How restaurants can use AI chat on their website, WhatsApp and Messenger to handle booking requests, menu questions and hours, with staff in control.
A restaurant's busiest hours for questions are rarely its quietest hours for staff. Guests ask about tables while the kitchen is mid-service, or at eleven at night when nobody is at the phone. This article shows how an AI chat agent can take the repetitive work off your team, which questions to let it answer, where to draw the line, and how to hand a guest to a person without losing the booking.
Why restaurants lose bookings between the lines
A restaurant is a business where the person best placed to answer a message is usually busy doing something else. The host is seating a party, the owner is on the pass, and the phone rings during the lunch rush.
A guest who does not get a quick answer rarely waits. They open another tab and message the next restaurant on the list. The loss is not dramatic, so it is easy to miss: a few unanswered messages a week, each one a table that went elsewhere.
The questions are also highly repetitive. Most of them have the same answers every day:
- Are you open tonight, and until when?
- Do you have a table for four on Saturday?
- Is there a vegetarian or gluten-free option?
- Can I bring a stroller, a dog, or a birthday cake?
- Where do I park?
An AI agent is well suited to this kind of repetition. It answers from your own information, at any hour, in the guest's language. It is much less suited to judgment calls, such as a serious allergy, a complaint, or a large-party negotiation. A good setup respects that boundary.
What guests actually ask, and who should answer
Before building anything, sort your incoming questions into three groups. This sorting decides almost everything about how the agent is configured.
| Type of question | Example | Best handled by |
|---|---|---|
| Fixed facts | Opening hours, address, parking, dress code | AI, from your knowledge base |
| Menu and policy | Vegetarian options, corkage, kids' menu, cancellation rules | AI, from your menu and policy documents |
| Booking requests | Table for six on Friday at 7 pm | AI collects the details, staff confirm |
| Sensitive or complex | Severe allergy, complaint, wedding buyout, refund | A person, with the AI preparing the handoff |
The third and fourth rows matter most. The AI should be good at gathering the guest's name, party size, date, time and contact details so that nobody has to ask twice. It should not promise a table it cannot see, and it should not give medical-grade assurances about allergens.
Bookings: collect the request, confirm with certainty
The most important design decision is what "booking" means in your chat. There are two honest models.
Booking request. The AI gathers the details and passes them to your team. Staff check the real table plan and confirm by message, phone or email. This works for almost every independent restaurant, because the table plan lives in your head, a notebook or a reservation tool the chat agent does not control.
Slot-based booking. For things that really do run on a calendar, such as a chef's counter seating, a tasting event or a private-dining viewing, a scheduling link lets the guest pick an open slot themselves.
Be clear in the chat about which one is happening. A message such as "I have your request for four people on Saturday at 7 pm; the team will confirm shortly" sets the right expectation. A message that says "You're booked!" when nothing has been checked creates double-bookings and angry arrivals.
- 01Guest asks about a table
- 02AI collects date, time, party size, contact
- 03Team receives the request and alert
- 04Staff confirm against the real table plan
Allergies, complaints and the handoff
Food allergies are the clearest case for a firm rule. A guest who writes "my daughter has a severe nut allergy" should not receive a casual chatbot reassurance. The right behaviour is to acknowledge the message, say that the kitchen team will respond personally, and route it to a human immediately.
You can encode this in the agent's persona and guardrails. For example, instruct it to:
- Never state that a dish is safe for a specific allergy.
- Share only the allergen information you have published in your own documents.
- Pass any mention of a serious allergy, a complaint or an incident to staff.
The handoff itself deserves care. If the guest is told "someone will be with you in five minutes" and nobody is on shift, trust is damaged. It is better when the message about what happens next reflects your actual staffing, for instance who is on duty now and when the team next replies. Evening and weekend coverage is where restaurants most often fall short, so plan a rota rather than hoping someone sees the alert.
Put the agent where guests already message you
Restaurant guests do not use one channel. A tourist uses WhatsApp, a regular uses Messenger or Instagram, and someone browsing at lunchtime clicks the chat bubble on your website. In Japan, LINE is often the first choice.
A single agent with a single knowledge base can serve several of these channels, which keeps your answers consistent. You write the opening hours once, not once per app.
Think about each channel's habits:
- Website chat: best for people deciding right now. Keep the welcome message short and offer the two things they most want, booking and menu.
- WhatsApp: common for international guests and group organisers. Messages are conversational and often come in bursts.
- Messenger and Instagram: guests often arrive from a photo or a post and ask a quick question. Fast answers matter more than long ones.
- Phone: some guests still call. An AI voice line can pick up when the restaurant is closed or busy, though it needs separate setup and different testing.
Serving guests in their own language
Restaurants in tourist areas, hotel districts and airports field questions in many languages. Staff cannot always match them, and a guest who cannot read your English-only page may simply leave.
An AI agent can answer in the language the guest writes in, drawing on the same menu and policy documents. Two habits keep this reliable. First, write your source documents clearly in one language, with plain wording for dish names and ingredients. Second, test with real examples in the languages your guests actually use, and have a speaker read the replies. Dish names, regional ingredients and dietary terms are where translation errors do the most damage.
Estimating how much chat you will handle
Credit-based pricing is easier to plan if you estimate your volume before choosing a plan. In AI SNS CHAT, one reply uses one credit, so the arithmetic is simple.
Here is a hypothetical example, not a benchmark. Suppose a restaurant receives 300 chats in a month and the agent sends about four replies per chat on average. That is roughly 1,200 replies, or 1,200 credits. Your own numbers will differ, so count a typical week of messages first and scale it up.
Seasonal restaurants should plan for the peak month, not the average. Holiday periods, local festivals and a mention in the press can all multiply your chat volume in a short time.
Keep the answers accurate over time
An AI agent is only as good as the information behind it. Menus change, kitchens close for a private event, and the summer terrace opens in May. Build a small routine:
- Review the knowledge base whenever the menu or hours change.
- Read a sample of real conversations every week at first, then every month.
- When the AI gives a weak answer, rewrite it as the answer you wish it had given, so it improves.
- Add temporary notices, such as holiday closures, as clearly dated text, and remove them afterwards.
Reading real conversations is the best training you have. You will quickly see which questions you never thought to document.
Doing this in AI SNS CHAT
Here is a practical path from nothing to a working restaurant agent.
1. Create the agent. Use the guided wizard to create an agent and choose a starter persona, then adjust the persona to your tone and your rules: never confirm allergen safety, always pass complaints to staff, and describe bookings as requests until the team confirms. You can test any change in the Playground before guests see it.
2. Add your knowledge. Under Data sources, upload your menu and policy files, add the website you already have, and enter short text snippets for things like parking and dress code. The Q&A source is useful for your top ten questions with exact wording. Keep wording consistent with your menu so the agent does not contradict it.
3. Capture booking requests. Use the Collect leads Action or a Custom form to ask for name, party size, date, time and contact details. Each submission appears under Activity → Leads, and every lead also creates a contact. Contact custom attributes can describe when the agent should ask for a field, for example asking about dietary needs only when relevant. If you offer slot-based experiences, the built-in Cal and Calendly Actions can provide scheduling. Integrations need the Hobby plan or higher.
4. Connect your channels. Under Channels you can deploy the chat widget on your website, and connect WhatsApp, Messenger, Instagram and LINE, among others. See the connections page for the full list of what is available and what is still being built.
5. Plan the handoff. Under Helpdesk → Scheduling, set shifts per person and timezone, and handoffs are assigned to whoever is on shift. Alerts can go to Slack, Teams, LINE WORKS, WhatsApp, email or a webhook. When a chat is handed over, the customer message about what happens next is built from your rota and never invents a wait time. Helpdesk needs the Standard plan or higher.
6. Step in and improve. In Activity → Chat logs you can filter, flag and escalate conversations. A staff member can take over a chat, which pauses the bot, and can edit a reply into the ideal answer to train the agent. The welcome message is translated automatically into every language the widget speaks.
7. Optional extras. Inbound AI phone calls use a Twilio number connected per agent. The Google reviews channel can draft replies to reviews, but it requires Google API approval. Outbound campaigns need Standard or higher and only reach people who messaged you first, and every message carries a STOP opt-out.
Pricing and plan limits are on the pricing page, and the features page gives the wider overview. You can start on the Free plan, which includes one agent, and sign up at /signup. Privacy practices are described on the security page.
Frequently asked questions
Can the AI confirm a table directly?
It is best used to collect the request: name, party size, date, time and contact details. Your team then confirms against the real table plan. For slot-based events, the Cal and Calendly Actions can offer a scheduling link instead, on the Hobby plan or higher.
How should the agent handle food allergies?
Set a firm rule in the persona that it never states a dish is safe for a specific allergy, shares only the allergen information you have published, and routes any mention of a serious allergy to staff. A person should always make the final call.
Which channels can a restaurant use?
The chat widget on your website, WhatsApp, Messenger, Instagram and LINE are all in the channel list, along with many others. One agent and one knowledge base can serve several channels so your answers stay consistent.
What happens when nobody is on shift?
Set shifts under Helpdesk → Scheduling, which needs the Standard plan or higher. The handoff message the customer sees is based on your rota and never invents a wait time, so the guest gets an honest idea of what happens next.


