By the time a customer formally cancels or fails to renew, the decision was usually made weeks or months earlier — churn prediction exists to surface that decision before it becomes final, while there's still time to intervene. This requires shifting from reactive reporting (counting churned customers after the fact) to leading indicators: behavioral and account signals that reliably precede churn, observed early enough that outreach or product intervention can still change the outcome.

Leading indicators of churn

Leading indicators fall into a few recurring categories across most subscription and recurring-revenue businesses:

Usage decline — declining login frequency, declining feature usage breadth (using fewer of the product's features over time), or a drop in a core usage metric relative to the account's own historical baseline (not an absolute threshold, since "normal" usage varies hugely by account).

Engagement decline — falling email open/click rates on product communications, reduced support ticket engagement, declining attendance at scheduled check-ins or training sessions.

Relationship/account signals — champion (the internal advocate who drove the purchase) leaving the company or changing roles, non-response to renewal or QBR (quarterly business review) scheduling attempts, a drop in the number of active seats/users on a multi-seat account.

Explicit signals — low NPS or CSAT scores, support tickets flagged as complaints or escalations, explicit downgrade requests.

Leading indicator categories

Category	Example signals	Typical lead time before churn
Usage decline	Login frequency drop, feature breadth narrowing, core metric below baseline	Weeks to months
Engagement decline	Falling email engagement, skipped check-ins, unresponsive to CS outreach	Weeks to months
Relationship signals	Champion departure, seat count decline, non-response to renewal scheduling	Often the strongest single predictor; can be very close to churn date
Explicit signals	Low NPS/CSAT, escalated support tickets, downgrade requests	Shortest lead time but highest confidence

Building a churn risk score

A churn risk score combines these leading indicators into a single number per account, the same additive-weighting logic used in lead scoring but inverted — instead of scoring toward sales-readiness, you're scoring toward risk. Each indicator is weighted based on how strongly and how early it has historically preceded actual churn in your own customer base, which means a churn score is only as good as the historical churn data used to calibrate it.

The calibration step is what separates a real churn model from a guess: pull a sample of accounts that churned over the last 12-24 months, and for each one, look back at what their leading-indicator values looked like 30, 60, and 90 days before they churned. Signals that consistently moved in a bad direction well before churn across many accounts deserve heavy weight; signals that were noisy or inconsistent should be weighted lightly or dropped.

churn_risk_score_example.txt

text

CHURN RISK SCORE (higher = more at risk)

  Usage drop > 40% vs. trailing 90-day baseline      +25
  No login in 21+ days                                +20
  Champion/primary contact changed role or left        +30
  Seat count declined in last 60 days                 +20
  NPS score <= 6 (detractor) in last survey            +15
  Missed or didn't attend last scheduled QBR            +10
  Support ticket flagged as escalation/complaint        +10
  Opened last 3 product emails (engaged)               -15
  Used a new feature in the last 30 days               -10

THRESHOLDS
  0-29   = Healthy
  30-59  = Watch (flag for CS review, no action required yet)
  60+    = At-risk (triggers intervention workflow)

A single bad signal rarely means churn; a cluster usually does

Treating any one indicator as a definitive churn signal produces constant false alarms — usage naturally dips during holidays, a champion changing roles doesn't always mean disengagement, one bad NPS score can be a bad day. The risk score's value comes specifically from combining multiple independent signals, so that a cluster of simultaneously declining indicators is a much stronger and more trustworthy signal than any single one in isolation.

Intervention timing: too early vs. too late

Once risk scoring identifies an at-risk account, timing the intervention is its own hard problem with real failure modes in both directions.

Too early intervention — reaching out the moment a single soft signal appears (a slightly lower login count one week) — trains customers to associate any minor dip with an alarmed vendor outreach, which reads as needy or presumptuous, and wastes CS capacity on accounts that would have recovered on their own.

Too late intervention — waiting until the risk score crosses a high threshold or until a cancellation request is already submitted — often arrives after the internal decision has effectively been made, when even a strong save offer has low odds of reversing it. By the time a customer formally requests cancellation, that conversation is a retention negotiation, not a proactive save; proactive saves depend on reaching the customer while they're still undecided.

Intervention timing trade-offs

Timing	Risk	What it requires to avoid the failure mode
Too early (single soft signal)	False alarms, wasted CS effort, customer feels surveilled	Require multiple corroborating signals before triggering outreach
Right window (sustained risk cluster, pre-decision)	Requires accurate, well-calibrated scoring to find reliably	Historical calibration of what the risk score looked like before past churns
Too late (post-cancellation-request)	Decision often already made; low save-rate even with strong offers	Catch the cluster of signals well before the explicit cancellation signal appears

The right intervention point is found by looking backward

The same historical-cohort analysis used to calibrate indicator weights also answers the timing question: looking at accounts that were successfully saved after an intervention versus accounts where the intervention failed, the saved accounts' interventions typically happened at a specific point in the risk trajectory — often when the score first crossed into a middle band and before it reached a high band. That inflection point, not an arbitrary "reach out as soon as possible" instinct, is where intervention timing should be calibrated to.

What's next

Not every at-risk account can be saved, and some will churn anyway — understanding how to re-engage those customers after the fact, rather than only trying to prevent churn beforehand, is the complementary discipline of win-back campaign design.

Next: Win-Back Campaign Design →

I build these systems professionally.

Whether it's a RAG pipeline, analytics migration, or AI workflow — let's talk.

Work With Me →

Need production AI or MarTech expertise? I build custom LLM workflows, server-side GTM tagging, and analytics infrastructures professionally.

Need custom AI or MarTech setup? Let's build together.

Let's talk →