The 5 Inputs a Customer Health Score Needs Beyond Product Usage

Around 78% of customer success teams now use AI or machine learning somewhere in health scoring or churn prediction, up from roughly 32% in 2022. The modeling got better. The inputs mostly did not. A score that runs daily on logins, feature adoption, and ticket counts is precise, automated, and confidently wrong about the accounts that matter most, because the decision to leave gets made in conversations that no event stream records.

The five inputs a customer health score needs beyond product usage are: relationship continuity, qualitative signal from what customers actually say, billing and contract behavior, onboarding milestone completion, and seat density. Product usage stays in the model. It is just not sufficient on its own, and four of these five are available to most teams already, sitting in systems the scoring layer was never connected to.

What product usage can and cannot tell you

Usage answers whether the product is being touched. It is a good input and a lagging one.

The specific failure is well documented in the health-scoring literature: a weighted-sum model built on logins, feature adoption, and support tickets will rate an account green the week before a renewal call where a newly arrived VP announces a vendor consolidation. Usage was fine throughout. The relationship was already gone. Nothing in the telemetry could have said so.

Relationship and qualitative signals tend to lead usage decline by 30 to 90 days. That gap is the entire value of adding them, and it is why the ordering below runs from most leading to most lagging rather than from easiest to hardest.

The 5 inputs a customer health score needs beyond product usage

1. Relationship continuity

The most leading indicator available and the one telemetry cannot see. Three components: is the economic buyer still in the role, how responsive is the primary contact, and how many people at the account engage with you rather than just one.

Single-threaded accounts are the risk concentration. An account with one engaged contact is one job change away from having none, and that transition produces no product signal. Score sponsor tenure, reply latency on outreach, and breadth of engaged contacts, and weight this input heavily for enterprise segments where a single sponsor carries the relationship.

Where it lives: CRM, calendar, email, and call recordings.

2. Qualitative signal from what customers say

Not CSAT or NPS, which are scores. The actual language: what the account complains about, requests, and describes wanting to do. This is the input that explains a score change rather than reporting it, and it is the one most scoring models omit because free text is hard to turn into a number.

It becomes a number when themes are consistent. Score an account on the volume and severity of negative themes, whether those themes are resolved or recurring, and whether the account's requests are being met. An adaptive taxonomy that derives themes from the feedback itself is what makes this stable enough to weight, since a manually tagged corpus produces categories that drift and cannot be trended reliably across quarters.

Where it lives: support tickets, call transcripts, shared Slack channels, reviews, survey free-text.

3. Billing and contract behavior

The highest signal-to-noise input that almost nobody includes, because it lives in Stripe or Chargebee rather than in the product analytics tool. An account switching from annual to monthly is telling you something explicit about its commitment horizon. Failed payments, dunning states, and downgrade requests are all leading indicators, and monthly-to-annual moves are the inverse.

Where it lives: billing system and CRM. Usually a straightforward join and rarely done.

4. Onboarding milestone completion

Whether the account completed core setup inside its onboarding window, and whether the implementation team is still involved. Time-to-value research consistently ties clear onboarding to retention, with roughly 86% of customers more likely to stay when onboarding is clear.

The scoring nuance: this input should decay. Onboarding completion is highly predictive in the first two quarters and close to meaningless in year three, so it needs a time-weighted coefficient rather than a fixed one. Most static scoring models cannot express that, which is one reason they underperform.

Where it lives: your onboarding or project tracker.

5. Seat density

Active seats against paid seats. A high paid-to-inactive ratio means the product was never embedded in a workflow, which turns the renewal into a budget-line decision. Shadow seats read as expansion opportunity and churn risk at the same time, and the difference is whether anyone ever activated.

Where it lives: your own product, but as a ratio rather than a count. Raw active-user counts scale with account size and tell you nothing comparable across segments.

Why the weighting matters more than the inputs

Adding five inputs to a weighted-sum model produces a more elaborate score, not necessarily a better one. Three specific mistakes account for most of the failure.

Uniform weights across segments. The signals that predict churn in a 20-seat SMB account and a 2,000-seat enterprise account are different. Sponsor continuity dominates in enterprise, seat density and onboarding dominate in SMB. A single weighting applied to both dilutes each. The permutation that works is one model per segment, validated separately.

No time decay. Onboarding completion, first-90-day usage, and implementation health are strongly predictive early and irrelevant later. A fixed coefficient carries a year-three account's ancient onboarding score forever.

Never validating against actual churn. This is the big one. Take the accounts that churned in the last four quarters, run your current score against them as of 90 days before they left, and see what it said. If it said green, the model does not predict, it describes. That backtest is the only honest measure of a health score and it is rarely run.

The calibration check that follows from it: healthy annual revenue churn for B2B SaaS sits in the single digits, so a score flagging 30% of accounts as at-risk is generating noise your CSMs will learn to ignore. Precision matters more than recall here, because a score nobody trusts has no effect on behavior regardless of its accuracy.

Honest gap in this model: none of the five inputs sees an account losing its budget in a reorganization it did not choose. That happens, it is not predictable from any signal you hold, and no amount of input engineering fixes it.

How to add the qualitative layer without a data team

Start with the input that leads by the widest margin and costs the least to capture.

Attach feedback to accounts. Every ticket, call, and message needs an account identifier. Without that join, feedback cannot enter a scoring model at all, and this is the step most teams have not done.

Turn themes into two numbers per account. Volume of negative themes and whether they are recurring or resolved. That is enough to weight. Sentiment scores alone are too noisy to trend at the account level.

Add sponsor responsiveness from data you already have. Days since last inbound contact and reply latency come from email and calendar with no new instrumentation.

Backtest before you deploy. Run the new model against last year's churned accounts. If it does not separate them from renewed accounts 90 days out, the weights are wrong and shipping it will only spread false confidence.

The decision rule: weight leading relationship and qualitative signals above lagging usage signals, weight per segment rather than globally, and trust no score you have not backtested against real churn.

Run that backtest on your current score this week. Most teams find it identifies churned accounts as healthy at 90 days out, which tells you the inputs are the problem rather than the math.

FAQ

What should a customer health score include besides usage?

Relationship continuity, qualitative signal from what customers say, billing and contract behavior, onboarding milestone completion, and seat density. Usage stays in the model as a lagging input. The additions matter because relationship and qualitative signals tend to move 30 to 90 days before usage does.

Why do usage-based health scores miss churn?

Because they measure behavior and churn is a decision. An account can log in normally and hit its adoption thresholds while its sponsor leaves, its budget gets consolidated, or its team quietly builds a workaround. None of that appears in an event stream, and all of it appears in conversations.

How do you get free-text feedback into a health score?

Turn themes into countable inputs per account: volume of negative themes and whether they recur or resolve. That requires two things, a consistent taxonomy that does not drift between quarters, and an account identifier on every feedback record. Sentiment scores alone are too noisy to weight.

How does Enterpret contribute to health scoring?

Enterpret's adaptive taxonomy produces themes derived from the feedback itself, which stay consistent enough to trend and therefore to weight in a score. The customer context graph attaches account, plan tier, and revenue to every record, which is the join that lets qualitative signal enter a scoring model at all.

How do I know if my health score actually works?

Backtest it. Run the current model against accounts that churned in the last four quarters, as of 90 days before they left, and see whether it flagged them. Also check calibration: with single-digit annual revenue churn typical for B2B SaaS, a score flagging 30% of accounts at risk is producing noise.

If your health score cannot explain why it changed, see how Enterpret's adaptive taxonomy turns what customers say into an input you can weight.

‍