On this page
The restaurant phone problem has a specific shape: demand arrives in a two-hour spike, exactly when every member of staff is least able to answer it. Between 6pm and 8pm the phone rings, and the person nearest it is carrying three plates.
This is close to an ideal case for an AI voice agent, because concurrency is the whole issue. Ten calls at 7:15pm are ten simultaneous conversations for an AI and a queue for a human.
The short version
Reservations are the easy, high-value starting point. Full menu ordering is genuinely harder than vendors imply — modifiers, allergens and background noise are where it gets difficult. Start with reservations, add ordering once the basics are boring.
Reservations: start here
Taking a table booking is a small, well-defined task: date, time, party size, name, phone number, and any note. The agent checks the table plan, offers real times, writes the booking, and confirms by text.
The rules a restaurant has to encode are fewer than a clinic's but not trivial:
- Table capacities and which combinations can be joined.
- Turn times by party size — a two-top and a party of eight are not the same slot.
- Service periods, last seating and kitchen close.
- Large-party rules: deposits, set menus, a manager callback above a threshold.
- Accessibility requirements, high chairs, and where those can be seated.
- What is done with a no-show, and whether a card is held.
The mechanics are the same as any other booking system — see AI appointment booking — with a table plan in place of a diary.
Phone orders: harder than it looks
Taking a takeaway order by voice is a genuinely difficult problem, and it is worth being clear about why before committing to it.
The modifier explosion
"A large pepperoni, but half with extra cheese, no onions on the other half, and can you make the base thin" is one sentence and four modifications against one item. Menus have hundreds of valid combinations and a smaller number of invalid ones the system also has to know about.
Background noise
Callers ring restaurants from cars, streets and parties. Transcription quality degrades exactly when it matters most, which is why order confirmation — reading the whole order back before taking payment — is not optional.
Allergens
This is the one to take seriously. An AI system must never guess about an allergen. The correct behaviour is to record the stated requirement verbatim, flag the order prominently for the kitchen, and — where the risk is significant — hand the call to a person. Treat this as a hard boundary in the same category as emergency triage in a clinic.
Allergen handling is a safety boundary, not a feature
Decide in advance exactly what the agent does when a caller mentions an allergy. "Repeat it back, flag the ticket, and transfer to a member of staff" is a defensible policy. "Answer from the menu data" is not.
What it fixes on a Friday night
The questions that are pure win
Before ordering or reservations, a substantial share of restaurant calls are simply questions. These are free to hand over and cost nothing to get wrong:
- Are you open now, and until when?
- Do you have availability tonight for four?
- Do you do takeaway or delivery, and what is the delivery area?
- Is there parking? Are you accessible? Do you allow dogs?
- Do you have vegan, vegetarian or gluten-free options?
- Do you take walk-ins?
- How long is the wait right now?
A sensible rollout for a restaurant
Week 1 — Information only
Hours, location, parking, dietary options, delivery area. The agent answers and, for anything else, transfers or takes a message. Zero risk.
Week 2–3 — Reservations
Connect the table plan. Start with standard party sizes and escalate large groups to a person.
Week 4 — Overflow and after hours
Route calls that ring out during service, and everything outside opening hours, to the agent.
Later — Ordering, if the menu allows
Start with the ten most-ordered items and no complex modifiers. Read every order back. Expand slowly, and keep allergen calls with humans.
Frequently asked questions
Can an AI take a full takeaway order by phone?
It can, but menu complexity determines whether it should. A short menu with few modifiers works well. A large menu with extensive customisation is a much harder problem, and the failure cases — wrong orders during peak service — are expensive. Starting with reservations and information calls is the lower-risk path.
What about allergies?
Treat this as a hard safety boundary. The agent should record what the caller says verbatim, flag it prominently on the ticket, and hand the call to a member of staff where the risk is significant. It should never reason about ingredients on its own.
Will it work with my reservation platform?
It depends on whether the platform offers an API. Most modern reservation systems do. Where one does not, the agent can still capture structured booking requests for staff to enter, which is weaker but still better than a missed call.
Can it tell callers the current wait time?
Only if something in your operation actually knows the wait time and can be queried. If wait times live in a host's head, the agent cannot report them — and it should say it does not know rather than estimate.
Does it handle group bookings?
Set a threshold — typically six or eight — above which the agent takes the details and hands off to a person. Large parties usually involve deposits, set menus and negotiation, which are human conversations.
Cafés, bars, salons and any other walk-in business with a phone face a version of the same concurrency problem. The complete guide to AI voice agents covers the general case, and the AI voice agents page shows how TensoraAI builds them.
