The win-back engine that cancels its own bonuses.

Short answer: the churn module builds a complete return plan for every quiet VIP worth pursuing — diagnosis of why they left, an offer priced by expected value, ready-to-send outreach in the player's own language, and the back-office bonus parameters behind a checklist gate. Its defining behaviour is restraint: when expected return doesn't clear the offer's cost, the bonus is cancelled by calculation and the plan says "service touch" instead. On a recent production run, that was two out of every three plans.

Part four of the Inside the Brief Center series (the model · the KPI brief · the VIP cockpit). The cockpit's job is to keep VIPs from going quiet; this module's job is the harder one — winning back the ones who did. Examples below are anonymised first-party observations from the scaling brand in our case study, not industry benchmarks.

Evidence note: the 34-plan run and its 21 no-bonus decisions describe one operational sample. A publication-grade comparison still requires the run date, unique-player count, eligibility and suppression rules, model version, forecast horizon, offer-cost definition and realised control-group outcomes.

Not a blast — a stack of individual plans

The nightly run scans every VIP silent between one week and four months, ranks them by potential loss × probability of return, and produces a full plan for each player at the top of the ranking — on a sample night, dozens of complete plans generated in minutes. Each plan opens with a diagnosis, because players leave for different reasons and the wrong response makes it worse: a heavy-loss burnout gets careful, no-pressure care; a KYC-friction case gets the verification unblocked before anyone mentions playing; natural cooling gets a light content touch. The plan states which case it believes this is, and shows the evidence.

The EV gate: every offer pays for itself, or it doesn't go out

The economics are computed per player, not per campaign. Expected return is the player's historical run-rate projected over a recovery window, weighted by the modelled probability of reactivation; the offer's cost has a hard ceiling relative to the player's lifetime value. If the expected return doesn't clear cost with margin to spare, the bonus is removed from the plan — the recommendation downgrades to a personal message or a content touch. This is the inversion that matters: most win-back programmes measure themselves by offers sent; this one treats a cancelled bonus as a correct outcome.

Outreach that reads like a person wrote it

Each plan ships two alternative outreach texts — different angles on the same offer — plus a chat script for the support line, written in the player's actual locale and register, referencing their favourite game by name. The channel is picked from evidence, not preference: a player who never opens email gets a push plan; an unsubscribed player is never emailed. Timing comes from the player's own best playing hour. A VIP host reviews, optionally edits, and sends — the drafting is automated; the relationship isn't.

Execution discipline: the checklist gate

The plan includes the exact bonus parameters for the back office — but the copy button unlocks only after the manager completes a checklist: protection flags re-checked, pending withdrawals confirmed clear, contact history reviewed. Outcomes feed a kanban and a suppress list; a player marked "do not contact" stays suppressed across future runs, and a player who reappears in the queue repeatedly without action is escalated instead of re-planned. Every report carries a freshness guard — if the data is stale, the plan says so before anyone acts on it.

What restraint buys

The relevant question is not only whether an offer is accepted, but whether it creates incremental value after cost. On the same anonymised brand, the cockpit reported around 60% return in its at-risk band and under 10% in its churned band. That rounded observational first-party curve is not proof that earlier contact caused the difference: cohort composition, eligibility and selection may differ. The safe operating principle is to measure incrementality with an eligible holdout and never promote to a player whom protection rules suppress. Next in the series: the support brief — LLM triage of every conversation, and how support tickets become a product backlog. Platform modules: /platform/.