ANSWER

An AI Guardrails Catalog is a centralized, reusable list of explicit boundaries for an AI-enabled feature—written as testable requirements, not vague policy statements.

In practice, the guardrails catalog is where the team documents:

• What the AI is not allowed to do (forbidden actions, restricted data, unsafe advice, bypass requests)

• What it must do instead (safe alternatives, escalation paths, clarifying questions, redaction, confirmation steps)

• What the user should see (approved refusal/safe-response wording)

• How you verify it (pass/fail validation steps QA can test)

I like to think of it as “acceptance criteria for safety.” Instead of relying on prompts or tribal knowledge, the catalog turns guardrails into observable product behavior that can be traced into user stories, acceptance criteria, Definition of Done, and regression tests. It also improves consistency across channels (UI, chat, exports, APIs) because the same guardrail IDs can be referenced everywhere.

A simple, high-value catalog entry typically includes: ID/title, scope, trigger, not allowed, required safe behavior, refusal copy, safe default, and validation. When it’s done well, the guardrails catalog reduces surprises, speeds up reviews with security/compliance, and gives QA something concrete to test—so “be safe” becomes something you can actually ship.