Customer health scores are useful precisely because they compress a lot of information into something a team can scan and act on.
Product usage, support activity, relationship data, implementation progress and commercial information can all contribute to a simpler view of the account. That is operationally valuable.
The risk is when the compression becomes the explanation.
Green does not mean we understand the customer. Red does not mean we know what is wrong. The score is a signal, and like any compressed signal it leaves detail behind.
The question is whether it leaves behind the detail needed for the next decision.
The same score can hide very different situations
Imagine two accounts that are both amber. One has low product usage because the rollout is deliberately phased. The other has high usage but weak evidence of the business result that justified the investment.
They do not need the same intervention.
Now imagine two green accounts. One is using the product heavily, has a clear business outcome and can demonstrate progress. The other is active and technically successful, but the sponsor who bought the product has left and the value case has not been rebuilt.
The colour can be the same while the route to renewal is very different.
That is not a criticism of scoring. It is a reminder to treat the score as a route into investigation rather than a verdict.
Every health score contains a theory
When we choose the inputs, weight them and decide what counts as green, amber or red, we are expressing a theory about customer success.
High usage means something. Support volume means something. Executive engagement may matter. Implementation delay may matter.
The useful question is what those signals mean in relation to the customer's outcome.
At IBM, working across a large software portfolio, common definitions mattered because different products and teams naturally used different language. Without a shared taxonomy, the same customer state could be described differently across the organisation.
I think the same principle applies to health. Before debating the formula, agree what the underlying states mean. What is adoption? What is value? What is risk? What is an outcome? What evidence would make us change our interpretation?
If those definitions are unstable, improving the arithmetic does not solve the problem.
Health should tell us where to look next
I prefer to think of a health score as a routing mechanism.
It should help a team decide where to look, what evidence is missing, which customer needs human attention and what has materially changed since the last review.
That means uncertainty can be useful information rather than something the model has to hide. An account might have strong usage evidence but weak outcome evidence. Another might have strong outcome evidence but a deteriorating stakeholder situation. A third may simply have too little evidence for a confident classification.
Those are different states, and they lead to different actions.
Product usage should not dominate simply because it is available
Telemetry is attractive because it is observable. We can instrument it, query it and build dashboards around it.
That can create an unintended hierarchy where product data is treated as more important simply because it is easier to collect. Sometimes it should be more important. Sometimes the most material signal sits elsewhere.
The customer may be using the product as expected while the business case has changed. A new executive may care about a different result. The product may be performing well while procurement is challenging the commercial model. The customer may already have achieved the original outcome and need a new reason to continue.
A useful health model needs enough context to distinguish these situations.
Common language matters more than a clever formula
One of the areas I examine through Routeability is common taxonomy. It sounds less exciting than predictive scoring or AI, but it is foundational.
If sales, implementation, Customer Success and product use different definitions for the customer's state, any score built on top of them will inherit the ambiguity.
A shared taxonomy does not mean every team needs the same dashboard. It means the important states have stable enough meaning to survive the journey across teams.
The outcome recorded in sales should still be recognisable to post-sale. Adoption should connect to the relevant customer objective. Risk should describe a condition rather than become a vague synonym for concern. Value evidence should have a meaning people can inspect and challenge.
Once those things are clearer, the score becomes more useful because the team can unpack what sits underneath it.
So the question I would ask about any health score is not only, "How accurate is it?"
I would also ask:
What decision are we making from it, and can we inspect the evidence underneath?
If the score helps answer that, it is doing useful work.


