Problem
Winning new customers in Indonesia is expensive — keeping the ones you have is far cheaper, yet almost nothing does it systematically. Churn arrives quietly: usage drops for weeks before payments stop, credit cards expire, VA bills go unpaid, and then the customer is gone without ever complaining. Local subscription businesses (SaaS, course memberships, standing contracts, franchise supply) only notice MRR leaking when the monthly report closes — by then it is too late to win anyone back.
Recognize this?
- Customers leave quietly — not one complaint message
- Card expires, VA payment fails — nobody checks
- MRR leak only shows when the monthly report closes
- Usage drops for weeks; nobody reaches out
✕ Now
- Churn noticed after the customer is already gone
- Payment failures met with silence
- Retention judged from a list of lost customers
✦ With Rigel
- Daily 0–100 risk score from real usage patterns, with reasons
- Win-back sequence H+1 / H+3 / H+7 via WA; stops when they return
- Monthly saved-MRR report vs an untouched cohort
What the agent does
The agent gives every customer a daily churn-risk score based on real behavior patterns, runs automated win-back sequences via WhatsApp/email before and after payment failure, and reports the revenue it saved — not just a list of customers who already left.
✦ Handled by the agent
- scores daily churn risk, runs the win-back sequence in the brand's tone, and compiles the saved-revenue report every month.
◆ Decided by you
- decide the bespoke retention offers (discount, downgrade, extension) and handle escalations from high-value customers.
Pipeline
webhooks from the billing aggregator (Midtrans/Xendit/iPayMu), the subscription database, product usage events (logins, API calls, transactions), and CS tickets — unified per customer.
features computed deterministically in SQL (usage-frequency decay, transaction-value decline, ≥2 charge failures, no response to promos), then the LLM writes readable risk reasons ("usage down 60% in 14 days") + a 0–100 score.
red (pre-churn: save offer), orange (payment failed: retry + help with payment method), yellow (dormant: reactivation). High-value accounts go to a human AM, not the bot.
WA-first, email fallback: H+1 usage insight → H+3 personalized offer (package downgrade / retention discount per the client's pricing policy) → H+7 polite last call. Stops automatically when the customer returns or clearly declines; a reply → human handoff.
monthly: at-risk MRR, saved MRR (compared against the untouched cohort), score accuracy (actual churn vs prediction).
Integrations
- Billing/subscription: Midtrans, Xendit, iPayMu, or the client's subscription DB directly
- WA Business API (Fonnte / WATI / Meta Cloud API) + email (Resend / SES)
- Usage events from PostHog/GA4/internal API; output wired into Mission Control
Model stack
- GLM 5.3 Flash for score narratives, win-back message drafting, and reply triage (prompt caching = package/offer table + brand tone)
- Score computation in SQL — deterministic, cheap, auditable; the LLM is only the explanation and conversation layer
- Cost estimate: < US$20/month for 5,000 active customers with daily scoring
Success metrics (tracked weekly)
Sprint deliverables
Billing + usage data ingest with a 30-day historical backtest (score validation)
✓3 risk segments + offer playbooks approved by the client
✓WA/email win-back sequence live with human handoff
✓Monthly "saved revenue" reporting dashboard
✓Flow documentation + code handover (perpetual)
✓Ready to light up this star?
Free 30-minute discovery call: we'll show you exactly how many credits this takes for your business.
Book a Discovery Call →