AI reservation agent for restaurants: party size, dates and the waitlist

What it is

A reservation agent answers the phone, works out whether the caller wants to book, change or cancel, resolves the date and time aloud, checks real availability for the party size in the reservation system, applies your policies for large parties, deposits and blocked times, writes the booking, and sends a confirmation by text. When you are full it offers the nearest real slots or the waitlist. When a caller asks about an existing booking, it identifies them before it says anything about it.

The call is short, two minutes or less when it goes well. The risk sits in three places. Relative dates: "next Saturday" means different things to different people and the agent must say the calendar date before it checks. Party size: "six, maybe seven" needs a decision, and a party above your threshold needs your policy, not a table. And changes: a caller moving a booking must end up with one booking, not two.

Independents call this an AI host or AI phone answering for bookings. Groups call it reservation automation and care about policy consistency across venues.

Who buys it

• Full-service independents where the host stand is the phone, and the phone loses to the guest standing in front of it every time.
• Restaurant groups with several venues on one reservation platform, who want the same booking rules and the same voice on every line.
• Reservation-platform customers choosing between the platform's bundled phone agent and a standalone product, who should run both on the same script.

Budget owner: the owner or general manager; in groups, the operations lead, with the reservation-platform administrator signing off on the integration and policy configuration.

KPIs

Sample fifty phone bookings against the book before you deploy, and log missed calls by half-hour for a week. Then track booking containment, date and time accuracy, answer rate at peak, large-party policy compliance and no-show rate on agent bookings against the staff cohort.

The trap is containment that rises because the agent books whatever it hears. Read accuracy and policy compliance alongside it. A fourteen booked as a standard table on a Saturday is a containment success and an operational failure.

Demo script

Bring a sandbox of your reservation system with real-looking availability, one blocked time, your large-party threshold and one existing booking under a test name.

1. Greeting and intent. Pass: restaurant name, AI disclosure, recording notice, and "new booking or changing one?"
2. Ambiguous date. Ask for "next Saturday at half seven for six". Pass: it names the calendar date and 7.30 p.m. and waits for a yes before checking. Fail: it checks availability without saying the date.
3. Interruption. While it reads out available times, cut in with "actually make it eight o'clock". Pass: it stops mid-sentence and checks eight.
4. Party-size change. After it offers a table, say "sorry, there are eight of us now". Pass: it rechecks availability for eight once and, if eight crosses your threshold, explains the policy. Fail: it keeps the six-person slot.
5. Large party. Ask for fourteen on the same night. Pass: policy explained, deposit link or hand-off to the events contact, no standard booking. Fail: fourteen booked as a table.
6. Eight seconds of silence after it asks for a phone number. Pass: one short prompt, then a graceful offer to text a link or call back.
7. Full at the requested time. Ask for the blocked slot. Pass: nearest real alternatives, then the waitlist with party size and phone written, and an explanation of the callback.
8. Modify an existing booking. Say you need to move "my booking on Friday". Pass: it asks for the name and phone or reference, matches, reads the booking back, then changes the same record. Fail: a second booking appears.
9. Out-of-scope question. Mid-booking, ask whether the duck is gluten free and whether you can bring a cake. Pass: menu answer from your data with no allergen guarantee, cake answer from your policy, then back to the booking.
10. Confirmation and write. Accept a slot, confirm name and phone with digit-by-digit read-back, and watch the booking appear in the sandbox with the SMS confirmation in the log.

Score each trap pass or fail. A large party booked as a standard table, or an existing booking disclosed before identity matched, is a hard stop.

Compliance notes

The booking call itself is inbound and outside the TCPA consent rule in the United States, but the confirmation text, the reminder and any waitlist callback are outbound contacts. The rule as published needs prior express consent for artificial-voice calls and texts, and the 8 a.m. to 9 p.m. window in the guest's local time applies to anything that counts as a solicitation; capture consent on the call and store it per number. Recording consent varies by state, so announce it. In the United Kingdom, recording is personal-data processing under UK GDPR and the agent should say so; a reminder is a service message, but a "come back and see us" call is marketing and needs specific prior consent under PECR Regulation 19. In the European Union, the AI Act's Article 50 duty to tell callers they are talking to an AI applies from 2 August 2026, and recording needs a GDPR lawful basis and notice. In Australia, state surveillance-devices laws differ, so announce recording; any promotional callback sits inside the telemarketing hours of Monday to Friday 9 a.m. to 8 p.m. and Saturday 9 a.m. to 5 p.m. with the Do Not Call Register applied. Informational, not legal advice; the compliance rows on this page carry the sources.

Build or buy

Buy if you are on a mainstream reservation platform; the bundled agents and the integrated standalone products have done the availability and policy plumbing, and your job is to test them on the same script. Consider a platform or a build only if you run a bespoke booking system or a policy model (ticketed dinners, prepaid menus, multi-room venues) that packaged products cannot express. In both cases the acceptance test is the same: "next Saturday" spoken back as a date before availability is checked, a party above your threshold routed to policy, and a change that changes one record.