AI voice agents for restaurants that handle every reservation, every night
Your hostess is with a guest. The phone keeps ringing. An AI voice agent answers every call — books reservations, quotes waits, takes takeout orders, and handles questions about the menu in whatever language the caller speaks.
One hostess, three phones, six people at the door.
Restaurants lose bookings to a simple operational problem: the hostess is physically busy seating a party, the phone is ringing, and the caller on the other end is deciding whether to call the next place on their list. Peak call times collide with peak seating times. After-hours calls don't get returned. Voicemail doesn't work for reservations — most callers don't leave one, and the ones who do get called back in the morning after the slot is gone.
The workarounds are worse. A traditional IVR tree that asks the caller to "press 1 for reservations" is a downgrade from a human. Staffing multi-language service is a staffing problem most single-location restaurants can't solve. Answering services miss the operational context — the chef's tasting on Tuesday, the private-event policy, the allergen matrix — and get every other detail wrong.
An AI phone agent answers every call, in every language, at $0.09 per handled call. It books directly into OpenTable, Resy, SevenRooms, or Toast Tables. It takes takeout orders and writes them into your POS where your POS exposes an order API — the matrix below says which do. It answers the hours, parking, and dietary questions that would otherwise tie up the floor. And when it can't, it hands the call to a named human with full context.
Three AI deployments that fit how a restaurant actually runs.
Metrics below are from live restaurant deployments, not pilots.
Reservation & booking agent
Inbound AI on Vapi/LiveKit that books reservations in OpenTable, Resy, SevenRooms, or Toast in under 90 seconds. Quotes real-time wait, offers alternatives when booked out, and holds the slot while checking availability.
Takeout & order-taking agent
Voice agent that takes takeout orders across phone and drive-thru, writes into your POS where an order API is available (Toast, Square, Clover, Lightspeed), and confirms the order with the customer before submit. Handles modifiers, allergen flags, and upsells without being pushy.
FAQ, hours, and location agent
Answers the calls that don't need to tie up the hostess — hours, location, parking, dietary menu questions, gift card purchases, private-event inquiries routed to the right manager. Frees the floor to actually host.
Milina — AI voice reservation agent running on LiveKit.
Milina — AI voice reservation agent running on LiveKit.
Milina, a reservation-driven NYC restaurant, needed a voice agent that could handle peak-hour reservation traffic without sounding robotic. We deployed LiveKit voice infra with Deepgram STT, GPT-4o-mini for dialog, and Cartesia for low-latency TTS. It handles 50+ calls a weekend night, books directly into the reservation system, and the unit economics land at $0.09 per handled call. Customers can't tell it's AI.
The stack we deploy when the phone has to be answered.
Low per-call cost, low latency, and direct writeback into the reservation platform and POS the restaurant already runs.
LiveKit · Vapi · Retell
For a single location with heavy reservation volume we usually deploy LiveKit for the lowest per-call cost; Vapi for fastest time-to-MVP; Retell for hosted convenience.
Deepgram · Cartesia · ElevenLabs
Deepgram for STT tuned to noisy restaurant backgrounds; Cartesia for the fastest TTS when latency shows; ElevenLabs when voice quality matters most.
GPT-4o-mini · Claude Haiku
Small models are the right answer for restaurant reservations. Faster, cheaper, and accurate enough for a bounded conversation.
OpenTable · Resy · SevenRooms · Tock
Direct API integration where available; browser automation fallback where it isn't (yes, we'll build both).
Toast · Square · Revel · Clover
For takeout and drive-thru, order write-back over the platform's official API so the ticket prints already confirmed. Scoped per restaurant — see the POS matrix.
LangSmith · Helicone · CloudWatch
Every call fully traced. Call recordings retained per your policy; cost and latency tracked per-call.
Restaurant calls are not spread evenly, and that is what breaks the pricing.
Why per-minute and per-call billing both fail a restaurant, and what to look for instead.
Restaurant call volume clusters. The phone is close to silent for long stretches and then produces most of the day's calls inside a 90-minute window that lands exactly when the floor is fullest. That is the same window in which a hostess cannot pick up, so the unanswered share climbs at the worst possible moment — 30–40% of peak-window calls going unanswered is the figure restaurant operators quote, and it is worth measuring against your own phone logs before you believe anyone's numbers, including ours.
Per-minute billing charges you most for the peak. A per-minute plan prices the exact minutes you were least able to answer, and it prices the queue too. Human answering services quote $0.75–$1.50 per minute; the reservation call centre the Milina owner used before us charged $1.80 per call, and the scripts were wooden.
Per-call billing punishes takeout. A reservation is a 60–90 second call. A takeout order with four items, three modifiers and an allergen question is not, and per-call pricing that averages the two either overcharges the reservation or loses money on the order — which the vendor fixes later, on you.
The labour saving evaporates if the order does not land in the POS. This is the part that decides whether any of the arithmetic matters. If the agent takes a perfect order and then emails it, or drops it into a dashboard, someone at the pass re-keys it into the till. That re-key is most of the labour you were buying back, plus a new source of errors at the busiest moment of service. An order that is not written into the POS is a message, not an order.
What that adds up to. A flat monthly cost with native POS write-back, priced on a volume band rather than on minutes. At $0.08–$0.18 per handled call, a single location on 1,500–3,000 calls a month runs $120–$540 of actual runtime; the number that varies between vendors is not the runtime, it is what they add on top of it and whether the order reaches the kitchen without a human in the middle.
Which POS take a real order write-back.
Same rule as everywhere else on this site: what we have shipped is marked as shipped, and what depends on your configuration is marked as that.
| Platform | Interface | What the agent can write | Status with us |
|---|---|---|---|
| Toast | Official Orders API, plus partner integration review | Order creation and updates; structured call notes against a booking | Note pipeline in production (Milina). Order write-back scoped per restaurant. |
| Square | Official Orders API | Order creation with line items and modifiers, sent to the ticket printer | Supported via official API; scoped in discovery |
| Clover | Official REST API and developer app | Order creation; menu and modifier sync | Supported via official API; requires a Clover app approval |
| Lightspeed Restaurant | Official API with event-driven webhooks | Order creation; inventory and order event subscriptions | Supported via official API; scoped in discovery |
| Revel | API access is tied to your Revel agreement | Depends on the endpoints your contract exposes | Assessed first — we check your access before quoting |
| Reservations: Resy, SevenRooms, OpenTable, Tock | Direct API where available | Availability, booking, modification, cancellation | Resy in production with SevenRooms fallback (Milina) |
| No usable interface | Structured handoff to your existing workflow | Order capture and confirmation, without write-back | Fallback — and we will tell you it is one |
The distinction that matters when you compare vendors: writing a note against a booking and creating an order that prints in the kitchen are different integrations with different approval paths. On Milina we shipped the first — Resy handles the booking, and a transcript-tagged note lands in Toast so the host has context when the guest walks in. Order write-back is a per-restaurant scope because it depends on your menu structure, your modifier tree, and in Clover's case on an app review. Any vendor who tells you order write-back is a switch they flip has not done it.
The caller shouldn't ask if they're talking to AI.
Four questions every restaurateur asks before signing. Here are ours, answered honestly.
What languages?
English, Spanish, French, Italian, Mandarin, Cantonese, Japanese, Portuguese are standard. Mixed-language calls (customer starts Spanish, switches to English) are handled natively with no reset. We use language-specific voices — not one voice with an accent layer.
How robotic does it sound?
The bar is "the caller doesn't ask if they're talking to AI." We get there with Cartesia or ElevenLabs voices tuned to the restaurant's tone, sub-600ms response latency, and interruption handling. We play a sample before you sign.
What about dietary & allergen questions?
The agent knows your menu, your allergen matrix, and your "ask the chef" flags. When it doesn't know, it says so and offers a callback from a human — never guesses on a peanut allergy.
Can it upsell without being annoying?
Yes, within rules you set — the Cabernet on Tuesday, the new dessert for first-time guests, the chef's tasting on slow nights. Off-theme upsells are blocked at the prompt layer. You approve the playbook before it goes live.
Single location to restaurant group.
Fixed project pricing. Per-call unit economics modeled in the discovery workshop so you know the monthly run-rate before you sign.
| Deployment | Scope | Price |
|---|---|---|
| Single-location voice agent | Inbound reservations + takeout, 1 POS, 1 reservation platform, shadow mode | $6,000–$12,000 |
| Multi-location rollout | Shared templates across locations, per-location config, consolidated analytics | $15,000–$28,000 |
| Enterprise restaurant group | Group HQ + franchisee config, approval flows, multi-language, custom voice | $25,000–$45,000 |
| Outbound reservation recovery agent | Callback agent that converts no-shows and waitlist, review-request flow | $7,000–$14,000 |
| Monthly retainer | Ops, menu updates, seasonal changes, new language, analytics | $1,500–$6,000/mo |
Per-call cost lands between $0.08 and $0.18 depending on language and call length. A single-location restaurant on 1,500–3,000 handled calls/month spends $120–$540 on AI voice runtime.
Restaurant questions we answer on every intro call.
How much does an AI voice agent for a restaurant cost?
Which reservation platforms do you integrate with?
Can the AI handle multiple languages?
Will customers know they're talking to AI?
What happens if the AI can't answer a question?
How fast can we go live?
Ready to answer every call, every night?
One 20-minute call for restaurateurs. We'll listen to a sample call, look at your reservation platform and POS, and tell you what's realistic — and what the unit economics will be. If it doesn't make sense for your volume, we'll say so.