AI for Restaurants: Orders, Reservations and FAQs on WhatsApp

AI for Restaurants: Orders, Reservations and FAQs on WhatsApp

AI for restaurants works best when it takes over the repetitive WhatsApp work: questions about the menu and opening hours, order collection and reservation requests. Decisions that carry risk, such as allergies, refunds and complaints, stay with a person. What separates a good setup from a problem is control: the AI answers only from approved information and checks the restaurant's own systems for anything that changes during the day.

Key takeaways

• AI handles repetitive volume well: opening hours, address, menu, order status and reservation requests.
• Table availability, prices and sold-out items must come from the restaurant's system, never from the model's memory.
• Allergies, intolerances and complaints need a clear rule for handing the conversation to a person.
• The menu and opening hours need an owner and an update routine, otherwise the AI repeats old information with great confidence.
• The safe path is to map current customer service, limit data and actions, pilot one flow, and only then decide whether to expand.

Where AI genuinely helps on a restaurant's WhatsApp

WhatsApp is already where customers ask, order and complain. The problem for the team is that the same questions arrive constantly, at the worst moments: the lunch rush, Friday night, a public holiday. Whoever is on the floor or in the kitchen cannot answer calmly, and the delay turns into a lost sale or an annoyed customer.

Realistic uses fall into three groups:

• Questions: opening hours, address, parking, payment methods, whether pets are allowed, whether there is a vegetarian option, whether you deliver to a given neighborhood.
• Orders: building the order through conversation, confirming items and address, and passing it to the ordering system or a staff member fully organized.
• Reservations: collecting date, time, party size and notes, and confirming only if the reservation system confirms.

Three things stay out at the start: negotiating compensation, stating that a dish does not contain a particular ingredient, and promising timings the kitchen does not control. When these topics come up, the conversation goes to a person with a summary already prepared.

How AI should handle orders without making anything up

An order has fixed parts and language parts. The language part is understanding that the customer wrote half pepperoni, half four cheese, no onion, delivered to a certain street. The fixed part is knowing whether the item exists, whether the restaurant delivers to that address, what the price is and whether the payment was accepted. AI is good at the first part. The second should be rules and system lookups.

In practice, that means a design where the model holds the conversation but the data comes from controlled sources:

• The menu is a structured database with name, description, extras, restrictions and availability status.
• Price and delivery fee are calculated by rule, not written by the model.
• The order is confirmed only after a summary that the customer explicitly approves.
• Sending the order to the ordering system uses a limited action: create an order with validated fields. The AI does not change prices or apply discounts on its own.

This technique of fetching information from an approved source before answering is called retrieval-augmented generation (RAG). A plain-language explainer is worth reading to understand why the menu should be looked up rather than memorized by the model.

Channel is an important decision. For small restaurants, the standard WhatsApp Business app may be enough at the beginning. As volume and integration needs grow, the API starts to make sense. If that question is yours, the article on WhatsApp Business API vs App for an AI Agent lays out the criteria.

Reservations: availability comes from the system, not the model

The most common mistake with reservations is letting the AI answer yes, we have a table without checking anything. The model does not know how many tables are free. It can only know if a tool or spreadsheet tells it. That is why the correct flow separates two steps: collecting the reservation request and confirming the reservation.

For each reservation, the AI collects:

• name and phone number (the WhatsApp number itself is usually enough)
• desired date and time
• number of guests
• occasion or note (birthday, high chair, accessibility)

Then it checks the booking calendar. If there is space, it confirms. If not, it offers alternatives that the calendar itself showed. If the group is large, there is a private event or the request is unusual, it passes the conversation to a person. The general logic is described in Automatic Appointment Scheduling on WhatsApp with AI, and it applies to table reservations too.

Two details prevent headaches. First, cancellations and changes must follow the same rule as creating a booking, with clear confirmation. Second, the AI should never promise a specific table or area of the restaurant if that is not recorded in the system.

Menu and opening hours: how to avoid outdated answers

The biggest source of error in a restaurant is not the model hallucinating. It is outdated information. The dish left the menu, the holiday hours changed, the kitchen ran out of an ingredient. If the AI answers from an old document, it is wrong in a consistent and polite way, which is worse.

The control points for this part are simple, but they need an owner:

• A single source of truth for menu, prices, opening hours and delivery areas.
• A named person responsible for updating that source, with the rule that a menu change only counts once it is recorded there.
• A sold-out alert that staff can trigger quickly, without opening a complicated dashboard.
• Special hours (holidays, events, early closing) entered with start and end dates.
• A safety rule: if the information is not in the database, the AI says it will confirm with the team instead of guessing.

That last rule is the most valuable one. A reply like I'll check with the team and get right back to you is far better than a wrong opening time. It is also worth testing the AI with questions that have no answer in the database, to see whether it admits its limit.

Allergies and dietary restrictions: the point you do not automate

A question about gluten, lactose, peanuts or shellfish looks like a menu question, but it carries a real health risk. The AI can list what is officially described on the menu, if the restaurant keeps that information on file. It should not state that a dish is safe for someone with an allergy, nor suggest substitutions the kitchen has not validated.

The prudent design is this: the AI reports what is on record, says that final confirmation belongs to the kitchen, and hands the conversation to a person when the restriction is serious. That handoff should carry the customer's question, the dish in question and the time, so nobody has to repeat everything.

Pay attention to privacy as well. Health information may require extra care under LGPD, Brazil's data protection law. Record only what is needed to fulfil the order and avoid keeping conversations longer than the operation requires. The guide on LGPD Compliance for AI on WhatsApp covers consent, legal basis and records, and ANPD, Brazil's data protection authority, is the official source for guidance.

Complaints: when the AI stops and a person takes over

A restaurant complaint is usually about something that has already gone wrong: a late order, a cold dish, a missing item, a double charge. The customer wants to be heard by someone with the power to fix it. An AI that insists on standard replies makes the situation worse.

The AI's role here is small and useful: acknowledge the problem, collect the facts and hand everything to whoever decides. That includes the order number, what happened, photos if the customer sends them and what they expect as a resolution. The AI does not offer a refund, coupon or compensation unless an approved, limited rule exists for it.

Triggers that should cause an immediate handoff:

• strong expressions of dissatisfaction, threats of a public review or mention of a consumer protection agency
• any report of feeling unwell, suspected food poisoning or an allergic reaction
• a request for a refund, reversal or a disputed charge
• a customer repeating the same complaint a second time
• an explicit request to talk to a person

The playbook on WhatsApp AI Agent Human Handoff explains how to build that transition without losing context. A good handoff means the customer does not have to tell the story again.

Control point checklist before you switch the agent on

Use this list as an exit criterion before the agent talks to real customers:

• Current customer service has been mapped: the most frequent questions, the peak hours and who answers today.
• There is a single, up-to-date source for menu, prices, opening hours and delivery areas, with an owner.
• The AI consults only that source and says it does not know when the information is missing.
• Price, delivery fee and table availability come from rules or systems, not from the model's text.
• Allergies, complaints, refunds and large groups have a written handoff rule.
• The customer knows they are talking to an automated assistant and can ask for a person.
• Conversations and decisions are logged, with a defined retention period and restricted access.
• The AI cannot perform actions beyond those planned (create an order, record a reservation).
• Someone reviews a sample of conversations every week during the pilot.

On security, remember that customer messages are untrusted input. A customer might, as a joke or in bad faith, try to talk the assistant into giving a discount or ignoring its rules. The OWASP Top 10 for LLM applications describes this kind of risk and reinforces the practical rule: permission lives in the system, not in the model's goodwill.

How to apply it: map, limit, pilot and hand over

The method for starting without creating risk has four steps.

Map: list what arrives on WhatsApp today, what is a repeated question, what is an order, what is a complaint and how much time the team spends on each type. Record the starting point too, so you can compare later.

Limit: define which data the AI can see and which actions it can take. For example, read the menu and opening hours, record a reservation request, create an order with validated fields. No refunds, discounts or promises about allergies.

Pilot: choose a single flow, such as opening hours and menu questions, or reservations. Run it with logs, weekly review and a metric agreed in advance, such as the share of conversations resolved without help or the errors found in review. The article on How to Measure an AI Automation Pilot helps define those criteria.

Hand over: at the end of the pilot, give the team the documentation, access to the logs and the menu update routine. Then decide with data whether to expand, adjust or stop.

Illustrative scenario: a neighborhood restaurant receives many messages asking whether it opens on the holiday and whether there is a table on Saturday. The pilot starts with only opening hours and menu, using a spreadsheet maintained by the manager. In the first week, the review shows that the AI answered the menu questions correctly but gave an old time for a holiday that had not yet been entered. The fix was not in the model, it was in the process: every change in opening hours is now recorded in the spreadsheet before it is announced. Only after that did the restaurant add reservation requests to the flow.

For teams that want to understand how to organize agents with simple tools and clear limits, Anthropic's guide on building effective agents recommends starting with the simplest design that solves the problem, which fits well with the reality of a restaurant.

Frequently asked questions

Does AI for restaurants replace the person answering WhatsApp? Not at the start and, for most restaurants, not entirely. It absorbs the repeated questions and organizes orders and reservations, while the team handles complaints, exceptions and anything involving judgment or risk. The gain is freeing up time at peak hours, not taking the person out of the conversation.

Can the AI answer questions about allergies and ingredients? It can repeat what is recorded in the official menu, making clear that final confirmation belongs to the kitchen. It should not state that a dish is safe or suggest unvalidated substitutions. For serious restrictions, the right move is to pass the conversation to a person with a summary of the question.

How do I stop the AI from giving the wrong opening time or price? Keep a single source for menu, prices and opening hours with someone responsible for updating it, and require the AI to answer only from it. When the information does not exist, the correct reply is to say it will check with the team. Testing questions that have no answer in the database and reviewing conversations every week completes the control.

Do I need to worry about LGPD when using AI for restaurant customer service? Yes, because the conversation involves personal data such as name, phone number, address and sometimes health information linked to allergies. Collect only what is necessary, tell customers the service is automated, define how long you keep conversations and restrict access. If you are unsure about the legal basis, seek legal advice and consult ANPD's material.

Sources

• AWS: What is retrieval-augmented generation
• OWASP Top 10 for LLM Applications
• Anthropic: Building effective agents
• ANPD: Autoridade Nacional de Proteção de Dados