Match the intervention to how at-risk a user actually is

Create an AI-derived attribute 'churn_risk_score' on the User object, refreshed daily. Inputs: engagement velocity (sessions trend over 30/14/7 days), feature usage decay, support sentiment trajectory, billing-status signals, peer-cohort churn patterns, account-level usage (for B2B). Output: numeric 0-100 with tier label — low (0-29) / medium (30-59) / high (60-79) / critical (80-100). NOT binary at-risk-or-not — the nuance is the point: medium and high get different treatments. Refresh on every significant behavioral event so a sudden engagement drop is caught fast.

Build a parent segment 'Churn risk - paying users' capturing all paying users with churn_risk_score >= 30. Implicitly partitioned by tier through the attribute. Excludes users in the first 14 days (early signals aren't reliable yet) and users already in active save flows (no double-intervention).

Generate per-tier email content. Medium tier: educational content matched to underutilized features that drive retention in their cohort + personalized 'pick up where you left off' deep-link. Tone: helpful, not alarmed. High tier: urgent value reinforcement, CSM availability signal, specific use-case help offer ('Want a 15-min walkthrough on [feature]?'). Critical tier: founder/CEO personal letter ('I noticed your account hasn't been used much — anything we can do?'), with direct calendar link and concrete asks. Send-from: success@ for medium, named CSM for high, founder/CEO for critical.

Generate in-app messages for high-tier users when they DO log in (rare and precious moments). Format: contextual banner referencing their most-likely-success-path ('Most users in your segment succeed by trying [feature] — want a 60-second walkthrough?'). For critical-tier users on rare logins: a personal-touch interrupt — full-screen modal from the CSM or founder offering a 1:1 call. Never blanket — only when AI signal is strong.

Configure a recommendation surface specifically for high+critical tier users: surfaces underutilized features that THIS user's cohort uses successfully but THIS user has not adopted (the 'features-that-stick-users' pattern). Renders in-app on dashboard, in email content blocks, and on personalization slots when the user does visit. The recommendation logic prioritizes features with high retention-correlation in this user's segment.

Build a tiered journey wired to the churn-risk segment. Branch on churn_risk tier at entry AND re-evaluate weekly: LOW tier — light-touch nurture email at Day 7 (no urgency, just value content). MEDIUM tier — Touch 1 email Day 0 (helpful tip), Touch 2 in-app at next login (deep-link to retention feature), Touch 3 email Day 7 with recommendation surface highlighting cohort-success features. HIGH tier — Touch 1 email Day 0 from CSM with offer of help, in-app urgent banner on next session, CSM task created same-day. CRITICAL tier — within 4 hours: founder/CEO personal email + agent handoff offering 1:1 call + urgent CSM task + executive-sponsor task. Tier RECOMPUTED weekly — users de-escalate (engagement returned) exit gracefully; users escalate get tier-appropriate next-touch. Exit on: churn_risk drops below 30 for 14+ days (recovered — log retention_win), subscription_canceled (handoff to post-cancel-winback), or unsubscribe.

Compose a tiered churn intervention dashboard: distribution of users across tiers (low/medium/high/critical) per week — leading indicator of book-of-business health, tier-level save rate (% of users in each tier who exit risk vs. churn — proves tiered strategy works), critical-tier CSM response SLA (target: <4hr for human-touch start), ARR-weighted save value by tier, and tier-migration tracking (high → medium = success, medium → high = early warning). Compare model accuracy: predicted churn vs. actual churn at 30/60/90 days.