AI Receptionist Voice Chat
Design a voice receptionist that conducts bounded conversations and hands calls to people or workflows.
Problem statement
For this exercise, assume the source-backed system identity and the explicitly bounded practice contract that follow. Design an AI receptionist for live voice calls.
It accepts an inbound call, carries a bounded conversation, uses approved business information, and either completes an allowed receptionist action or hands the call to a person with context.
Exact call volume, latency target, business integrations, and retention are intentionally open interview inputs; state assumptions and define what the automated agent may never do without confirmation.
What to build
- Establish a live audio session and maintain a coherent turn-by-turn conversation.
- Use only approved business information and expose when the answer is unsupported.
- Execute an allowed receptionist action with confirmation and idempotent external side effects.
- Transfer to a person with a concise transcript and the action state preserved.
At scaleExercise boundary
- Concurrent-call volume, end-to-end response target, integration mix, and retention are intentionally unspecified; choose bounded assumptions.
- Audio and model output are streaming, while business actions require durable exactly-once business semantics.
- A call may outlive or reconnect across failures in speech, model, telephony, or an external action dependency.