Skip to content
Giallo

Restaurant chatbot: agents that answer guests while the floor team works.

What can a restaurant chatbot do besides taking bookings?

A well-built restaurant chatbot is a team of AI agents: one handles tables and the waitlist, one answers allergen questions by reading only the kitchen's allergen sheet, one drafts replies to reviews, one reads supplier delivery notes. Giallo Studio builds them with Claude, GPT and Gemini; menu, prices and public replies remain the owner's decisions.

We write for owners, floor managers and chefs who get requests on WhatsApp, Instagram and the phone right in the middle of service. Here you'll find the agents we design for a restaurant, the software they talk to, the model we pick for each task and the points where you have the final say.

Three round white plates stacked and offset. On the front plate, a yellow sphere, the symbol of the AI agent, gives off a thin wisp of steam

The AI agents we design, one by one.

The dining room from above: the tables agent reads “4 of us, outside” and takes the patio table; when a table for two opens up, it messages the first person on the waitlist and, with no reply within the set time, moves on to the second, who takes the seat.

Tables and waitlist agent

Takes bookings from written channels, offers alternatives when a seating is full and contacts the waitlist when a spot opens up.

When a table request comes in and when someone cancels.

  • Booking system
  • WhatsApp Business Platform
  • Instagram Direct
  • Website chat
Claude Haiku
Table requests are short, frequent and often typed in a hurry: we need a fast model that understands “tomorrow night, four of us, outside if there's room” without a high cost per message.
GPT
During testing we compare it with Claude Haiku on the same anonymized chats from last season, to pick the one that gets fewer dates and party sizes wrong.
How it works, step by step
  1. Picks out date, seating, party size and special requests such as a high chair or an outdoor table from the chat.
  2. Checks availability in the booking system, following the restaurant's rules on joined tables and seatings.
  3. Confirms the table or offers the closest free time and, if the guest agrees, puts them on the waitlist.
  4. On the day of the dinner it sends the reminder using the approved template and records who cancels.
  5. When a table opens up, it messages the first person on the list and moves on to the next if no reply arrives within the time the restaurant has set.

Who approves. The manager sets capacity, seatings and group rules once. Large parties, special nights and deposit requests always go to a person.

Menu and allergens agent

Answers on ingredients and allergens by reading only the sheet written by the kitchen, and passes every doubtful case to the chef.

Every time a guest asks what's in a dish, before booking or ordering takeout.

  • Kitchen allergen sheet
  • Digital menu
  • WhatsApp Business Platform
  • Website chat
Claude Sonnet
It strictly follows a long instruction that forbids inferring ingredients that aren't written down: when the right answer is “let me ask the chef,” we prefer a model that knows how to stop.
How it works, step by step
  1. Receives from the kitchen the dish sheet with allergens, signed by the person in charge and dated.
  2. When a guest asks about a dish, it looks up the entry in the sheet and lists the allergens present without adding anything.
  3. If the dish isn't on the sheet, if the recipe is marked as under review or if the guest mentions a severe allergy, it doesn't answer on the substance.
  4. Forwards the question to the chef with the dish name and the exact request, and lets the guest know a person will reply.

Who approves. The chef approves the allergen sheet at every menu change; the agent doesn't answer about a dish until the new version is confirmed.

Reviews agent

Collects new reviews, groups them by topic and drafts a reply that the owner reviews before publishing it.

Every morning, with the reviews received since closing the day before.

  • Google Business Profile
  • Owner's dashboard
  • Restaurant email inbox
Gemini Flash
It sorts all new reviews by topic in a quick, low-cost pass, so only the ones that need a reply go on to the writing model.
Claude Sonnet
It writes measured replies even to harsh criticism, with no automatic apologies or defensive tone, and respects the words the owner never wants to use.
How it works, step by step
  1. Downloads new reviews from the Google Business Profile through the official APIs and collects those from the other platforms the restaurant specifies.
  2. Sorts them by topic: service, wait, dishes, price, cleanliness, booking.
  3. Writes a draft for each one in the tone chosen by the owner, without promising discounts or freebies.
  4. Flags separately the reviews that seem to refer to another business or that mention staff members by name.
  5. After approval it publishes the reply on Google; for the other platforms it prepares the text to paste.

Who approves. No public reply goes out without a click from the owner or whoever they delegate.

Delivery notes and suppliers agent

Reads delivery notes and invoices, matches them to orders and flags quantities or prices that don't match.

When a delivery reaches the kitchen and when the supplier's e-invoice arrives.

  • Delivery note photos
  • E-invoices
  • Order sheet
  • Custom connector
GPT
It extracts structured data from very different documents, including crooked photos of printed delivery notes, with output checked field by field.
Codex
We use it to build the nightly sync between supplier invoices in the tax portal and the order sheet: it generates the script, and we review it line by line before it runs on your management software.
How it works, step by step
  1. Receives the photo of the delivery note taken when the goods arrive, and the e-invoice from the tax portal (cassetto fiscale) or from the accountant.
  2. Extracts supplier, items, quantities and unit prices into a table with fixed fields.
  3. Compares the data with the order sent and with the agreed price list.
  4. Prepares a list of discrepancies for the purchasing manager, with the document attached to each line.

Who approves. The agent never disputes anything with the supplier: the purchasing manager assesses the discrepancy and decides whether to call, request a credit note or let it go.

Before and after the agent, in everyday work.

At half past eight nobody can check the phone

Today. Table requests come in by DM, on WhatsApp and to voicemail just as service starts. Whoever is on the floor replies at the end of the night, when the guest has already booked somewhere else.

With the agent. An agent reads availability from the booking system and replies right away, offering other times if needed. The floor team only sees the requests that break the rules, already summarized.

The gluten question reaches the wrong server

Today. A guest asks in chat whether the risotto contains milk or tree nuts. Whoever is holding the phone answers from memory, even if the recipe changed last week.

With the agent. The agent answers only by quoting the allergen sheet kept up to date by the kitchen. If the sheet isn't enough, or the guest mentions celiac disease or a severe allergy, the question goes to the chef.

Reviews go unanswered for weeks

Today. The owner reads reviews on the phone between one order and the next and puts off the negative ones because they need calm and time. Meanwhile, the restaurant's page stays silent.

With the agent. Every morning they find the new reviews grouped by topic, each with a draft reply in the restaurant's tone. They read, edit and publish: the judgment on what to say stays theirs.

Paper delivery notes and invoices that don't add up

Today. Deliveries arrive with documents in a different format for each supplier. At month end someone compares ordered, delivered and invoiced quantities by hand, and discrepancies surface late.

With the agent. An agent reads photos of delivery notes and e-invoices, matches them to the order and flags only the lines that don't match. Whether to dispute anything with the supplier is up to the purchasing manager.

Frequently asked questions, straight answers.

Yes, if it works on the booking system the restaurant already uses. The agent checks real availability, confirms or suggests another time and records the table; large groups and nights that require a deposit are always confirmed by a person. For service messages it uses approved templates on the WhatsApp Business Platform.

Usually not. We check what your software offers through APIs or exports and build the connection with limited permissions. If the system exposes nothing, the agent prepares the booking and the floor team enters it with one click.

Only by repeating what the kitchen sheet says about that dish. If the guest has celiac disease or mentions a severe allergy, the question goes to the chef, because only the people at the stove know about cross-contamination and shared preparation.

It depends on the task. For table requests we use a fast model like Claude Haiku, for allergens and reviews Claude Sonnet, and for reading delivery notes and invoices GPT. The final choice is made during testing, on the same anonymized chats.

No. It drafts a reply in the restaurant's tone and queues it: the owner reads it, edits it and decides whether to publish. For reviews that seem to refer to another restaurant, the agent flags the doubt without replying.

Yes, with different rules. The agent collects pizzas, variations and pickup time, shows allergens from the sheet before confirming and forwards off-menu requests to the pizza chef. Delivery times are set by the kitchen, not by the chatbot.

Almost always tables and the waitlist, because that's the work that interrupts service today. On the first call we look at which channels requests come from and what your booking system allows, then we choose together.

How we work with you.

  1. A Saturday service from the pass

    We stand between the dining room and the kitchen during a full service to see when messages come in, who answers and which questions interrupt the work.

  2. The sheet first, then the chat

    With the chef we put dishes, allergens and possible variations in writing. Without this foundation the menu agent doesn't start.

  3. Written floor rules

    With the manager we set seatings, tables that can be joined, groups, deposits and phrases the agent must never use with a guest.

  4. A test on last season's chats

    We run the agents on real, anonymized conversations and compare the models before opening a channel to guests.

  5. One channel at a time

    We start with the channel that brings in the most requests, usually WhatsApp or Instagram, and add the next one when the floor team trusts the first.

Covers, seatings, suppliers: what's your dining room like?

Tell us which channels bookings and menu questions come from today and which booking system you use: we'll start from the service that has your phone ringing mid-order.

Prefer the full form? Go to the contact page.

No spam or sales pitches. Response within 1 business day.

We use this data to contact you and, if we work together, to manage the project. We store it in Giallo's internal management system, on servers in the European Union, where only our team can see it; by email we only receive a notice, without your data. If you do not become a client, we delete it 3 months after our last contact. Details and your rights: privacy policy.

Do you prefer to write or call us?

AI agents for restaurants

Which messages do you want to take off your floor team's hands?

Tell us which channels bookings and menu questions come from today and which booking system you use: we'll start from the service that has your phone ringing mid-order.