Restaurant messages arrive exactly when nobody is free to answer them. At 7:30 pm, in the middle of the dinner rush, a guest asks on WhatsApp whether there is a table for four in an hour. At 10:45 pm a delivery customer wants to know if the sauce contains nuts. Just after midnight someone sends a photo of a cold biryani with the caption "is this what you call delivery?". The host is seating guests, the manager is in the kitchen, the delivery coordinator is on the phone with a rider. The messages wait, and in the meantime the guest books somewhere else or leaves a one-star review on Google Maps.
This article shows how to split a restaurant's incoming messages into three streams — bookings, menu questions and complaints — and what you can safely hand to AI in each one, and what has to stay with your team. At the end there is a checklist of the data you need before you switch anything on.
Three streams that should never share one set of rules
A common mistake is a single "bot for everything" that answers a parking question and a food poisoning complaint in the same cheerful tone. In hospitality, the cost of a wrong answer varies hugely between message types, so the rules have to vary too.
- Bookings. High volume, structured data: date, time, party size, area, occasion. A mistake (a double booking) is annoying but fixable.
- Menu questions. Ingredients, portion size, spice level, halal status, vegetarian and vegan options, allergens. Most questions repeat, but a wrong allergen answer can put someone in hospital.
- Complaints. Late riders, wrong items, food quality, staff behaviour. Rare compared with bookings, yet they decide whether a guest comes back and what they say in public.
Bookings: where AI can work almost on its own
Table reservations are the most automatable process in a restaurant — provided the AI can see live availability. That might be your reservation system, a module in your POS, or at minimum a shared sheet the host keeps up to date during service. Without live data, the AI will promise tables that do not exist.
What AI handles well:
- collecting booking details in one or two messages, and asking follow-up questions when the guest writes "tomorrow evening, there's a bunch of us";
- offering the nearest available slot when the requested time is full, instead of a flat "fully booked";
- sending a confirmation and a same-day reminder, and processing changes and cancellations;
- answering the side questions: dress code, kids' menu, valet or parking, whether guests can bring their own cake.
What goes to a person: private dining and venue hire, groups above a threshold you set, special requests ("we're proposing, can we have the window table?"), and anything involving deposits or prepayment. The AI gathers the details and hands the manager a ready summary, not just "someone messaged you".
A sample exchange:
Guest: Hi, do you have a table for 6 this Saturday around 8?
AI: Good evening! On Saturday at 8:00 pm we have a table for six in the main dining room. The earliest slot on the terrace is 9:30 pm. Which would you prefer? Could you also share a name and phone number for the booking?
Guest: Main room please. Sarah, +44 7… It's a birthday — can we bring a cake?
AI: Booked: Saturday, 8:00 pm, 6 guests, main dining room, under Sarah. On the cake: our manager will confirm the policy for outside desserts — I've passed your question on and they'll reply before the end of today. I'll send you a reminder on Saturday morning.
Notice that when there is no rule for a question, the AI does not invent an answer. It hands the question to a person and tells the guest when to expect a reply.
Menus and allergens: answer only from the dish card
Menu questions are the second-largest stream, especially for delivery: "how big is the portion", "is it spicy", "anything gluten-free", "is the chicken halal?". AI answers these instantly — if it has a recipe card, or at least a structured record of each dish with its ingredients.
The key rule is simple: the AI never guesses. A language model "knows" that classic pesto contains pine nuts, but your chef may have swapped them for sunflower seeds — or added cashews. So every answer is built only from your own dish card. If the card lacks the relevant field, the AI says: "The description of this dish doesn't confirm whether it contains nuts — let me check with the kitchen," and passes the question on.
In the UK, food businesses must be able to tell customers about the 14 regulated allergens in their dishes, and the Food Standards Agency guidance is the place to check what applies to you. An AI assistant can make that information easier to reach, but it does not take over your legal responsibility for it — the data behind it has to be correct.
Also plan for:
- Cross-contamination warnings. If the same kitchen handles nuts, sesame or gluten, the AI should say so rather than declare a dish "safe".
- The 86 list. If a dish has run out and the AI keeps recommending it, the guest is disappointed twice. Sold-out items must be updated in the same place the AI reads the menu from.
- Languages. In Dubai or London, guests write in English, Arabic, Russian, Hindi and more — often mixing languages in one message. Give dish names in your main languages so menu search does not break.
Delivery complaints: acknowledge fast, route correctly
A late-night complaint left unanswered until morning very often becomes a public review. Yet the AI should not decide on refunds or vouchers: that depends on the customer's history, the order value and whose fault it was — the kitchen, the rider or the delivery partner.
A workable flow:
- Acknowledge and apologise within minutes. Not "your feedback is important to us", but something specific: "I'm sorry your biryani arrived cold — that shouldn't happen."
- Collect the facts. Order number, what went wrong, a photo, delivery time. If the AI is connected to your order system, it pulls up the items and timings itself, so the customer does not have to repeat everything.
- Classify. Late delivery, wrong or missing items, food quality, rider behaviour, possible health issue.
- Route by priority. A suspected allergic reaction or food poisoning goes straight to the duty manager by phone, not to a shared inbox. Everything else goes into a queue with a ready-made summary.
- Say what happens next. "Our manager will contact you before 10 am" — and make sure that actually happens.
If you deliver through aggregators such as Deliveroo or Talabat, many complaints go to their support team, and you may not see those conversations at all. In that case AI is most useful in your own channels — WhatsApp, Instagram, your website — and in a weekly review of which complaint reasons repeat and on which shifts.
Who does what: AI versus your team
| Process | What AI does | What stays with people |
|---|---|---|
| Table booking | Collects details, checks availability, confirms, sends reminders | Private dining, deposits, special requests |
| Changes and cancellations | Processes them and frees the table | Disputes involving prepayment |
| Menu questions | Answers strictly from dish cards and the 86 list | Keeps cards up to date, answers what isn't on them |
| Allergens | Answers from ingredients, warns about cross-contamination | Confirms every non-obvious case with the kitchen |
| Delivery status | Reports status from the order system | Deals with the rider when something breaks |
| Complaint | Acknowledges, gathers facts, classifies, routes | Decides on compensation, calls the guest |
| Health-related report | Escalates immediately to the duty manager | Everything else |
Pre-launch checklist
A restaurant AI assistant is only as good as the data you give it. Before launch, check that:
- there is a single place that shows the floor plan and bookings — and it is updated during service;
- every dish has a card: ingredients, portion, spice level, allergens, halal status — and someone owns keeping it current;
- the 86 list lives in one system, not in the head chef's memory;
- house rules are written down: maximum party size without a manager, outside cake and corkage policy, last kitchen order time;
- a named person is on duty for urgent overnight escalations, and it is clear how they get alerted;
- you know where guest names and phone numbers are stored and on what legal basis. Data protection rules differ — UK GDPR, the UAE's federal data protection law, Uzbekistan's data localisation requirement — so confirm with a lawyer how they apply to the tools you choose.
If half of these items still live "in the manager's head", start there — it pays off even without AI. To estimate how many enquiries you lose to slow evening replies, try our loss calculator with your own message volume and average bill.
How a rollout usually goes
The sensible order is from low-risk to high-stakes. First the AI answers questions about opening hours, location and the menu, and takes bookings, while the manager can see every conversation. Then complaint intake is added, always with a hand-off to a person. Each week, review where the AI passed questions to staff — that is how dish cards and rules improve. We describe a similar approach for patient scheduling in our article on AI booking for clinics, and the difference between a simple bot and an AI agent in our piece for online stores.
The cost of a project like this depends on the number of channels and locations, and on whether you already have reservation and order systems; a first conversation is there to work that out. See how our AI agents work, or browse more sector breakdowns in the Industries section.
FAQ
Won't guests mind talking to AI instead of a person?
Waiting tends to annoy people more. What matters is not pretending the AI is human, keeping answers polite and to the point, and giving an easy route to a real manager — a keyword like "staff" or a button.
Can AI take delivery orders?
Yes, if it is connected to the menu, the 86 list and your order system or POS. Without integration it can collect the order and pass it to a coordinator, which still saves time but does not remove manual entry.
What about voice notes?
Guests often send voice messages on WhatsApp. Current speech models handle them reasonably well, but accuracy drops with background noise and mixed languages, so for bookings the AI should confirm the key details — date, time, party size — in text.
How will the AI know the kitchen closed early tonight?
Only if someone tells it. Give the shift manager a simple button or command: "kitchen closed", "dish sold out", "terrace closed due to weather".
We have several sites. Can we start with one?
Yes, and it is the better route: one site is where you refine dish cards and escalation rules, and then you roll the setup out to other branches, adjusting for their menus and floor plans.