WallaB concierge & guardrails

WallaB is the AI cancel-save concierge: when a shopper heads for the cancel button, WallaB starts a real conversation, diagnoses the reason, and proposes a save offer — strictly inside limits you set.

#Guardrails you configure

Everything lives in Settings → WallaB concierge:

• Enabled — the master switch (off by default; with it off, WallaB makes no offers and simply confirms cancellations kindly).
• Allowed actions — which offer types WallaB may propose: a discount, a pause, a product swap, or skipping the next delivery.
• Maximum discount — a hard percentage cap; WallaB can never offer more. A new install has no cap saved, so WallaB offers no discount at all until you set one: the settings form suggests 15% the first time you open it, and onboarding recommends a cap based on your margins. Nothing is stored until you save.
• Maximum pause — 1–12 whole months (default 3). A dated pause resumes by itself: when its end date arrives, billing picks the subscription back up on the promised schedule — neither you nor the shopper has to do anything. (An indefinite, manually-created pause stays paused until someone resumes it.)
• Swap products — optionally restrict swaps to a specific product list; left empty, any active product in your catalog is eligible.
• Tone — free-text guidance (up to 200 characters), e.g. "Warm and casual — we're a small family roastery."
• One-off add-on limits — an optional price cap and product allow-list for "add to your next delivery" add-ons. A one-off (or a reordered past item) is added to the subscriber's next order and billed with it — never charged on its own — and only on pay-as-you-go subscriptions; a built-in per-day velocity limit stops add-ons from stacking up.

#The server is the enforcer, not the model

The AI only ever proposes an offer. Before anything reaches a shopper, the server independently validates it against your guardrails — and when a proposal is out of policy, dropping the offer card is only half the job, because the model's own wording would still be promising it. So the whole reply is replaced with vetted copy: it names only the saves WallaB can genuinely make on that turn, never names the guardrail that fired, and always restates the plain cancel path. A withheld offer is never promised in the text.

If the offer type was fine and only the terms were out of bounds — a discount above your cap, a pause longer than you allow — the replacement says WallaB offered more than it can deliver and corrects itself, without naming your limits, so it can come back with the same kind of offer at terms you permit.

The model is never trusted to enforce its own limits.

Discounts follow a replace, never stack rule: a new concierge discount replaces the active one, so WallaB will only propose a discount that is strictly better than what the subscription already has — and once a subscription is at your cap, no further discount offers are valid. An accepted concierge discount does not expire: it applies to every order from then on, until you change it or a later save offer replaces it.

#Moderation, both directions

Every message is screened by a moderation filter in both directions — what the shopper sends before the AI ever sees it, and what WallaB writes before the shopper ever sees it. Screening covers all five portal languages (English, Spanish, French, German, Brazilian Portuguese) at once. Flagged shopper messages get a calm de-escalation reply instead of an AI turn, and after repeated flags the plain cancel confirmation is surfaced directly so nobody is ever trapped in a conversation.

#Reliability without an AI key

WallaB runs on a three-tier brain: the live AI model when configured, a library of your own proven past saves as the second tier, and a deterministic fallback that always answers. The cancel-save flow — and every guardrail — works even with no AI provider connected.

#Everything is auditable

Every conversation is stored with its full transcript, and every AI decision (replies, proposed offers, dropped offers, moderation events) is logged. You can review exactly what was said and offered in your dashboard's Conversations and Activity views.

#No dark patterns

A plain "just cancel" path stays visible throughout the entire conversation — no countdown timers, no guilt copy, no pre-checked add-ons. That's a product rule, not a setting.