Most customer health scores are built from the things that are easy to measure: login frequency, feature usage, support ticket volume, contract value, days since last contact. Then someone weights them, and the result is a number that goes red before a customer churns.

The number works. That's the trap. It works well enough to be trusted, and it's systematically blind to a specific category of departure that in my experience accounts for a large share of the ones that actually hurt.

The failure I keep seeing

A customer stops using a feature because it doesn't fit how they work. They don't complain, because complaining implies a relationship where complaints get acted on, and they aren't sure they have one. They build a workaround, usually a spreadsheet. Usage on that feature drops, but overall usage stays fine because they are still doing everything else.

The health score stays green. It stays green for months. Then renewal comes up, someone senior asks what they're paying for, and the answer is a list of things they use plus one large thing they replaced with a spreadsheet. Now it is a price conversation, and you have no defense because you never knew.

The signal was there the whole time. It just wasn't the kind of signal the score is built to see, because the score measures activity, and what happened was a quiet substitution.

Silence is data, and it's treated as neutral

The deeper issue is that every health metric I have seen treats absence of negative signal as positive.

No complaint is scored as satisfaction. Steady usage is scored as engagement. Low support volume is scored as success, and it's often the opposite: a customer who has stopped asking for help may have stopped believing help would arrive.

I ran into the same structure in a completely different domain. I audited public transportation data quality and found that when the information a rider needs is missing, the rider checks the app, gets nothing useful, and stays home. No complaint, no ticket, no data. From inside the agency, the service looks fine, because the failure produces silence and every reporting system reads silence as satisfaction.

That is exactly the shape of the churn nobody sees coming. The failure mode does not generate the signal your instrument measures.

Three things I would add

Track what they stopped doing, per feature, per customer. Not aggregate usage. A customer whose overall activity is flat but who abandoned two workflows is in a different position from one whose activity is flat and stable, and most dashboards can't distinguish them. Abandonment is a much earlier and much sharper signal than decline.

Count the workarounds. Ask, in plain words: what are you still doing outside our product that you would rather do inside it? People answer this honestly because it isn't framed as a complaint, it's framed as a wish. Every answer is a precise description of a gap, and the accumulation of them is your actual churn risk register. Nobody keeps a shadow spreadsheet for fun.

Watch for the disappearance of your champion. Health scores track the account. Relationships live in people. When the person who advocated for you internally moves teams or leaves, your renewal is now in the hands of someone with no history with you, and no metric based on account activity will register that anything changed. In my experience, this single event predicts more surprise churn than any usage pattern.

Why the number stays green

It's worth being honest about why this persists, because the incentive is structural rather than anyone's failing.

A health score exists partly to reassure. It's reported upward. A model that frequently flags healthy-looking accounts as at-risk generates work and gets tuned toward calm. Over a few quarters it converges on a version that's quiet, which is comfortable and which has quietly optimized away the thing you built it for.

If your health score has never surprised you, it's not measuring anything you did not already know.

The cheap version of the fix

You don't need a better model. You need one conversation, and it isn't the quarterly business review, which is where customers tell you what they think you want to hear.

Ask a customer who looks healthy: what's the thing about working with us that you have stopped mentioning because you assumed it would not change?

That question gets an answer, reliably, and the answer is usually something specific that has been true for a long time. It is not in any dashboard because it never produced an event. It's the actual state of the relationship, and it's available for the price of asking.