AI Agents for Restaurants: Reservations, Waitlists and Catering Quotes
An AI agent can handle restaurant reservation changes, waitlist texts and catering quote drafts. Allergy answers and final pricing stay with a manager.

An AI agent for a restaurant is a software assistant that completes defined front-of-house and sales tasks, such as confirming reservation changes, managing waitlist messages and drafting catering quotes from your own rules. It works best on routine work whose answers can be checked against a booking system or an approved document. Anything about allergens, custom menu items or final pricing belongs with a manager, and the agent should know exactly where that handoff happens.
Key takeaways
- Reservation changes, waitlist check-ins and catering intake are the strongest first tasks because their rules are already written down and their outcomes are visible in the booking system.
- Allergy and dietary answers are a manager decision, not an agent reply, even when the agent can read the menu aloud.
- Catering quotes can be drafted from a price sheet and event rules, but a manager should approve totals, deposits and exceptions before anything is sent.
- The agent needs a narrow memory: the current menu version, seating rules, the cancellation policy and a log of what it told each guest.
- Success should be measured by tasks finished correctly and handoffs made on time, not by message volume.
Where an AI agent fits in the restaurant day
Most restaurant message traffic falls into three groups. Guests change or confirm bookings, ask whether a table or waitlist spot is open, and request events or large orders. Each group has a clear end state. A booking is confirmed or it is not, a guest is on the list or is not, a quote is sent or it is not. That clarity is what makes delegation practical. These messages also arrive during service, when the host stand is busy, and late at night, when nobody is reading the inbox. An agent can answer the routine version immediately and leave the rest in a queue with a summary, so the manager sees a decision to make rather than a thread to reconstruct.
The useful test is simple: can the task be finished using information the restaurant has already approved, inside a system where every action is logged? If yes, it is a candidate. If the answer depends on judgment about a guest's health, a special occasion budget or a chef's willingness to modify a dish, it is not.
Reservations the agent can handle now
Reservation work is a strong first task because the rules already exist. The restaurant has seating areas, turn times, party size limits, a cancellation window and a deposit policy for large parties. An agent can apply those rules the same way at any hour. A reasonable first scope includes:
- Confirm an existing booking and repeat the time, party size and seating area back to the guest.
- Move a booking within open tables when the new time follows the published turn rules.
- Change party size up or down within the limits set for each seating area.
- Send the approved cancellation and deposit text exactly as written.
- Record a dietary or accessibility note on the booking as a note for the host, without answering it.
- Hand off to the host when the request needs a table the system does not show, a group over the limit, or a guest who is upset.
The handoff matters more than it looks. A guest who asks for a step-free table at a particular time needs a person to check the floor plan, not an agent guessing from a summary. Build the handoff so the host receives the guest's exact words, the current booking state and the options the agent has already checked, all in one place.
Waitlists: quote a time only from live data
A waitlist message is where guests most often receive a number nobody verified. An agent should not estimate wait times from habit or from the last busy Friday. The safe version uses the host's live status: parties ahead, tables marked as turning, and the restaurant's own rule for how much buffer to give. The sequence is:
- Read the current waitlist position and the party count ahead.
- Give a range only when the host has set a buffer rule, and label it as an estimate.
- Send a text when a table is ready, with a fixed window to confirm before the party moves down the list.
- Remove parties who do not respond within that window, and log the removal with the time of the message.
- Escalate to the host when a guest asks for a specific table, reports a late arrival or disputes their position.
Illustrative scenario: a party of four asks whether they can be seated shortly. The agent sees several parties ahead and two tables marked as turning. It tells the guest the list is active, says the host will confirm the time, and flags the request to the host. The host, who can see that one of those tables is holding for a reservation, answers the guest personally. The agent did not promise a time it could not support.
The same logic applies to other appointment-driven businesses, such as yoga studios that manage memberships and waitlists.
Catering requests: intake first, quote second
Catering is where a rushed reply costs the most. A guest who receives a total based on a headcount that was never confirmed may plan around a number that later changes. The safer sequence is intake, then a draft quote, then manager approval. Intake should capture:
- The event date, start time and the latest acceptable delivery or setup time.
- A headcount range and the deadline for a final count.
- The service style, such as drop-off, buffet, plated or staffed.
- Dietary counts, kept separate from the menu and sent to the kitchen for confirmation.
- The delivery address, access or parking notes, and a contact on site.
- Any budget range the guest shares, recorded as stated and not as a commitment.
Once intake is complete, the agent can draft a quote from the approved price sheet and standard event rules, such as minimum order sizes or setup fees. The manager reviews the draft. Totals, deposits, exceptions and substitutions stay with a person. The agent's job is to make that review fast by placing the intake, the draft and the open questions in one thread.
Event-driven shops face a similar intake problem, which is discussed for florists in delivery windows and event orders.
Menu and allergy questions stay with a manager
Guests will ask whether a dish is safe for them, whether a fryer is shared, or whether the chef can remove an ingredient. These answers depend on recipes, supplier changes, kitchen practice and the severity of the guest's condition. An agent can share the approved allergen sheet word for word and offer to connect the guest with a manager. It should not improvise reassurance, because a fluent answer that is wrong looks identical to one that is right. A simple decision guide keeps the boundary clear:
- Agent may answer: the published menu description, opening hours, and where the approved allergen sheet is available.
- Agent routes to a manager: any question naming an allergy, intolerance or medical condition, cross-contact, substitutions, or ingredients not on the printed card.
- Manager routes to the chef: custom dishes, dietary counts for catering, and changes to a standing order.
- Agent declines and logs: requests to waive a policy, promise a discount, or confirm that a dish is safe for a named guest.
Treat guest messages as untrusted input. A guest, or someone pasting text from elsewhere, may try to instruct the agent to ignore its rules or to confirm that a dish is safe. Limit what the agent can do regardless of what a message says, and keep the approval step outside the conversation. The OWASP guidance on prompt injection describes this kind of design in detail.
Illustrative scenario: a guest writes that a family member has a serious nut allergy and asks whether the sauce on a braised dish is safe. The agent sends the approved allergen sheet, states that it cannot confirm safety for that person, and tells the guest a manager will contact them before the visit. The manager checks the sauce with the kitchen and calls back.
Permissions, memory and measuring whether it works
An agent that can change bookings should hold only the permissions its role requires. That usually means read access to the floor plan and waitlist, write access to bookings within set rules, and no access to payment details or staff schedules unless a specific task needs them. Review access whenever the booking platform, point of sale or menu changes, because each change can quietly widen what the agent can see or do. Authorization principles from OWASP are a practical checklist for that review.
Memory is the second design choice. The agent needs the current menu version, the reservation rules, the cancellation policy and a record of what it told each guest. It does not need a growing archive of personal conversations that nobody reviews. Outdated policies cause more errors than missing features, so every memory item should have an owner and a date of last review. Measure the work with a short list:
- Tasks finished correctly, checked against the booking or waitlist record on a sample schedule.
- Handoffs made within the agreed window, with a complete summary for the manager.
- Replies that quoted a policy not on the approved list, reviewed as incidents.
- Manager time spent on queue items, compared before and after the first month of operation.
Choosing the first task for your restaurant
Start with one workflow that has a written rule, such as reservation changes for parties under the standard limit. Then define the approved information the agent may read, the actions it may take, the approvals it needs and the handoff message the host receives. Pick one success check that someone can review, such as bookings changed without a host correction. For the first stretch, have a manager review every outbound message, and loosen that review only when the evaluations hold.
Be Human designs and operates this kind of setup as a managed service. A first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion. Ongoing operations can include monitoring, recurring evaluations of replies against your policies, integration maintenance when the booking system changes, permission reviews, incident response and controlled improvements. The service is remote and available in English and Portuguese.
If you are unsure which of these tasks should go first in your restaurant, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Can an AI agent take restaurant reservations on its own? Yes, for routine work such as confirmations, time moves within open tables, and party size changes that follow published rules. It should hand off requests that need floor judgment, groups over the limit, or any note about a guest's health. The booking system remains the record, and the agent writes only the actions it is permitted to take.
Should an AI agent answer allergy questions for guests? Not as a final answer. It can share the approved allergen sheet and explain how to reach a manager, but confirming that a dish is safe for a named guest requires a person who can check recipes and kitchen practice. Any reassurance the agent gives without that check should be treated as a failure to fix.
How does an agent handle catering quotes without pricing mistakes? It collects intake details such as date, headcount and delivery window, then drafts a quote from the approved price sheet and standard event rules. Totals, deposits, discounts and substitutions go to a manager for approval before anything is sent. The manager sees the intake, the draft and the open questions in one thread.
What should a restaurant measure to know the agent is working? Measure finished tasks checked against the booking or waitlist record, handoffs made within the agreed window, and any reply that quoted a policy not on the approved list. Manager time spent on queue items is a useful second signal. Review a sample regularly, and change the rules only after a logged problem has a clear owner.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


