Churn doesn't announce itself. My Monday Slack digest does it instead.

The Monday digest

Every founder finds out about churn the same way: the cancellation email. By then the user has been gone for weeks - the cancel is the paperwork, not the decision. The decision happened earlier, quietly, in their usage: the daily user who went weekly, the power feature that stopped appearing, the API calls that flatlined. That decay is fully visible in metering data. The problem was never detection - it is that nobody reads per-user usage tables on a Monday morning. So I made the tables write me a memo. Vevee meters every user's AI usage (@vevee/sdk), and its compose() generates structured output from any user's history - including a churn-risk assessment. A Monday cron runs it over my paying users:

import { createClient } from "@vevee/sdk";

const vevee = createClient({ apiKey: process.env.VEVEE_SECRET_KEY! });

interface ChurnAssessment {
  risk: "healthy" | "cooling" | "at_risk" | "gone_quiet";
  evidence: string[];      // concrete usage facts, with numbers
  probableCause: string;   // best inference from the pattern
  suggestedAction: string; // one thing, doable this week
}

const flagged: Array<{ email: string; a: ChurnAssessment }> = [];

for (const user of payingUsers) {
  const result = await vevee.compose<ChurnAssessment>(
    "churn-assessment",
    user.id,
    { plan: user.plan, mrr: user.mrr }
  );

  if (result.status === "generated" && result.output.risk !== "healthy") {
    flagged.push({ email: user.email, a: result.output });
  }
}

await postToSlack("#health", formatDigest(flagged)); // sorted by MRR at risk

What Monday looks like now

Here is what lands in #health. Two things make this readable where a dashboard is not. First, the comparison is to each user's own baseline - 50 events/week is healthy for one account and a 90% collapse for another; absolute thresholds cannot see that, the per-user composition can. Second, it arrives ranked and reasoned - not "engagement score: 42" but "she failed three times and stopped," which is a sentence a human acts on.

• #health - Week 24 · 3 accounts flagged · $94 MRR at risk
• dana@- (Pro, $15) - at_risk. Was 40 generations/week for 2 months → 4 last week → 0 this week. Last session ended on three failed style: "anime" requests in a row.
• Probable cause: hit a quality wall on her main use case, not loss of interest - frequency was stable until the failures.
• Action: personal email asking about the anime-style results; fix or workaround likely saves the account.
• team@- (Max, $49) - cooling. API volume -30% over 3 weeks; dashboard logins unchanged. Pattern fits their end-users dropping, not them leaving.
• Action: offer a 20-min call about their conversion funnel - their churn is your churn one step removed.

Close the loop or it's theater

Each flagged account gets the suggested action; the outcome is one analytics event away (same SDK, reserved event names). After a couple of months you know your save rate per risk tier and per action type - which converts "founder gut feel" into the retention playbook you will someday hand a hire.

await vevee.capture({ distinctId: userId, event: "feature_used",
  properties: { source: "churn_save_action" } });
// vs. the eventual subscription_cancelled if the save fails

Cost and trust

A weekly sweep across a few hundred paying users costs cents - every result reports usage.costMicroUsd, and the dashboard's monthly compose budget hard-caps the spend. Users who opted out of AI personalization return opted_out and simply don't get assessed; for a consent-scoped system that's the correct trade, and you can still eyeball their raw charts by hand. The cancellation email is a lagging indicator with a 100% detection rate and a 0% save rate. The Monday digest is the opposite trade. Take it.