AI Front-Desk / Booking Agent (demo project)
Demo on sample data. Not a client project. The run below used a local test stand-in for the AI, which only exercises the tool-calling path. It does not show real conversational quality.
The problem
Small service businesses answer the same questions and booking requests all day.
What it does
A Claude-style tool-using agent that can check availability, create a pending booking request, and escalate to a human. It can only use those three actions.
Design choices
- Bookings are created as pending; staff confirm them.
- Refunds, complaints and out-of-scope questions escalate with a reason.
- Unknown tool requests are rejected, and customer text is treated as data.
What we tested
Three automated tests cover the tool layer: a valid booking becomes pending, an unavailable slot is refused, and an unknown tool is rejected. The full tool-use loop ran through the real Anthropic Python library against a local stand-in, which called the availability tool and returned a reply.
What we did not test
Behavior with a live AI model, real calendars or CRMs, or long conversations. A client job starts with a behavior spec the client approves, then scripted test conversations against the real setup.
Files: `demos/3-support-agent/`.