Churn signal in support conversations

When to use it

Reach for this when someone says any of these — they are the phrases the skill itself triggers on:

• “which customers are at risk of leaving”
• “find cancellation intent”
• “customers asking how to export their data”
• “top reasons customers asked to close their account”

How it works

The method, in the order the skill runs it. The full procedure — tables, worked examples and the edge cases — is in the skill itself.

1. Most churn never appears in support

   The majority of customers who leave never contact you about it. They stop using the product, the subscription lapses, and the first record of the decision is a billing event.

2. Three signal classes with opposite error profiles

   1. Explicit intent — asking to cancel, close, downgrade, or not renew. High precision, low recall, and usually too late to act on. By the time someone asks how to close the account, the decision is generally made. Still worth capturing for reason coding.

3. Lead time is the metric, not accuracy

   A detector that fires the day someone cancels is a very accurate, completely useless detector.

4. The base rate will ruin a naive classifier

   Churn is rare per period. At a 2% monthly churn rate, a detector with 90% sensitivity and 90% specificity produces roughly one true positive for every five false positives — the precision is about 15%, not 90%.

5. Validate against outcomes, not sentiment

   The signal must be labelled against what actually happened — did the account churn or downgrade — which means joining support data to billing or CRM outcomes. If you cannot make that join, you cannot validate the signal, and you should say so rather than validating against a proxy.

6. Reason coding is the durable output

   The at-risk list decays; the reason taxonomy compounds.

Related skills

• Upsell signalsCatch upgrade hints nobody is routing anywhere
• Revenue at riskSize the revenue sitting behind unresolved problems
• Customer health scoreAdd support signals to a health score that predicts something
• Crediting support for NRRWork out what support can honestly claim credit for
• Renewal risk reviewPull an account’s whole support record before a renewal
• Why customers leavePull competitor and switching evidence out of support