Menu
In short. A customer health score aggregates several signals, such as product usage, engagement, support, contract and sentiment, into a single indicator of how likely an account is to renew, churn or expand. The catch: a good score should predict a customer's departure, yet most only record it too late. Customer Success Platforms fall into two families. Deterministic platforms (HubSpot, Planhat, ChurnZero…) rely on rules and thresholds set by hand, based on what has already happened. Self-calibrating AI platforms (Skalin, Hook, Customerscore.io…) learn from your own churn patterns and predict future behaviour from the past. Rules offer transparency and control, provided you have the data resources to know which rules actually predict churn. For most mid-market Customer Success teams, AI builds a reliable score faster and catches signals that fixed thresholds miss.
An account is green on your dashboard. The team logs in every day, the CSM has flagged nothing alarming, the health score shows a reassuring dot. A few weeks later, that customer doesn't renew. The scene is so common it has come to define the whole debate around health scores: almost everyone uses one, but very few actually predict churn.
The problem comes from two places at once: the signals you choose to track, and the method that turns them into a score. And it is now measured in black and white: in the Customer Revenue Leadership study run by ChurnZero in 2025 among nearly 800 post-sale leaders, 73% say their current health score does not reliably predict churn. In other words, three teams out of four are steering with an instrument that fails at its main job.
This article first sets out what a good health score is, then compares the two approaches CSPs use to build one, before giving a clear decision criterion.
A customer health score is a composite score: it aggregates several families of signals, such as product usage, engagement, support tickets, contract data and customer sentiment, into a single indicator meant to reflect how likely an account is to renew, churn or expand. The idea isn't new; what has changed is how much is now expected of its reliability.
The most important distinction comes down to two verbs: record and predict. A bad score is a rear-view mirror: it turns red once logins have already dropped, which is often once the customer has already mentally decided to leave. A good score is a leading indicator: it catches weak signals while there is still time to act. That is the whole difference between a dashboard that documents churn and a dashboard that helps prevent it.
In practice, a well-built health score serves three purposes: prioritising at-risk accounts so CSMs focus their attention where it counts, triggering actions (playbooks, outreach, check-ins), and forecasting renewals and NRR with a little lead time.
Here's the counter-intuitive part. If scores fail so often, it isn't only a question of method: it's that the signals people believe are predictive aren't, and the ones that truly predict churn are often the opposite of business intuition.
The most striking data on this comes from ChurnRX (Greg Daines), based on cohorts of tens of thousands of SaaS customers and expressed as half-life, the time it takes for half of a cohort to churn. To give just two examples:
The rest of these counter-intuitive factors (discounting, monthly billing, downsell, onboarding…) is covered in our dedicated article on the lesser-known drivers of SaaS churn.
The lesson isn't that these exact figures should become your rules; they come from other bases than yours. It's that business intuition is a poor guide for choosing a score's signals. If the best predictors are often counter-intuitive, then a score built by hand, on rules that "seem logical", mostly reflects the assumptions of whoever wrote it, and misses the real signals for months.
Customer Success Platforms fall into two broad families depending on how they compute customer health. The real divide isn't "with or without AI" (nearly all have added AI layers), but how the score is determined: by hand-written rules, or by a model that calibrates itself on your data. Here are both, treated fairly.
This is the historical approach, used by HubSpot, Planhat, ChurnZero or Totango. The principle: the user defines the conditions. "If fewer than N users log in for Y days, remove X points"; "if NPS drops below a given threshold, turn the account amber"; and so on. The score is the weighted sum of rules and thresholds written by hand.
Its merits are real and shouldn't be dismissed. A rule-based score is transparent (you know exactly why an account is red), controllable (you adjust a rule knowingly) and explainable to a leadership committee. For a company that is very mature in how it measures customer health, that has already analysed its past churns and has data resources, it's a perfectly defensible approach, and sometimes the most relevant one.
Its defining trait, however, is also its main weakness: the score is based exclusively on what has already happened. It compares the account's current state to fixed thresholds, without anticipating any trajectory. Its limits come down to four points:
The recent approach, adopted by players such as Skalin, Hook or Customerscore.io, flips the logic: instead of asking you to write the rules, the score learns from your own churn patterns. Its fundamental difference from the deterministic approach: it doesn't just record the past state, it predicts future behaviour from the past. An account declining gradually, or whose usage turns erratic, is spotted on its trajectory, not on crossing an arbitrary threshold. Three principles define it.
First, benchmarking without preconceptions: the AI continuously scans a large number of indicators across the whole portfolio, segment by segment, without assuming in advance which ones matter. Second, trajectory awareness: an account can be below average but improving quickly, in which case it is valued, where a static rule would penalise it. Third, a self-learning model: when a customer churns, the AI re-weights the variables that proved most discriminating, so the score sharpens over time instead of staying frozen.
This approach has two practical strengths. It covers a broad spectrum of signals from the start; at Skalin, four families: interactions (with multilingual sentiment analysis on emails and tickets), product usage, contract data and the "CSM Pulse" (the CSM's qualitative read on the account). And it is quick to get running: two to three weeks of usage data are usually enough to produce a reliable score in under a month, with no rule-configuration phase.
Beware a common misconception: "self-calibrating" doesn't mean "frozen" or "black box". A good AI score stays configurable: you can add your own business rules to bring in a specificity you know about (a contractual signal, an event specific to your product), and adjust the weighting of the different criteria to reflect what matters most in your business. The AI provides the self-calibrated baseline; you keep control to fine-tune it. We detail this in our article on how AI helps implement a customer health score.
Here is how the main Customer Success Platforms position themselves, according to what each one highlights on its site.
| Platform | Scoring approach | Positioning |
|---|---|---|
| HubSpot | Deterministic (manual rules and thresholds, no product usage) | All-in-one CRM, Service Hub |
| Planhat | Deterministic (rules defined by the user) | Enterprise-focused platform, Swedish |
| ChurnZero | Deterministic (highly configurable ChurnScores) | Mid-market / Enterprise CSP, US |
| Totango | Deterministic (configurable rules) | Modular platform focused on large accounts |
| Custify | Deterministic (rules to configure) | European mid-market CSP (Romania) |
| Hook | Self-calibrating AI (predictive agents) | B2B SaaS mid-market to Enterprise, high data volumes, strong data culture |
| Customerscore.io | Self-calibrating AI (predictive scoring) | Product-led SaaS (low-touch CS), small teams, from growing startup up to mid-market |
| Skalin | Self-calibrating AI + optional business rules | French CSP for B2B SaaS mid-market |
Two observations. First: the real divide isn't "AI or no AI", but deterministic (fixed thresholds) vs self-calibrating (learning). Second: these two approaches are not mutually exclusive, and the market consensus is that the strongest setups combine both, a model that automatically finds the patterns, plus the option to add your own rules to fine-tune.
That is exactly the category Skalin sits in: a self-calibrating AI score, operational with no rules to write, on top of which you can layer optional business rules and adjust the weighting of criteria to refine it. You start from a predictive score on day one, and you keep control to adapt it to your specifics, without being forced to build everything yourself or to accept a black box. To go further, see our Top 15 Customer Success Platforms ranking and our detailed comparisons Skalin vs HubSpot, Skalin vs Planhat and Skalin vs ChurnZero.
Neither approach is universally superior. The right choice mostly depends on your resources and your maturity. The table below sums up the trade-offs.
| Criterion | Deterministic (rules) | Self-calibrating AI |
|---|---|---|
| Time to set up | Long (defining + calibrating rules) | Short (reliable score in under a month) |
| Resources needed | Data team / analysis of past churns | No data team required |
| Signal coverage | Limited to written rules | Broad, from the start |
| Reading of time | Records the past state (fixed thresholds) | Predicts the future trajectory |
| Adaptation over time | Constant manual maintenance | Self-learning model |
| Explainability | Total (each rule is readable) | Good, but more complex logic |
One line of this table clearly favours rules: total explainability. If you have a data team able to model your past churns, justify each weighting and maintain the model over time, a deterministic approach gives you a level of control no AI matches on that specific point.
The decision criterion therefore comes down to one question: do you have a data team able to model your past churns and maintain the model over time? If yes, a deterministic approach can work perfectly well. If not, which is the case for the vast majority of mid-market CS teams, typically 1 to 15 people, AI does the job better, faster, and without inheriting the received ideas that weigh manual scores down.
At Skalin, the health score applies the three principles of the AI approach described above: benchmarking without preconceptions across a broad set of KPIs, awareness of each account's trajectory, and a model that re-weights itself as customers churn. It aggregates four families of signals (interactions with multilingual sentiment analysis, product usage, contract and CSM Pulse) and becomes reliable in under a month, from two to three weeks of usage data. It stays fully customisable: weighting by segment, custom business metrics, and business rules you can add optionally, at no extra cost, to refine the self-calibrated score. That is what places Skalin in a category of its own: the predictive baseline of AI, with the control of rules when you need it.
Where Skalin goes further is with Insights, the feature that bridges Customer Success and Product to make the voice of the customer heard. Where the health score works account by account, Insights reads the entire customer base and surfaces the converging signals, including those that preceded past churns: recurring feature requests, bugs reported by several accounts, shared support needs. Each topic is categorised by theme, number of occurrences, accounts affected and share of MRR affected, and a cross-customer AI you can query in natural language lets you ask, for example, what the main churn reasons or the most awaited features are. Concretely, this gives Product and the roadmap inputs to tackle churn at the source: rather than handling departures one by one on the CS side, you identify and fix the structural causes, quantified by the revenue they put at stake. No other tool in the panel offers this aggregated read in this form.
Want to see a health score that predicts churn instead of recording it? Book a Skalin demo: we'll show you the Smart Health Score and Insights on a case close to yours, in under an hour.
A customer health score is a composite indicator that aggregates several signals (product usage, engagement, support, contract, sentiment) into a single score reflecting how likely a customer is to renew, churn or expand. It is used to prioritise at-risk accounts, trigger actions and forecast renewals.
A deterministic score relies on rules and thresholds set by hand: it compares the account's current state to fixed values, based on what has already happened. A self-calibrating AI score learns from your churn patterns and predicts the account's future trajectory. The practical consequence: a gradual or erratic drop in usage often slips under a deterministic score's thresholds, whereas AI catches it on the trend.
Two methods coexist. With a deterministic score, you define by hand the conditions that raise or lower the score (usage thresholds, login frequency, etc.). With a self-calibrating AI score, the model learns from your own churn patterns and automatically weights the most predictive signals. The first needs data resources to know which rules to write; the second produces a reliable score faster, with no rule configuration.
The decisive question is: do you have a data team able to model your past churns and maintain the model over time? If yes, a deterministic approach offers full transparency and control and can work. If not, which is the case for most mid-market teams, a self-calibrating AI score builds a reliable score faster and without relying on often-wrong assumptions. The most advanced approaches combine both: AI first, optional rules to fine-tune.
Usually for two reasons: the chosen signals aren't truly predictive (many scores rely on usage and NPS, whose predictive value is weak or nil), and the method is frozen (rules written once and never recalibrated drift over time). In the ChurnZero 2025 study, 73% of post-sale leaders said their score does not reliably predict churn.
With a self-calibrating AI approach, a reliable score can be obtained in under a month, from two to three weeks of usage data. With a deterministic score, it takes far longer: you have to analyse past churns, write and calibrate the rules, then maintain them continuously as the product and segments evolve.