Agentic Multi-Tool Voice Concierge

A customer-facing assistant that takes both text and voice messages and calls dedicated tools for live data — checking today's menu or booking a table — instead of guessing.

How it works

The problem

Text-only FAQ bots can't handle voice messages or real actions like checking live availability or making a booking — so businesses still need a human for anything beyond the simplest question.

The approach

1. 1

   Every incoming message is normalized first; voice notes are transcribed through Whisper before they ever reach the conversational agent, so the same flow handles text and audio identically.

2. 2

   The core agent is wired with dedicated tools, each an isolated, independently testable sub-workflow: one reads live availability from a source of truth, one records a booking, one alerts a human for anything the agent shouldn't decide alone.

3. 3

   Session memory keyed per contact keeps multi-turn conversations coherent without mixing up different customers.

4. 4

   The agent is instructed to answer only from what its tools return — never inventing availability, prices, or details it wasn't given.

Highlights

• One entry point handles both voice and text messages.
• Tool-calling architecture keeps the agent grounded — no invented answers.
• Each tool is an isolated sub-workflow, easy to test and extend independently.

Built with