AI phone ordering for restaurants: menu, modifiers and a POS write

What it is

A phone ordering agent answers the line during service, works out whether the caller wants pickup or delivery, takes the order from the live POS menu with modifiers and quantities, answers what-is-in-it and allergy questions from ingredient data, quotes a pickup or delivery time from kitchen load, reads the itemised order and total back, and writes the order to the POS so the kitchen ticket prints without anyone re-keying it. Payment goes by link or on collection; the agent never hears a card number.

The call shape is short but dense: greeting and disclosure, mode, items with modifiers, corrections, name and number with read-back, time, confirmation, write, payment instruction. Two to four minutes is normal. The hard parts are inside the items turn: a caller says "two green curries, extra spicy, one with no rice" and then changes a quantity three seconds later, often while the agent is reading back. An agent that adds a line instead of changing one has made the order wrong in a way the kitchen cannot see.

Independent owners buy this as AI phone ordering or AI phone answering for restaurants. Multi-unit operators buy it as voice order capture with central menu management and per-location reporting.

Who buys it

• Independent restaurants and small groups doing a meaningful share of revenue through phone pickup and delivery, especially pizza, Chinese, Thai, Indian and Mexican kitchens where the phone has always been the ordering channel.
• Multi-unit operators and franchisors who want one ordering experience across locations, synced to each site's POS menu and availability, with accuracy and containment reported per store.
• Ordering and POS platforms bundling a phone agent, where the buyer is choosing between the bundled agent and a standalone product on the same demo.

Budget owner: the owner or general manager; in groups, the operations or digital lead, with the POS administrator signing off on the integration and menu sync.

KPIs

Sample fifty of today's phone orders against the kitchen ticket before you deploy, and log missed calls by half-hour for one week. Those two numbers are the business case. Then track the strip above: order accuracy at the ticket, answer rate at peak, order containment, allergy handling, pickup-time honesty and voice-to-voice latency on the turns that read the menu or write the order.

Two traps. "Accurate" means the ticket matches the caller's final confirmation, not the agent's first read-back, so audit against the recording. And a containment number that rises because the agent stopped asking about modifiers is an accuracy problem wearing a containment costume; read the two together.

Demo script

Run the full script from the demo guide. The condensed version, with the traps that separate a product from a demo:

1. Greeting and mode. Call during a simulated service. Pass: restaurant name, AI disclosure, recording notice, and "pickup or delivery?" before any item is taken.
2. Modifiers with interruption. Order two mains with four modifiers between them, then cut in during the read-back with "no, the curry is the extra spicy one". Pass: it stops mid-sentence, moves the modifier to the right item and restarts the read-back from there. Fail: it finishes the wrong read-back.
3. Quantity change within ten seconds. Three seconds after ordering one curry say "actually make that two". Pass: one line updated, confirmed in a few words, no duplicate line and no second confirmation of the same change.
4. Ambiguous time. Ask for the order "tomorrow evening". Pass: it names the date and offers the slots the kitchen actually has, such as 6.30 or 7.15, rather than accepting "evening" as a time.
5. Allergy question. Say your child has a peanut allergy and ask whether the satay is safe. Pass: ingredients read from menu data, your configured wording spoken, the allergy written as a kitchen note, alternatives offered, no guarantee. Fail: "yes, that should be fine."
6. Discontinued item. Order an item from an old printed menu. Pass: it says the item is no longer available and offers the closest current items with prices. Fail: it accepts an order the POS cannot hold.
7. Eight seconds of silence after it asks for a name. Pass: one short prompt, then a graceful hold or offer to call back. Fail: it hangs up or repeats the whole read-back.
8. Mumbled quantity with noise. From a car with the radio on, trail off on "and two of the...". When the agent reads back three, say "yeah". Pass: the itemised total before the write gives you a second chance and the correction is taken. Fail: the wrong quantity reaches the POS.
9. Out-of-scope request. Ask it to book a table for twenty next month. Pass: it says it handles orders, offers to transfer or take a message, and does not invent a booking.
10. Pickup time, confirmation and write. Pass: a clock time from the sandbox load, an itemised read-back, a plain yes, the order on the POS screen with the allergy note on the ticket, and the payment link in the SMS log. Then say "can I give you my card now?" Pass: it declines per your rule and points to the link.

Score each trap pass or fail. A safety guarantee in turn 5, a card number requested in turn 10, or an order written that differs from the final read-back is a hard stop.

Compliance notes

This use case is inbound, and a call the customer places is outside the TCPA consent rule in the United States. Three other things are live. Recording: federal law is one-party consent but roughly a dozen states require all-party consent, so announce it. Payment: a card number spoken to the agent puts the recording and every vendor in the audio path inside PCI DSS scope; the PCI Security Standards Council's telephone-payment guidance describes pause-and-resume recording and DTMF masking, and a payment link avoids the question. Allergens: federal law defines nine major allergens, with sesame since 1 January 2023; the labelling rule does not cover food wrapped after a customer orders, but a wrong answer is still your liability, so the agent reads data and never guarantees. Any order-ready text or callback from the agent is an outbound artificial-voice or text contact and needs prior express consent and the 8 a.m. to 9 p.m. window. In the United Kingdom, the business must provide allergen information on 14 allergens, and since 1 October 2021 prepacked-for-direct-sale food carries full ingredient labelling under Natasha's Law; recording is personal-data processing under UK GDPR, and any marketing callback needs specific prior consent under PECR Regulation 19. In the European Union, the AI Act's Article 50 duty to tell people they are talking to an AI applies from 2 August 2026, and the same 14-allergen duty applies to restaurant food. In Australia, state surveillance-devices laws differ, so announce recording, and 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 a packaged product if you are a single-site or small-group restaurant on a mainstream POS; the menu sync, the modifier model and the payment-link flow are the hard parts and vendors have done them, including several POS and reservation platforms with a bundled agent. Consider a platform or a build if you run many locations with per-site menus, a custom upsell policy or an allergen policy that legal has signed off word for word. In both cases the acceptance test is the same: an order on your own sandbox POS during the demo with the allergy note on the kitchen ticket, and a quantity change that produces one line, not two.