What AI Changes About Customer Health Scoring
A lot of health scores are built from some mix of last login, an NPS response and the CSM's gut feel. They look fine on a dashboard right up until a green account churns and everyone scrambles to explain why.
AI can help with that. You don't need a data science team to get value from it, but you do need to understand what's wrong with the traditional model first, or you'll just automate the same blind spots.
Why traditional scores miss
The first problem is vanity metrics. A customer logging in doesn't mean they're succeeding, and filling out an NPS survey doesn't make them loyal. These are lagging signals, and a score built on lagging signals keeps CS reactive. They're the same support-era metrics most CS teams should be retiring. I've made the longer case for measuring outcomes instead of logins, and it applies directly here.
The second is gut feel. Experienced CSMs have good instincts, but bias creeps in. "They sounded happy on the call." "They really like working with me." Churn doesn't care how the last call felt.
The third is that most models are static. They get updated quarterly, if that, while customers change month to month. By the time the score moves, the risk has usually been sitting there for a while.
What AI actually adds
More signals, updated constantly. A model can watch far more inputs than a person can track by hand: depth of usage rather than login counts, feature adoption milestones, sentiment in emails and call notes, whether expansion conversations are happening, and patterns in support tickets. And it can do it close to real time.
Patterns before people see them. The real value is prediction. A model can find combinations of behavior that tend to come before churn or expansion in your own customer base. As an illustration, it might surface that accounts whose usage falls off after the first several months churn far more often, or that customers who attend onboarding sessions expand more. Whatever the specific patterns turn out to be for you, they let the team step in earlier instead of reacting at renewal.
Different definitions of healthy. Enterprise and SMB, high-touch and tech-touch, one industry and another: they don't look the same when things are going well. A model can learn what healthy means for each cohort instead of forcing one formula on everyone. If you've already segmented into scaled and strategic motions, those segments are a natural place to start.
What it looks like in practice
Picture a team with a few hundred accounts. The model flags a group of them trending down: lower use of a few core features, support tickets taking longer to close, no engagement with recent product updates. That's an illustrative example, but the shape is common.
Instead of waiting for the renewal scramble, CSMs reach out with a specific re-engagement plan for each account. Product sees the same data and adjusts the adoption journey for the features that are lagging. Fewer surprises, and more of those accounts still around to expand.
How to start without overwhelming the team
Audit what you have. What are you tracking today? Which signals have actually preceded churn or expansion, and which are just comforting to look at? Keep the ones tied to outcomes. The fundamentals of building a health score that predicts churn are the same whether or not a model is involved.
Add AI carefully. Don't try to model everything at once. Usage drop-offs, changes in ticket volume and sentiment, and dips in executive engagement are good starting points. The model gets better over time; it doesn't replace your health score overnight.
Then train CSMs to act on the flags, not just watch them. That means prioritizing flagged accounts, tailoring outreach to the specific risk, and treating the score as a signal rather than a verdict. This matters even more once you're using AI to cover a larger book of accounts, because the flag is often the only thing that gets a CSM's attention.
Mistakes to watch for
The most common one is trusting the first model. Check its predictions against what actually happened, and keep checking. The second is building a dashboard nobody can read; keep the score visual and simple enough that a CSM can act on it in a minute. The third is ignoring small signals. Early risk usually looks minor right up until it isn't.
A healthy amount of skepticism, plus managers who coach to action, is what gets these models used well.
AI won't replace a CSM's judgment about an account. What it can do is make sure that judgment gets applied to the right accounts, early enough to matter.
If you're rethinking your health score and want a second opinion, reach out.
