per account that says, roughly, how likely this customer is to stay and grow. The trap founders fall into is either never building one because it feels like enterprise machinery, or building an elaborate one on day one that they cannot maintain. Both are wrong. The right health score is simple, honest, and revised against reality.

The three layers

A trustworthy score is built from three layers, weighted by how predictive each one is. The usage layer is the heaviest, around 40 per cent of the score, because product behaviour is the leading indicator and tells you soonest. It captures login recency, the breadth and depth of feature adoption, and whether the account is using the things that correlate with staying. The relationship layer captures the human signals: is your sponsor still engaged, is there more than one champion inside the account, what is the sentiment in support interactions. The commercial layer captures the hard facts: the renewal date, payment health, contract size and tier. A healthy usage picture with a quiet sponsor and a renewal sixty days out is a different account from a healthy usage picture with an enthusiastic champion and a year to run, and the score should know the difference.

The point of weighting usage most heavily is that it earns you time. The commercial layer tells you an account is at risk when the renewal looms, which is late. The usage layer tells you weeks or months earlier, while you can still act. Build the score so that the leading signal carries the most weight and you build yourself a head start.

Static scores rot, so re-weight against real outcomes

Here is the discipline that separates a health score you can trust from one that quietly lies to you. A static score rots. The weights you guess at on day one are hypotheses, not facts, and the only way to know whether they predict churn is to check them against accounts that actually churned. Each quarter, take the customers you lost and the customers who expanded, and ask whether your score saw them coming. If healthy-scored accounts are leaving, your weights are wrong, and you re-weight them against what actually happened. The score is a model of churn, and an unmaintained model drifts away from the truth it was built to capture.

Which is also why you should not wait for a perfect model before you start. You do not need clean event data and a data team. Start with two or three signals you already have, login recency, a count of one key event, support sentiment, and ship a crude score this week. A rough score you tighten against real outcomes each quarter beats a perfect score you never build, because the rough one is already surfacing your at-risk accounts while the perfect one is still a plan. Begin imperfect, then let reality sharpen it.