Design a Voice and Chat AI Receptionist Platform

Design a customer-facing AI receptionist platform for businesses. Customers can call a business phone number or contact it through chat on a website, mobile application, or supported messaging channel. The receptionist answers using business-approved information, performs permitted actions, and transfers the customer to a human when needed. ### Constraints & Assumptions - Different businesses and industries require configurable behavior and approved knowledge. - Preserve tenant isolation and distinguish answering a question from executing an action. - A human handoff must transfer control of the live conversation, not merely suggest that the customer contact someone else. - Define an initial scope before choosing traffic or latency targets; none are supplied. ### Clarifying Questions to Ask - Which actions are allowed, and which require customer confirmation or human approval? - How are business knowledge, policies, hours, and escalation contacts approved and updated? - Does each business have a staffed agent queue, and what happens outside its available hours? - Must voice and chat share conversation history across channels? ### Part 1 — Channels and conversation state Describe voice and chat ingress, identity, session state, and response delivery. #### What This Part Should Cover - Channel adapters and a shared conversation model without losing channel-specific timing behavior. - Tenant routing and authorized customer context. - Voice transcription/synthesis, interruption, and partial-message handling. ### Part 2 — Knowledge, configuration, and actions Explain how the platform uses approved information and executes only permitted operations. #### What This Part Should Cover - Versioned business configuration and permission-aware retrieval. - Grounding, uncertainty, content-quality evaluation, and unsupported requests. - Typed action contracts, authorization, confirmation, idempotency, and audit. ### Part 3 — Human takeover Walk through escalation, agent assignment, transcript/context delivery, and transfer of response authority. #### What This Part Should Cover - A real agent queue and an atomic ownership transition. - Voice bridging or chat routing after acceptance. - Prevention of simultaneous bot and human replies and behavior when no human is available. ```hint Treat takeover as a state transition The model deciding to escalate is only the start. Specify who owns the conversation before assignment, during transfer, and after the human accepts. ``` ### What a Strong Answer Covers - A coherent multi-tenant voice/chat platform with explicit control boundaries. - Business-specific customization that cannot bypass shared security and action policies. - Measurable answer/action quality and a functioning human handoff path. ### Follow-up Questions - How would you handle a customer interrupting an AI voice response? - What happens if an action succeeds but the model or channel times out before acknowledging it?