Who hurts
Customer success leads at SaaS companies with 100–2,000 paying accounts.
The problem
Churn prediction from usage data catches accounts too late — by the time logins drop, the decision is made. The earlier signal is in support tickets, unread in aggregate.
What you build
Score every account weekly on ticket volume, sentiment drift, unresolved threads, and specific phrases that precede cancellation in your own history. Deliver a Monday list of five accounts and the sentence that got each flagged.
Why now
Helpdesk APIs are universal and the analysis is per-ticket cheap. The historical labels — who actually churned — are sitting in the billing system already.
Validate it this week
Backtest on one company's last twelve months. If you'd have flagged the churned accounts and not flagged the healthy ones, they'll buy on the spot.
Why you'd keep winning
Per-customer language models of what 'about to leave' sounds like in their product, tuned on their own outcomes.
The honest risk
Sample size. Under a few hundred accounts there aren't enough churn events to learn from, which cuts out the customers most eager to try it.
The prompt is written to make an AI argue with you before it writes code — that first round of pushback is worth more than the scaffold.