Ask a VP of Customer Success how the health score is calculated and you usually get a pause, then "it's a mix of usage, support tickets and NPS." Ask which account turned red last month and whether the score caught it first, and the pause gets longer.
The customer health score is the most-built and least-trusted artifact in customer success. Almost every team has one. Almost nobody makes a decision on it alone. This post is about why, and about what a score has to look like before someone will bet a renewal forecast on it.
Why most health scores fail
They measure the vendor's idea of value, not the customer's
The typical score weights login frequency, feature adoption, ticket volume and a survey. Those are things you can measure, so they went in. Whether they track what the customer bought you for is a different question.
A customer who bought your product to produce a quarterly board report is healthy if the report ships four times a year. Their logins will look terrible. A customer whose team uses you daily but never got the integration they were promised is unhealthy, and their usage will look great. A score that cannot tell these two apart is not measuring health. It is measuring activity.
They are tuned once and never again
Somebody picked the weights in a workshop in 2024. The product has shipped forty releases since. The book has doubled. Nobody has re-opened the spreadsheet because nobody owns it, and because changing the weights would make this quarter's numbers incomparable to last quarter's.
They hide their reasoning
A score of 62 tells a CSM nothing about what to do. Is it 62 because usage is dropping, because there are three open P1 tickets, or because the survey came back lukewarm eight months ago? Each of those calls for a different action. A number that compresses them into one digit has destroyed the information a CSM needed.
They lag
Most scores are recomputed weekly or monthly from data that was already days old when it landed in the warehouse. By the time the account turns red, the champion has been gone for a month and the renewal conversation has already happened without you.
What a defensible health score looks like
"Defensible" means a specific thing: a CSM can walk into the renewal review, point at the score, and answer "why that number?" with a sentence that the account exec, the CFO and the customer would all recognize as true. That requires five properties.
1. It is graded against the customer's own definition of success
Before anything else is measured, somebody writes down, in that customer's words, what they bought you to achieve. Not "increase adoption." Something like "cut the monthly close from eight days to three" or "stop losing deals because reps can't find the latest pricing."
Every signal after that is judged against that sentence. Logins only matter if logins are how this customer gets to their outcome. This is the step most teams skip because it is manual, and it is the step that makes the rest of the score mean something.
2. It separates the dimensions instead of averaging them
Health is not one thing. At minimum, a useful score keeps apart:
- Is the customer getting the outcome they bought? (Value realized.)
- Are the people who champion you still there and still engaged? (Relationship.)
- Is the product doing what it was supposed to, technically? (Delivery.)
- Is the commercial relationship clean? (Payments, contract, scope.)
An account can be green on three and red on one, and the one tells you what to do. Averaging them to 71 tells you nothing. Keep the dimensions visible; show a roll-up if you must, but never only the roll-up.
3. Every grade comes with its evidence
The score should be able to show its work: "Relationship: at risk. The executive sponsor who signed the contract left in August (LinkedIn, confirmed by the ticket reassignments). The new VP has not been on a call. Last meaningful email from the account was 41 days ago."
That paragraph is the score. The color is a summary of the paragraph. If your tooling only stores the color, you have thrown away the thing the CSM needed.
4. It is recomputed as the evidence changes, not on a calendar
A health score that updates when a champion leaves, a ticket escalates or usage shifts is useful. One that updates on the first of the month is a report. The difference between them is a renewal.
5. It is scored as a sequence, not a snapshot
A single reading of 62 means little. A reading of 62 that was 80 six weeks ago, dropped when the analyst left, and has not recovered since the replacement started, is a story a CSM can act on. The score should be judged by how early its trend caught a change that later mattered, not by whether the number looked reasonable in isolation.
How to build one without a data science team
You do not need a model to get most of the way here. You need discipline about what you write down.
- For every account, write the success sentence. One line, in the customer's words, dated. If you cannot write it, that is your first finding: you do not know why this customer bought.
- Pick four or five dimensions and keep them separate. Do not average. Give each one a grade and a one-paragraph justification.
- Make the justification mandatory. A grade with no evidence is not allowed to be saved. This is the rule that keeps the score honest a year from now.
- Re-grade on events, not on a schedule. Champion change, escalation, a missed outcome milestone, a usage shift in the module that matters for this customer's success sentence.
- Backtest against last year's churn. Take the accounts you lost and ask: on which date would this score have gone red, and what did we actually know on that date? If the answer is "two weeks before the non-renewal email," the score is not early enough to be useful.
Where AI fits
Everything above is hours of reading per account per week, which is exactly why most teams do it once and never again. This is the part that machines are good at: re-reading the usage data, the tickets, the email, the call transcripts and the public web every day, for every account, and writing the evidence paragraph for each dimension against the success sentence a human wrote.
The human still writes the success sentence. The human still looks at the grade and the reasoning and decides what to do. But the score stops being a number somebody tuned in 2024, and starts being a thing you can defend in the room because it says, in plain words, why.
That is the only health score worth having. If yours cannot do that today, the fix is not a better weighting. It is deciding to keep the reasoning.