Customer Health Scores for SaaS: How to Predict and Prevent Churn

A health score that just tells you who is red is half the value. The full value is using the score to predict churn risk on a rolling 30/60/90-day horizon and intervening before the customer asks to cancel.

Most SaaS teams stop at the dashboard. They build the weights, color-code accounts, send a Slack alert when something drops to red, and wait for the CSM to react. That works only when red is already late — when the customer has already drafted the cancellation email or scheduled an internal call to evaluate alternatives. The intervention misses by weeks.

The predictive version of the same score catches the customer 60 to 90 days earlier, when the signals are quieter but the trajectory is clear. The math is not machine learning. It is a few weighted signals, a forward-looking trend window, and a triage cadence that maps every score band to a specific human action. This guide walks through exactly how to build that — the signals, the model, the playbook, and the failures that look like progress but quietly bleed retention.

A detection score answers a present-tense question: is this customer healthy right now? You take the latest snapshot of usage, support, and billing, run them through weights, and output a number.

A prediction score answers a future-tense question: where is this customer likely to be in 30, 60, and 90 days? You take the same signals, but you measure their trend over time and project forward. The difference is small in math and enormous in outcome.

A detection score sees a customer with 200 active users today and labels them green. A prediction score sees that same customer had 240 active users 60 days ago, 220 thirty days ago, and 200 today — a 17% engagement drop on a clean linear slope — and flags churn risk for 60-90 days out, while the relationship is still warm enough to save.

By the time a detection score turns red, the customer has often already made the internal decision to leave. Predictive scoring buys you the window where intervention still works.

You do not need 30 inputs. You need 5 that move ahead of churn instead of with it. These are the signals that operators across B2B SaaS consistently report leading cancellation by 1-3 months.

Active user count and login frequency are the obvious metrics, and the obvious mistake is using them as levels rather than trends. A customer with 50 weekly active users could be healthy or terminally ill depending on whether that 50 is up from 30 or down from 80.

Measure the slope: weekly active users this period vs. the rolling 8-week median. A drop of more than 15% on that ratio for two consecutive weeks is the most reliable single early-warning signal in B2B SaaS. It typically appears 45-90 days before a cancellation email.

Segment by license tier. A customer paying for 100 seats with 60 active is leaking. A customer paying for 100 seats with 95 active and dropping is leaking. The seat-utilization ratio matters more than raw login counts because it directly maps to the renewal conversation: if they are not using what they bought, they will buy less or nothing.

Most teams track ticket volume and CSAT. Both are weak predictors on their own. A customer with no tickets could be perfectly happy or quietly disengaged. A customer with many tickets could be deeply integrated and demanding, or actively suffering.

The predictive variant is the shape of ticket activity. Look for two patterns:

- Sudden silence after sustained engagement. A customer who logged 3-5 tickets per month for six months and goes to zero for 30 days is rarely satisfied — they have stopped trying. - Escalation in tone or seniority. When the person filing tickets shifts from a power user to a director or VP, you are watching the renewal conversation start without you. Tag tickets by sender role and watch for the executive escalation curve.

If you have an LLM-powered helpdesk, sentiment classification on inbound tickets adds a layer: a quiet but increasing share of negative-sentiment tickets is a 60-day warning even when CSAT scores stay flat.

Customers who churn rarely just stop using the product. They first stop using the edges of the product — the integrations, the advanced reports, the workflows their team set up six months ago. The core feature usage often holds until the very end.

This is why feature breadth is more predictive than feature depth. Count the number of distinct features a customer touches in a rolling 30-day window. When that count contracts by 30% or more relative to their 90-day baseline, you are watching a customer simplify their relationship with you in preparation for leaving it.

The nuance: define features at the right grain. Counting button clicks is noise. Counting meaningful product surfaces — integrations connected and actively syncing, reports run, automations executed, mobile sessions, API calls — is signal.

The single biggest non-product churn driver in B2B SaaS is the person who championed the purchase leaving the company. New leadership reviews vendor contracts, runs build-vs-buy debates, and wants to put their stamp on the stack. If your champion left and you did not know within a week, you are already behind.

This signal does not come from your product. It comes from LinkedIn job-change data, email bounces, and the moment a CSM cannot get a meeting with the original sponsor anymore. Operators who track this systematically — through CRM enrichment or a job-change monitoring tool — get a 90-120 day head start on the renewal risk that follows a sponsor change.

The playbook when the signal fires: identify the new decision-maker within 14 days, run a re-discovery session with their priorities, and produce a fresh ROI summary that ties your product to the new leader's stated goals — not the old champion's.

Finance teams know what customer success teams often do not: a 30-day-late invoice is a churn signal even when the customer assures you the AP delay is purely administrative. The pattern that matters is change: a customer who has paid on time for 18 months and suddenly pays 21 days late is signaling friction somewhere — internal budget review, executive scrutiny, a cancellation conversation that has not surfaced yet.

Other billing signals that predict churn:

- Removing seats below initial purchase quantity - Declining the auto-renewal opt-in or pushing renewal to monthly - Asking for an itemized usage report (the precursor to a downsize negotiation) - A new procurement contact appearing on a previously direct relationship

None of these guarantee churn. All of them shift the probability meaningfully. They are also some of the most underweighted inputs in standard health-score models — billing data lives in finance, the score lives in customer success, and the two systems often do not talk.

You do not need a data scientist. You need a weighted average that updates daily and a calibration loop that adjusts the weights every quarter based on what actually predicted churn in your last cohort.

Start with a simple linear model. Each signal gets a weight that sums to 1.0. Each signal is normalized to a 0-100 scale where 100 is best and 0 is worst. The output is a single 0-100 risk score, where lower means higher churn risk.

Two adjustments matter once you have the basic model running.

Segment by ARR tier. A $500/month customer and a $50,000/month customer should not run on the same weights. For SMB accounts, engagement decline carries more weight (no champion structure to lean on). For enterprise accounts, executive sponsor and billing signals weigh more (the contract is the relationship). A reasonable split: SMB uses 40/20/15/10/15, enterprise uses 20/15/15/30/20.

Calibrate quarterly. Pull the last quarter's churned accounts. Look at where their score landed 30, 60, and 90 days before cancellation. If your model was already showing them as risk by day 90, your weights are roughly correct. If most churns came from accounts that scored healthy at the 60-day mark, increase the weight of whichever signal showed a divergence first. Two or three calibration cycles will get you to a model that catches 70%+ of churn 60+ days out.

A score that does not trigger a defined human response is not a system. It is a dashboard. The intervention playbook below ties every score band to a 30/60/90-day rolling action — same cadence, different intensity by tier.

Most CS teams measure churn rate and stop. That tells you whether last quarter's interventions worked in aggregate; it does not tell you which interventions did the work. Three measurements matter.

NPS is a satisfaction snapshot from the people willing to answer surveys. It is biased toward champions and biased away from the operational users whose disengagement actually drives churn. Use NPS for product feedback. Do not let it carry more than 10% weight in a churn-prediction model — and arguably do not include it at all.

Customer success teams often anchor scores on what they can see directly — emails, calls, NPS, QBR attendance. Product engagement lives in a different system, and when integration is hard, it gets reduced to a token weight or dropped. The result: scores that move with relationship quality but not with whether the product is being used. Engagement trend is the single most predictive signal in B2B SaaS — give it the weight it earns, even if it requires a data engineer to wire up the pipeline.

Late invoices, seat removals, and procurement involvement are some of the loudest churn signals in any portfolio, and they are routinely missing from the score because the data sits in finance. Connect the billing system to the CRM (or use a platform where they share a database from day one), and let the model see the signals that are already obvious to your CFO.

If your dashboard shows the current score but not the trend, you have built a thermometer, not a forecast. The score must be stored as a daily time series, not overwritten. The 60-day slope of the score is more predictive than the score itself. Without history, you cannot run the cohort recovery measurement above, and you cannot calibrate the model — you are flying with the same weights you set on day one.

A red account is not a lost account; it is an account that needs a defined response in the next 48 hours. Teams that treat red as a cancellation queue stop investing in the playbook because the outcome looks predetermined. Teams that treat red as a 90-day rescue window post recovery rates above 30% — meaning roughly one in three near-cancellations gets saved, which over a year is meaningful ARR protected.

Most SaaS teams hit one structural problem: the signals live in five different tools. Engagement is in product analytics. Tickets are in the helpdesk. Sponsor changes are in the CRM. Billing data is in Stripe or finance. Email response patterns are in the email tool. Stitching them into a single rolling score requires either a data engineer or a months-long Zapier graph that breaks every time a tool updates its API.

Deelo collapses the systems. CRM, helpdesk, analytics, invoicing, and email run on the same database, which means a customer health score can reference engagement, ticket sentiment, executive contact stability, and billing health without an integration layer.

In practice that looks like:

- A custom field on the company record stores the daily health score, written by a scheduled workflow that reads from CRM, helpdesk, analytics, and invoicing. - A daily automation compares today's score to the 30-day average; a drop of 10+ points emits an event. - That event drives the playbook: a CSM task is created, a Slack alert fires, an executive briefing email is queued for accounts above an ARR threshold, and the account is added to a rescue cohort for cohort-level reporting. - The Analytics app charts score-by-cohort, recovery rate, and save velocity without anyone exporting a CSV.

The point is not the tooling. It is that a predictive churn system needs the data unified and the actions automated. When the signals and the playbook live in the same workspace, customer success teams stop debating data quality and start running the loop that actually retains revenue.