A customer health score you can build this quarter.

You can build a customer health score this quarter. Not a vendor template with a red-amber-green light bolted on the side - a real one, built from the accounts you already renewed and the ones you already lost.

Most scores are built the wrong way round. Someone picks a few inputs that feel important, weights them by instinct, and is then genuinely surprised when the green accounts churn and the red ones renew. Build it backward instead, from what actually happened.

Start with the accounts you already lost

Pull two years of renewals and two years of churns. Line them up side by side.

Now look at the months before each decision, not the month of it. The signal is always earlier than the churn. The customer who left in month eleven usually stopped using the thing around month three. You are hunting for what separated the two groups back when nobody in the room was worried yet. You are not looking for the reason each account left. You are looking for what the leavers had in common that the stayers did not.

The four inputs that usually survive

Across the accounts I have looked at, four behaviours tend to do most of the separating. None of them is exciting. All of them move before the renewal does, which is the only property that matters.

Adoption depth. Real active users, spread across the team - not seats sold, and not one power user carrying the whole account on their back. Breadth is what makes you hard to remove.

Time to first value. How long before they hit a real outcome, not their first login. The faster the first genuine win, the stickier the account stays for years after.

Champion present. Is the person who fought to buy you still in the building. When the champion leaves, a clock starts, and the person who inherited the account never wanted it in the first place.

Support pattern. Not ticket count - the shape of it. A flood of tickets is an engaged customer working hard to succeed with you. Silence from an account that used to talk every week is the one that should scare you.

Keep the first version embarrassingly simple

The instinct is to build something clever - fifteen inputs, a weighted model, a machine-learning layer someone read about on a flight. Resist it.

Three or four inputs you actually trust will out-predict fifteen you half-measure. You can compute the whole thing in one CRM property or a spreadsheet this month. The goal is not elegance. The goal is a number that moves before the customer does, and that a real person will look at on a Tuesday.

Weight by what separated them, not by what feels important

Do not weight by gut. Look at which of the four best split your own renewals from your own churns, and weight that one heaviest.

If you have no time to model it, equal weights still beat a feeling. Start crude. A rough score you act on is worth more than a precise one you are still designing when the renewal arrives. Tune it every quarter as more accounts close and you learn which inputs were carrying the prediction.

Wire it to act before the renewal, not after

A score that only lives on a dashboard changes nothing. The point is the trigger.

A score that drops below your line should create a task, with an owner and a date, in month three - while you can still do something about it - not in a review three weeks before the renewal, when you are just narrating the loss. This is the whole game, and it is why retention is decided far earlier than the renewal date. Read the leading signal, act while it is still leading.

Give that trigger a real owner and a real threshold. Not the whole customer success team, which means nobody. One person, one number that means act, one place the task lands. The account manager should hear that an account is slipping from the system, not from the customer's cancellation email.

A score built on stale data lies

Your score is only as honest as the fields feeding it. If the champion field points at a contact who left eight months ago, the score glows green while the account quietly dies.

And data goes stale on its own. HubSpot's own data puts B2B contact decay at about 22.5% a year. One CRM audit I ran found 722 custom properties, 203 of them completely empty. Feed a health score from a CRM in that state and you have not built a predictor - you have automated a guess. Clean the inputs first, because trustworthy data is the thing every downstream number quietly depends on.

Clean the inputs before you trust the output.

Check whether the score was right

Every quarter, look back. Did the accounts it flagged red actually struggle, and did the green ones actually renew. If the green accounts keep leaving, an input is lying or a weight is wrong - and the fix is to go back to the accounts you lost, not to bolt on three more inputs.

Most health scores tell you an account is red the week before it leaves. That is not a health score. That is an obituary.

Common questions