What are retention guardrails?

Retention guardrails are the limits a merchant sets on what a save offer may contain — which kinds of offer are allowed at all, and how far each may go — enforced by the system rather than trusted to whoever is making the offer. In an AI-assisted flow they are what separates a concierge from an open-ended discount machine.

#Why the enforcer cannot be the model

A language model can be asked to stay inside a limit; it cannot be relied on to. The only durable enforcement is a server that independently checks every proposed action against the merchant's settings before the shopper ever sees it, and drops anything out of policy. That is a security property, not a prompt.

#The guardrails WallaB.AI enforces

• A master switch. With the concierge off, it makes no offers at all and simply confirms cancellations kindly.
• Allowed actions — which of discount, pause, swap, and skip may be proposed.
• A hard discount cap — a percentage the merchant sets, which no offer can exceed; with no cap set, no discount is offered at all.
• A longest pause, from one to twelve whole months.
• A swap allow-list, or the whole active catalog when left empty.
• Tone guidance in the merchant's own words, up to 200 characters.
• One-off add-on limits — a price cap, a product allow-list, and a built-in per-day velocity limit.
• Skip limits on the everyday self-serve skip — see Skip limits.

#What happens to an out-of-policy proposal

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 that names only the saves that can genuinely be made on that turn, never names the guardrail that fired, and always restates the plain cancel path. Every decision — replies, offers proposed, offers dropped, moderation events — is logged and reviewable.

Configure them on WallaB concierge & guardrails.