◆ Supervised Specialist
Detects the churn signal … closing balances, a 'how do I cancel' search, a competitor's rate quote in chat … and derives a suitable retention offer within policy and value caps. Proposes the offer; for high-value or sensitive saves, an oversight agent re-derives it before it commits, with fair-treatment checks so it never tips into a dark pattern.
Memory
Working The churn signal + the customer's value and current products.
Episodic Prior retention attempts and what worked or backfired.
Semantic Retention-offer policy, suitability rules, value-of-customer models.
Procedural Save playbooks per segment and churn driver.
Store Feature store + offer ledger
Orchestration
MCPA2A
Harness · Managed Agents … event-driven + live-session; offer rationale logged.
Tools
{ } Churn-signal analytics API ⌕ Offer + pricing policy Retrieval ›_ Customer value model Code exec ⇄ Oversight-agent save gate A2A
Evals & guardrails
- UDAAP / fair-treatment guardrail: no manipulative friction or misleading offers.
- Offer-suitability + value caps; high-value saves require an oversight agent's re-derivation.
- Champion/challenger on save strategies; long-run value vs. complaint rate tracked.
- Fairness checks so offers aren't skewed across protected segments.
Frontier edge
- ▲World-model simulation: simulates the customer's likely response to each save option, and the long-run value of keeping them, before proposing an offer.
- ▲Causal reasoning: counterfactual churn analysis ('what's driving the exit, the rate or the bad fee experience?'), so the save addresses the real cause.
- ▲Proactive / anticipatory: catches the drift early (a 'how do I cancel' search, a balance sweep) and stages a fair offer before the customer commits to leaving.
In numbers
26,000
At-risk customers flagged / day
+19%
Save rate vs. baseline