Calls peak when the floor is busiest
One phone, one host, every caller at once.
AI VOICE RESERVATION AGENT
The restaurant phone that never goes unanswered.
A voice agent that answers the phone, checks live table availability, books the table, and handles cancellations.
01 · How it works
How it works
Voice in, a live table lock out.
Caller
Picks up the phone
ElevenLabs
Hears, speaks, handles turn taking
Anthropic Claude
Decides what to ask and call
Make
Runs the logic and the guards
Airtable
Holds every table and seating
02 · Problem
Six ways the phone loses bookings
Nobody knows what is actually free
Availability lives in one person's head or one paper book.
Two parties, one table
Shared sheets let two bookings land in the same slot.
Large parties refused by default
Single-table thinking misses combinations that would seat them.
Booking calls pull staff off the floor
Every minute on the phone is a minute away from guests.
Bookings lack detail for the floor
Notes stay in someone's memory, not with the reservation.
03 · How I build
How I build voice AI agents
01
Voice experience
Diners judge the call in the first few seconds. STT and TTS settings, latency, and interruption handling are tuned so the agent sounds like a real host, with guest names, dates, times, and party sizes coming through clearly.
02
Prompting and context
The agent knows the restaurant, its three nightly seatings, and its booking rules, and nothing it doesn't need. Persona and tone match a friendly host, and context stays tight so every turn moves toward a confirmed reservation.
03
Conversation design
Happy paths cover new bookings and read-backs before confirming. Error paths cover fully booked slots, mid-call changes, and unclear dates. Guardrails keep other guests' details private and route large parties or complaints to staff.
04
Tools and integrations
Every agent tool is a Make.com webhook scenario that checks availability or writes a booking before the agent answers. Airtable holds the backend, with tables, seatings, and bookings read and written live, so every answer reflects the real floor plan.
05
Evals and testing
Each change runs through scenario tests and tool call tests before going live. Coverage includes date and time accuracy, party size changes mid-call, booking data reaching Airtable correctly, fully booked alternatives, and data fishing attempts. Live calls are tracked for latency and booking accuracy.
06
Monitoring and iteration
Call performance is reviewed after every round of live calls. Failures are root-caused by layer (speech, prompt, tool, or flow), agent settings and prompts are updated to fix them, and the full test suite is rerun before the next version ships.