Customer health scores. Predicting churn before it happens.

The honest evolution: nothing, then too much, then simple

Early on, we never really had a formal health score. We may have looked at logins, but that was it — churn mostly announced itself. As we matured, we got more sophisticated: we implemented our own customer health score and prediction algorithm, and along the way we also implemented the heavyweight customer success tooling — ChurnZero, in our case. I'm not a fan, and it's nothing against that vendor specifically; it's the whole category, and I'll make that full case in the dedicated tools piece later in this series. What matters here is where the journey ended: the thing that actually worked was the simple system we built ourselves.

The score that was really an email

Our method was a simple email, sent company-wide, with a prediction score and a color code for the account. Behind it sat four inputs: logins, product usage data, the percentage of seats utilized, and customer tickets. That's it — no boil-the-ocean model, no fifty-factor weighting exercise.

And here's the part most health-score articles miss, because they obsess over the algorithm: the power of ours wasn't the math. It was the distribution. Because the score went to the whole company, when a large-ARR account went red or yellow, it raised alarm bells instantly, everywhere — not in a dashboard one CSM might check on Thursday, but in front of every person who could do something about it. That visibility is why it worked. We saved several large accounts that way: the company-wide email went out with the churn risk on it, and we intervened early enough precisely because everyone saw it at the same time.

The signals, ranked by what actually predicted churn

If you're choosing what feeds the score, here's the ranking our sixteen years produced. Declining usage and logins is one of the best signals there is — the top of the list. Contracting seats is right behind it: a customer quietly reducing how many people use the product is telling you something with their budget. And a champion leaving is a very strong signal, almost always in the bad direction — when the person who brought you in or runs you internally walks out the door, the account's gravity changes. Bug counts are important too, but not as much as declining usage and logins; a frustrated account that's still using you heavily is fighting with you, and fighting beats leaving.

Notice what those four have in common: none of them require a survey, a sentiment model, or an integration project. They're behavioral, they're already in your systems, and they show up before the churn number does — which is the whole point. The churn metric tells you what already happened; these signals are the usage story turning before it becomes a cancellation.

The lead time is weeks — so the response must be instant

Here's the uncomfortable truth about how much warning you actually get: the signals show up in weeks. A pattern of a few weeks of declining logins and usage is all you need to know — and all you're going to get. That means you need to be on these accounts almost instantly. The failure mode is treating the early warning as something to keep watching: give it another few weeks "to confirm the trend" and you may already be too late, because the customer's internal decision is forming on the same timeline your data is.

So the response play has to be ready before the signal fires. At PipelineCRM, the success team almost always owned the response — and if the account was big enough, a co-founder got involved directly. The first move was almost always a call or an email. And if the success team felt the account was genuinely at risk, it went onto our IDS list as an issue — which plugged the at-risk account straight into the weekly operating machinery, with an owner and a resolution path instead of a worried Slack thread. The responses themselves ran a ladder: a simple "how can we help," re-training the admin or the team, or getting on the phone to dig into configuration, data, or process issues — because a lot of "declining usage" turns out to be a fixable setup problem, not a verdict on the product. (What happens from there is the save play — the full diagnosis and the menu of responses gets its own piece.)

The advice: the minimum viable health score

If you're a founder at $1–10M ARR, my advice is to resist the temptation to buy a customer success platform like ChurnZero or Vitally — and again, this is a category critique, not a vendor one. The reasons stack up fast: they're way more than you need. They're very expensive. They will absolutely compete with your CRM system. They require deep integrations to build and maintain. And the return we saw, given the investment, was extremely low.

There's a cultural cost too, and it's the part I genuinely disliked: these tools turned our CSMs into email-drip-campaign managers. I'm a bit old school on this — picking up the phone and building the relationship the old-fashioned way is much better in the long term than blasting your customer base with campaigns trying to get them to engage with your product. A health score should trigger a human conversation, not an automation sequence.

So here's the minimum viable health score: start with login data — frequency and users — because that alone is a huge signal and every SaaS business already has it. Add seat utilization and ticket volume when you're ready for four inputs instead of one. Roll it up to a score and a color. Then do the part that actually made ours work: put it in front of the whole company on a schedule, and treat a red large account as an issue with an owner, not a dashboard widget. The algorithm can stay simple forever. The visibility and the speed of response are the product.