Create an AI support agent

You are helping me design an AI support agent for my B2B SaaS (CommentSell.com). The agent lives in our app and on our website. It must: 1) Answer support questions conversationally using our documentation + product knowledge. 2) Resolve issues end-to-end by calling backend actions, not just linking articles. 3) Those backend actions are Cloud Functions that write to Firestore (or call other APIs). 4) The AI must NEVER directly write to Firestore. It can only request actions through a strict whitelist of function calls. 5) Every action must have: validation, permissions, audit logs, and clear user confirmations. 6) Risky actions require human approval (like refunds, deletions, plan cancellations). 7) It should gracefully fall back to a human support agent when unsure or missing data. Design the architecture, data model, and prompting strategy. Include: - A Firestore schema for conversations and action requests - A list of recommended Cloud Functions (actions) - How the AI decides when to call an action vs ask a question - Guardrails against hallucinations / unauthorized changes - Example system prompt(s) and JSON tool schemas - A minimal MVP plan and then a scale plan