The Problem
A customer emails to cancel a subscription. Your support agent has about ninety seconds to decide: full refund, partial refund, or a retention offer — and they make that call differently every time. A high-lifetime-value customer who just needs a nudge gets treated the same as a churned account, or worse. The decisions are inconsistent, undocumented, and impossible to optimise because nobody can see what was offered to whom.
Why a Black-Box LLM Fails Here
An LLM that just emits "offer 30% off" gives you no policy, no consistency, and no record of why. You cannot audit whether high-value customers are being retained, you cannot prove fair and consistent treatment, and you cannot improve a policy that does not exist in inspectable form.
The Governed Architecture
- The LLM reads. It extracts structured fields from the cancellation message and the account context — tenure, plan, lifetime value, stated reason, recent engagement. It never picks the offer.
- A DMN policy table decides. Your retention policy as a deterministic table: LTV bands, engagement signals, and cancellation reasons map to a refund decision and a retention offer tier.
- BPMN routes. The decided offer is presented or actioned, and the outcome (accepted or declined) is logged.
Try It Live
Go to archiet.com/agents and open the Refund & Retention agent. Type:
Customer Emma Torres, account since March 2023. Monthly plan $299. Lifetime
value to date $8,670. Reason for cancellation: too expensive, not using it
enough. Last login 6 days ago. No support tickets in the last 90 days. Usage
last 30 days: 4 sessions.Click Run governed agent:
1. The LLM extracts (shown read-only):
{
"customer": "Emma Torres",
"tenure_months": 27,
"monthly_value": 299,
"lifetime_value": 8670,
"reason": "price_and_underuse",
"recently_active": true,
"complaints": false
}2. The DMN policy table evaluates and fires: LTV > $5,000 AND recently active AND no complaints → Tier 1 retention: offer 40% discount for 3 months plus a personal onboarding session. Rule RET-HIGH-LTV-002.
3. BPMN routes to present the Tier 1 offer and records whether the customer accepts.
4. The audit trail records the firing rule, the conditions met, the extracted context, and the timestamp.
Now you can see exactly what offer every customer received and why — and you can tune the policy table as you learn what retains.
The Regulatory and Operational Angle
Consistent, documented customer treatment is both a consumer-protection safeguard and an operational asset. When a customer confirms cancellation, GDPR's right to erasure may apply — and the audit trail documents exactly what offer was made and whether it was accepted, which protects you if a treatment-fairness or consumer-protection complaint arises. Beyond compliance, a deterministic policy table is the only version of this decision you can actually measure and improve: you change a rule, you watch retention move.
Generate This for Your Own System
Describe a subscription, billing, or customer-success workflow in Archiet's Blueprint Wizard and the governed refund/retention agent is generated with your LTV bands, offer tiers, and eligibility rules extracted from your PRD. Engine, policy model, endpoints, and UI come wired and auditable.
The same governed pattern powers invoice approval, loan adjudication, FNOL triage, prior authorisation, and KYC/AML in the gallery.