Data Analytics · Read + Write

Churn prediction + outreach

Automate Rep · High Reason · High

The problem

By the time a customer cancels, the save-call is too late. Churn lags 60–90 days behind the disengagement signals.

The AI approach

Predict churn risk from usage + support + billing signals. For high-risk, generate a personalised outreach (CSM email, in-app nudge, or call task).

The outcome

~24% saved

Save rate on flagged customers: ~24%. Most-impactful intervention: a CSM call within 7 days of the flag.

Try itInput → Process → Output

Input — customer signals
Acme Logisticscustomer #C-1138 · CSM: P'Ploy · MRR ฿84k
logins / wk2 ↓ 78%
NPS (last)5 ↓
billingpaid
tickets/30d14 ↑
API calls−62%
contractrenews Aug
Process — AI pipeline
1Aggregate signals (usage / billing / tickets)ReadSymbolic
2Score churn probabilityReadPredictive
3Pick intervention from playbookRulesSymbolic
4Draft personalised outreachWriteGenerative
Output — risk + recommended action
Click Run demo to score the customer and recommend an intervention.
Churn risk · 30-day window HIGH · 0.81
Probability of churn
81%
Recommended intervention: CSM outreach call within 7 days. Talk track: address the 14-ticket spike (mostly API auth issues — see #4521), offer custom integration check-in.
Drafted email (CSM tone): "Hi Khun Anan — noticed the team's been running into API auth hiccups recently. Can I grab 30 min next week to make sure your June launch goes smoothly?"

Three AI types in this use case

SymbolicSignal aggregation pipeline (last-30-day usage drop, NPS dip, ticket volume); intervention playbook (which action for which segment + risk-level).
PredictiveGradient-boosted classifier (XGBoost) on tabular signals; uplift model picks who responds to which save offer.
GenerativeLLM personalises outreach with the customer's actual usage drop and recent ticket history.

The stack

  • Feature store · Feast
  • Model · XGBoost / LightGBM
  • Uplift · CausalML
  • Sink · Salesforce / HubSpot

When this works

  • ≥ 12 months of churn history to train on
  • Intervention playbook is documented
  • CSM team has bandwidth to act on flags within SLA

When it fails

  • Concept drift — model fit on pre-pandemic signals fails post-pandemic
  • Self-fulfilling: aggressive outreach annoys retainers
  • No control group — can't measure uplift
  • Bias: model penalises new accounts for low usage