A customer health score is a composite metric that estimates the likelihood a customer will renew, churn, or expand. Most health scores fail for one reason: they are built from lagging indicators that describe what already happened, so by the time the score turns red, the decision to leave has already been made. A predictive health score is built from leading indicators and tied to actions, not dashboards.
Why Most Health Scores Fail
The typical health score averages login frequency, support ticket volume, and NPS into a single color. Each input is lagging. Logins drop after users have disengaged. Tickets stop when the customer stops caring enough to complain. NPS captures sentiment months after the sentiment formed. A score built on these tells you a customer churned before the invoice does, which is not prediction. It is narration.
The second failure is averaging. A customer with perfect usage but a departed champion is at severe risk, yet an averaged score shows yellow. Averages smooth away exactly the signals that matter.
The Four Pillars of a Predictive Score
Build the score on the same four pillars that drive retention itself: onboarding, usage depth, success criteria, and stakeholders. Each pillar contributes leading indicators, and any single pillar failing should be able to turn the account red on its own.
Pillar 1: Onboarding velocity
Churn is often decided in the first ninety days. Measure time to first value against your own benchmark: days to go live, days to first real workflow completed, percentage of purchased seats activated in month one. An account that is three weeks behind your median onboarding pace is a churn risk twelve months before the renewal date, and that is when the signal is actually useful.
Pillar 2: Usage depth, not usage volume
Logins measure attendance. Depth measures habit. Track how many of the workflows the customer bought are actually in weekly use, whether usage is spreading to new users or concentrating in fewer, and whether the features tied to the customer's original business case are the ones being used. A customer using one feature heavily is one reorg away from churning.
Pillar 3: Success criteria progress
Every customer bought to achieve something. If those success criteria were captured at closing, the health score can track progress against them: is the customer measurably closer to the outcome they bought? If success criteria were never captured, that absence is itself a red flag, and the first CS motion is to establish them retroactively. A customer achieving their outcome tolerates bugs, price increases, and reorgs. A customer who cannot connect your product to a result tolerates nothing.
Pillar 4: Stakeholder coverage
The strongest usage cannot survive losing the only executive who sponsored the purchase. Track named champion and executive sponsor, whether each has engaged in the last sixty days, and single threading: accounts where every relationship routes through one person. Champion departure is among the most predictive churn events in B2B software, and it is fully detectable the day it happens if your system watches for it. It is also the most reliable precursor in our own book: a change in the decision maker or champion, whether leaving the company or switching roles, has preceded renewal risk more consistently than any usage signal we track.
Scoring Mechanics That Keep the Score Honest
Use gates, not averages. Score each pillar separately and let the worst pillar cap the overall score. A red pillar means a red account, whatever the other three say.
Weight by time to renewal. Onboarding signals dominate early life; stakeholder and success criteria signals dominate the two quarters before renewal.
Backtest before you trust it. Run the score against the last eight quarters of churned and renewed accounts. If the score would not have flagged your actual churn at least a quarter early, adjust inputs until it does.
Attach an action to every state change. A score that turns red should create a play: an executive outreach, a success review, a re onboarding. A score nobody acts on is a dashboard decoration.
Where AI Changes the Mechanics
The hardest part of health scoring has never been the formula. It is the data collection: stakeholder changes live in email threads, sentiment lives in call recordings, success criteria live in a closed deal from a year ago. Historically that meant CSMs typing updates into fields, which meant the score ran on stale data. An AI native system reads the calls, emails, and meetings directly and keeps the pillars current without manual entry, which is the difference between a score that is theoretically predictive and one that predicts.
This is how health scoring works in Dreamhub: the four pillars are native objects, and Dreamer keeps them updated from the communication stream itself.
FAQ
What is the difference between leading and lagging indicators in a health score?
Leading indicators change before the renewal decision forms: onboarding pace, champion engagement, success criteria progress. Lagging indicators change after: login drops, ticket silence, low NPS. Predictive scores are built on the former.
How many inputs should a health score have?
Eight to twelve, organized under the four pillars. Fewer misses risk categories; more turns the score into a black box nobody trusts or debugs.
Should health scores be shared with the customer?
The underlying facts, yes: success criteria progress and adoption make excellent executive business review content. The internal risk rating itself, generally no.
How often should the score refresh?
Continuously if your system automates data capture, weekly at minimum if it does not. A monthly health score in a business with a 30 day sales cycle for churn decisions is a rearview mirror.