Skip to main content

AI Agent Structured Guardrails & Decline Quality

Typed Guardrails (Forbidden vs Not-our-market, with enforced per-type behavior and one canonical refusal template) plus dead-end-free capability declines and a quick answer-correction layer — so the AI Agent refuses only what is truly forbidden and steers every winnable customer toward sales.

Born from the Jun–Jul 2026 telesales quality audit: ~50 genuine businesses refused like abusers, 525 declines with no alternative offered, 3 inconsistent refusal wordings.

Hard blocking dependency: DSAI (AI Service) must consume a structured guardrail schema and render the canonical refusal template verbatim — contract agreed and dated before the PRD moves to READY. DSAI committed 2026-07-21 to add a structured guardrails array in AI service (target date TBD); see thread note.

Scope Changes​

Backend (typed guardrail persistence + sync, answer-correction store/injection, events) · Frontend (Guardrail type selector + refusal template field, Answer Corrections page) · Design (both surfaces — no existing pattern for the corrections page).

QA Lane​

Lane B — human QA gate. Basis: customer-facing refusal behavior with a cross-team (DSAI) contract; automation entry bar unmet (no E2E specs yet). Owner TBD; assessed 2026-07-23 at PRD authoring time.

PRDs​