Conversational AI in restaurants, in plain terms

Conversational AI is software that holds a back-and-forth conversation with a guest in everyday words, by voice or by text, and then acts on it: it puts an order into the POS, books a table or passes the guest to staff. In restaurants it answers phones, takes drive-thru and kiosk orders, and replies to chat messages. It is built from a handful of parts, and each part fails in its own way.

Our guide to AI agents for restaurants covers what these agents do in each channel, what they cost and how to trial one. This page opens the box: the parts, the words vendors use for them, and the failure each part can cause, so you can follow a sales demo and test the right thing.

What are the parts of a conversational AI system?

A voice agent runs through these parts on every turn of the conversation. A text chatbot skips speech recognition and voice output.

Part	What it does, in plain words	What vendors call it	How it can fail
Line or microphone	Carries the guest's voice in: a phone line, a speaker post or a kiosk microphone	Telephony, audio capture	Road noise, a bad line, two people talking at once
Speech recognition	Turns the guest's speech into written words	ASR, speech-to-text	Mishears an accent, a dish name or a number
Language model	Works out what the guest wants and writes the next reply	LLM, NLU, intent detection	Misreads the request, or gives a confident answer that is wrong
Restaurant knowledge	The menu, prices, hours and rules the model may use	Knowledge base, grounding	An old menu, a missing rule, a price from another store
Order memory	Keeps track of the cart and what the guest has changed	Dialogue state, context	Keeps an item the guest removed, or adds one twice
Actions	Sends the order to the POS, books the table, transfers the call	Integration, API, tool use	The POS rejects the order, or it goes to the wrong store
Voice output	Reads the reply aloud	TTS, text-to-speech	Replies slowly, or talks over a guest who interrupts
Handoff	Passes the guest to a person	Escalation, transfer, human in the loop	Nobody picks up, or the guest cannot find the way out

What do the other common terms mean?

• Intent: what the guest is trying to do, such as order, ask the hours or book a table.
• Slot or entity: a detail the AI must fill in, such as the size, the pickup time or the number in the party.
• Large language model (LLM): a model trained on a huge amount of text that can read a request and write a reply.
• Hallucination: a fluent answer the model made up. It is the main reason an agent must be tied to your own menu and rules.
• Grounding: tying the model's answers to your own data, such as the POS menu, so it answers from your facts.
• Latency: the pause between the guest finishing and the AI starting to reply.
• Barge-in: the guest cutting in while the AI is still talking, and the AI stopping to listen.
• Containment: a vendor's word for conversations finished with no staff involved. Ask how they count it.
• API: the connection one system uses to send data to another, such as an order into your POS.

Square's help article on its AI-powered voice ordering shows grounding and actions in a real product. It says its phone AI combines generative AI with your business information, and that an order reaches the POS only after the guest pays through a link sent by text during the call.

How does it differ by channel?

Channel	What is different	Example from a vendor's own page
Phone	No screen, so the AI must read the order back; calls arrive at all hours	Maple's phone product answers questions, transfers calls and, on its Pro plan, takes orders into supported POS systems
Drive-thru	Outdoor noise, engines and wind; the lane sets the pace	Maple's drive-thru unit has its own directional microphone and a screen showing the cart, with a one-press switch back to the crew; it is in early access
Kiosk	Indoor crowd noise and other guests talking nearby	Maple's voice kiosk uses a microphone built to pick out the guest in front and builds the cart on screen; it is in early access
Chat	Typed words, so no speech recognition or voice; the guest can scroll back and reread the order	Our chatbot guide compares website, text and social message tools

Where does conversational AI fail in restaurants?

Public cases line up with the parts above.

Knowledge and grounding. In March 2024, The Markup found New York City's business chatbot telling owners they could make a restaurant cash-free, though a city law requires businesses to accept cash. The model gave a fluent answer that no source backed.

Order rules. A Taco Bell guest ordered 18,000 water cups at a voice AI drive-thru for a social media video, and an employee took over, Nation's Restaurant News reported in September 2025. The checks this case points to are a quantity cap and a rule that sends strange orders to a person.

Handoff. In September 2026, Burger King said some guests in its pilot drive-thrus drove off rather than talk to a bot, and that it is building an easier way to order from a person, Nation's Restaurant News reported.

Our review of public restaurant AI setbacks covers more dated cases, including McDonald's test with IBM.

How can you test each part in a demo?

1. Line or microphone: call from a car with the window down, or order at the kiosk while two people talk beside you.
2. Speech recognition: order your hardest dish names and a number such as "two twelve-inch".
3. Language model: ask the same question two ways, then ask something off topic.
4. Knowledge: change a price in the POS, then order that item and listen for the new price.
5. Order memory: add an item, remove it, change a modifier, and compare the final ticket line by line.
6. Actions: check that the order reaches the POS and the kitchen screen at the right store, with the right pickup time.
7. Voice output and barge-in: interrupt the AI mid-sentence, then time the pause before each reply.
8. Handoff: ask for a person with staff free and again with nobody free, and note what the guest hears both times.

Our voice AI comparison turns these into a shared script to run with every vendor. If the phone is where you would start, the Maple phone ordering page sets out a trial from call to kitchen, and Maple pricing shows both phone plans.

Published by Maple. This AI-assisted explainer combines vendors' own product and help pages, reporting by The Markup and Nation's Restaurant News, and Maple's published product pages with original definitions and a test list. It does not report a test Maple ran or promise any result.