kai-retention

Objective

A retention system the business can run: a churn diagnosis that names which churn type is actually costing money, an engagement health score with defined signals and thresholds, intervention playbooks per risk tier, a win-back sequence, involuntary-churn prevention, and a 90-day implementation roadmap with the metrics to watch.

Diagnose before prescribing. A loyalty program aimed at voluntary churn does nothing when the real leak is failed payments, and a dunning sequence does nothing when customers leave in week three because onboarding never delivered value.

Done when

Work type strategy-plan — floor E3/C3/O1 (harness/eco-floors.yaml). The system is a plan; nothing leaves the workspace until a human sends it.

• E3 — a named human approved the retention playbook, the engagement scoring spec, and the email sequences.
• C3 — banned_word_check clean on all customer-facing copy, every email sequence at 10+/16 on Four U's, and a non-author read the system end to end. Max 2 auto-retry cycles on gate failures for email content.
• O1 — the plan names its metric with baseline, threshold, window, and owner: monthly churn rate, cohort retention curve, health score distribution, NPS trend, and expansion versus contraction revenue.

Sending any sequence is separate work under email-lifecycle (E5/C3/O3), where unsubscribe and sender-identity compliance is a C4 field-standard item, not a lint.

Constraints

• Read MARKETING.md from the project root first. If it does not exist, build it from the codebase — CLAUDE.md, README.md, PROJECT.md, package.json, landing pages, email/ad/analytics config — using the template from /kai-email-system, and confirm the draft. Do not ask the user what the product is.
• Seven things must be known before diagnosis: business model (SaaS, ecommerce, services, marketplace); current churn rate; customer count and average revenue per customer; retention efforts already running; known churn reasons from exit surveys, support tickets, or the cancellation flow; whether product usage and feature adoption are tracked; and customer segments (free vs. paid, plan tiers, cohorts).
• No banned Tier 1 words in any customer-facing copy.
• Win-back emails comply with CAN-SPAM — see harness/references/cold-email-rules.md.
• Loyalty rewards must not erode margins below profitability. Redemption options drive retention, not margin destruction.
• Discounts in rescue plays cap at 20% unless the user approves higher, and all email sequences target 10+/16 on Four U's before they count as deliverable.
• Match the prescription to the maturity level. Predictive churn scoring proposed to a Level 0 business is a plan that never gets built.

Level	Current state
0	No retention effort beyond the product itself
1	Basic cancellation flow plus occasional check-in emails
2	Lifecycle emails, usage tracking, support triggers
3	Predictive churn scoring, proactive intervention, loyalty program

Context

Need	Load
Retention mechanics, churn tactics, loyalty design	knowledge/playbooks/customer-retention.md
Retention as a growth loop	knowledge/playbooks/growth-loops-applied.md
Lifecycle email structure and triggers	knowledge/channels/email-lifecycle.md
Which persona is churning	knowledge/personas/_persona-index.md
CAN-SPAM compliance for win-back sends	harness/references/cold-email-rules.md
Lifecycle email format contract and gate minimums	harness/skill-contracts/email-lifecycle.yaml
Product, ICP, voice, current channels	MARKETING.md (project root)

Churn types — separate them before proposing anything. Voluntary: the customer actively cancels (dissatisfaction, budget, switched). Involuntary: payment failure, expired card, billing issue. Passive: stops using without cancelling (ghost users).

Churn timeline — most exits cluster: first 30 days (onboarding failure), 60–90 days (value not realized), at renewal (annual decision point), or after a price increase or feature change. Leading indicators: login frequency decline, feature usage drop, support ticket spike, NPS/CSAT decline, billing page visits.

Engagement score (0–100) — the rubric:

Signal	Weight	Scoring
Login frequency (last 14 days)	25%	Daily=100, Weekly=60, Monthly=20, None=0
Core feature usage	25%	Used all=100, Used some=50, Used none=0
Support interactions	15%	Positive=80, Neutral=50, Negative=20
Account expansion signals	15%	Upgraded=100, Stable=50, Downgraded=10
Referral/advocacy	10%	Referred=100, NPS promoter=60, Passive=30
Billing health	10%	Current=100, Late=30, Failed=0

Risk tiers and the play for each:

Tier	Score	Play
Red	0–39	Immediate rescue: personal outreach within 24 hours, a concession (discount, extended trial, premium support), escalation to customer success, win-back sequence
Yellow	40–69	Proactive nurture: usage tips for unused features, office hours or webinar invite, relevant case study, a short feedback survey (not NPS)
Green	70–100	Expansion and advocacy: referral or testimonial request, early access, advisory board or beta invite, relevant cross-sell or upsell

Win-back for already-churned customers: a 3-email sequence at Day 1, Day 7, Day 30, each addressing a different churn reason, each carrying a specific offer or product update.

Involuntary churn prevention: a dunning sequence of 3–5 emails over 14 days, smart retry logic for failed payments, and a card-update reminder before expiration.

Loyalty program, where the model supports one: reward mechanics (points, tiers, milestones, or referral credits), earning actions mapped to business goals, redemption options that drive retention, and a launch communication plan.

Gates, on every email file:

python scripts/quality_gates/banned_word_check.py <file>
python scripts/quality_gates/four_us_score.py <file>

Escalate when

• Churn rate, cohort data, or usage data is unavailable and the diagnosis would be guesswork.
• The stated churn reason from the user conflicts with what exit surveys or support tickets show.
• A rescue play needs a discount above 20%.
• The proposed loyalty mechanics would push unit economics negative.
• Churn is driven by the product rather than by marketing — say so; a retention campaign cannot fix broken onboarding.
• Win-back targets contacts whose consent basis or suppression status is unclear.