PROBLEM
Churn was a lagging indicator, not a warning
At 7.5% quarterly churn against Series B growth targets, churn came up in every board meeting. Usage events, support tickets and seat activity lived in three disconnected systems nobody reconciled into a single risk view. Twelve CSMs covered 150+ accounts each with no way to tell who needed a call this week — so the first real signal was usually the cancellation email itself.
✕7.5% quarterly churn against Series B growth targets, flagged at every board meeting;
✕Usage events, tickets and seat activity sat in three disconnected systems nobody reconciled;
✕12 CSMs covered 150+ accounts each with no way to prioritize who to call first;
✕Renewal risk surfaced at the cancellation email — zero early-warning window.
SOLUTION · 8 WEEKS
From raw usage data to a playbook a CSM can act on
We built a churn-signal pipeline instead of another dashboard nobody opens: usage and billing data feed a risk score, an LLM turns the score into a plain-English account brief, and the brief triggers the right CS playbook automatically in Salesforce.
WEEKS 1–2
Signal audit & data model
Mapped 9 usage and billing signals — logins, feature adoption, seat utilization, support sentiment, invoice delays — into a unified account-risk table refreshed weekly in Snowflake.
WEEKS 3–4
Risk scoring & LLM summaries
Built a weighted risk score plus GPT-based account briefs: a 5-sentence summary of what changed and why, regenerated automatically every Monday for every account above threshold.
WEEKS 5–6
CS playbooks in Salesforce
Mapped 6 risk tiers to 6 playbooks — exec check-in, win-back call, training nudge and others — auto-created as Salesforce tasks with the LLM brief attached.
WEEKS 7–8
Pilot, tuning & rollout
Piloted with 3 CSMs, cut false-positive alerts from 35% to 9%, then rolled out to all 12 CSMs with a live Slack feed and a leadership dashboard.
RESULT · 2 QUARTERS
Churn became something CS could see coming
| METRIKA | PŘED | PO | ZMĚNA |
|---|---|---|---|
| Quarterly customer churn | 7.5% | 5.7% | −24% |
| At-risk accounts flagged 3+ weeks early | 0% | 80% | NEW SIGNAL |
| CS team NPS | 36 | 54 | +18 |
| Net revenue retention (NRR) | 91% | 97% | +6 pp |
| CSM manual account-review time / week | 6 hrs | 1.5 hrs | −75% |
“We used to find out an account was unhappy when the cancellation email arrived. Now the system taps a CSM on the shoulder three weeks earlier, in plain English, with a reason and a next step.”