AI for Hospitality: Which Jobs to Delegate and Which Stay With Staff

Hand over repeatable, low-stakes work: answering routine questions from a written fact sheet, drafting replies, posts and emails for a person to approve, and collecting booking details for your own booking system. Keep complaints, allergen and dietary-safety questions, and refunds with staff. A one-page rule sheet, printed below, makes that split explicit for every shift.

The dividing line is not how hard a job is. It is what a wrong answer costs. An assistant writes the most plausible reply from the text it has been given; it cannot walk into the kitchen, check a fridge or phone a guest. So it can safely do anything whose answer is already written in a document you control, and it should never be the last word on anything a guest could be harmed or short-changed by.

Follow me on Instagram@sagnikteaches

How an assistant answers a guest, and why that sets the limits

When a guest message arrives, the assistant predicts a helpful reply from three things: the message, your instructions and any documents you have attached. It has no live view of the kitchen, the till or the room list unless you connect it to a system that provides them. That is why the same tool is excellent at "what time does breakfast finish?" and dangerous at "is the risotto safe for my son?".

Connect on LinkedInSagnik Bhattacharya

Five misconceptions are worth clearing up before anything else in a small hotel, cafe or restaurant:

Subscribe on YouTube@sagnikteaches

• "It learns my business as it goes." Not by default. It knows what is in its instructions and the files you gave it. Improvements come from you editing the fact sheet.
• "If it says it checked with the kitchen, it did." It cannot. A line such as "I have confirmed with the chef" is invented, so ban that wording in the instructions.
• "A friendly reply is a safe reply." Warm wording slips in promises. "We will sort that for you" can commit you to a refund nobody approved.
• "The facts stay true." Menus, prices, hours and suppliers change. A dated fact sheet and a weekly check keep replies current.
• "Guest details are fine in any chat tool." Keep names, phone numbers and booking references out of consumer chat apps. Consumer plans let users switch off training on their chats in privacy settings, and business plans do not train on your content by default, but the safer habit is to paste nothing identifiable at all.

For a catalogue of the jobs a small hotel or cafe could give an assistant, see how to use AI in the hospitality industry. The harder question is where each job sits once you have the list.

Four questions that sort any job into a lane

Three lanes cover almost everything. Lane A means the assistant replies on its own and a person spot-checks a sample. Lane B means the assistant drafts and a person approves every time. Lane C means staff only, and the assistant's single task is to say that a person will reply, and when. Four questions place a job:

1. Is the answer written in a document you control? The menu, the policy, the rate card, the booking system. If it depends on what the kitchen did today, the answer is no.
2. Could a wrong answer hurt someone or cost real money? Think of an allergic reaction, a double-charged deposit, a wedding party sent to the wrong room.
3. Does the reply promise something? A table, a price, a discount, a refund, a time.
4. Would a regular guest expect a person here? A complaint, an anniversary, a bereavement, a first visit after a bad review.

The scoring is short. A yes to question 2 or 4 means Lane C. Otherwise, a no to question 1 or a yes to question 3 means Lane B. A clean sheet (yes, no, no, no) means Lane A. One exception: when your booking system itself confirms the promise, the assistant can take the details in Lane A and the system makes the promise.

Nine guest messages, sorted

These nine are invented for illustration, but each is the kind of message a small hospitality inbox receives.

Two details are worth noticing. The vegan question is Lane A only while the message contains nothing about allergies; the moment "allergy" or "intolerance" shows up, the whole message moves to Lane C, even if the rest is routine. And the table request never ends with the assistant saying "booked". It says the request has been sent to the booking system, and the confirmation wording comes from the system.

The one-page rule sheet, ready to print

Copy this, fill in the brackets, pin it in the office and give the same text to the assistant as its standing instructions. Keep it to one page; a sheet that runs to three is a sheet that nobody reads.

AI RULE SHEET - [Business name] - reviewed [date] by [name]

LANE A: the assistant replies on its own. We spot-check one reply in five each week.
- Opening times, address, directions, parking, Wi-Fi, check-in and check-out times
- What is on the current menu and its price, from the menu document only
- Whether a table or room is free, only as shown by the booking system
- Taking booking details for the system: name, date, time, party size, phone

LANE B: the assistant drafts, a person approves every time.
- Special requests: late check-out, early arrival, cot, high chair, cake
- Group and event enquiries, quoted from the rate card
- Changes to an existing booking
- Replies to reviews and social comments
- Newsletters, posts, menu descriptions, pre-arrival and thank-you emails
- Translations of guest messages and of our replies
- Rota drafts, supplier order drafts, training notes

LANE C: staff only. The assistant writes only the holding line.
- Complaints of any size
- Any allergy, intolerance or "can't eat" question, however small
- Refunds, compensation, deposits, chargebacks
- Illness, injury, safety, harassment, anything involving a child
- Accessibility needs that depend on the building or the kitchen
- Press, legal or insurance matters

TRIGGER WORDS (any of these moves the whole message to Lane C):
allergy, allergic, intolerance, coeliac, anaphylaxis, epipen, nut, peanut, sesame,
gluten, dairy, shellfish, "can't eat", refund, compensation, deposit, chargeback,
"money back", complaint, disgusting, ill, sick, "food poisoning", injured, unsafe,
solicitor, lawyer, press

CONTACTS (who owns each Lane C reply)
Complaints: [name], reply within [hours]
Allergens: [head chef or duty chef], same day
Refunds and deposits: [manager]
Out of hours: the assistant says a person will reply from [time]

NEVER: say we have checked with anyone; say a booking is confirmed unless
the booking system says so; promise a discount, refund or upgrade.

Notice how few Lane A jobs there are. That is deliberate. The sheet is not trying to automate as much as possible; it is trying to make the boundary obvious to a new member of staff on their first shift.

One warning if a keyword filter applies the trigger list rather than the assistant reading it, as happens in some chatbot builders and automation tools: set it to match whole words. As plain fragments of text, "ill" fires on "will" and "bill", "nut" fires on "minute" and "sick" on "sickly sweet", so nearly every message lands in Lane C and staff stop trusting the lanes. Whole-word matching fixes most of it. The false alarms that remain, such as "nut-brown ale" on a drinks enquiry, are the safe way round: a needless staff reply costs a minute, a missed allergy costs far more.

What the assistant says when a message belongs to staff

The assistant's most important skill is declining politely. Give it a short triage prompt that sits on top of the rule sheet:

You triage guest messages for [business name]. The rule sheet and fact sheet are below.
For every message output three lines:
LANE: A, B or C, with the reason in five words.
REPLY: Lane A, the answer from the fact sheet only. Lane B, a draft starting "DRAFT FOR APPROVAL:".
Lane C, only the holding line: thank them, say a named person will reply by [time], give no advice.
NOTE FOR STAFF: one line with what the guest wants, any booking details and any deadline.
If a message mixes lanes, use the highest one (C over B over A).

Run the allergy message from the table through it (illustrative output):

LANE: C (allergy mentioned, safety)
REPLY: Thank you for telling us about your daughter's allergy. A member of our kitchen team will reply personally by [time] today so that you get accurate information. Please do not rely on the menu until then.
NOTE FOR STAFF: Guest asks whether the lamb tagine is nut-free. Her daughter has an allergy. Booking Saturday 7pm, party of four.

That reply does three things right. It acknowledges the allergy, gives a specific time, and says nothing about which dishes are fine. A weaker version would add "our chefs take allergies very seriously", which sounds reassuring and answers nothing. If you see sentences like that in your own tests, add a line to the instructions: "In Lane C, write nothing about the food, the kitchen or safety."

The holding line needs a real time and a real person behind it. "We will be in touch soon" breaks the moment a guest waits two days. If nobody is on shift until 9am, the line should say so: "Our team will reply after 9am."

Lane B in practice: a draft that promises too much

Lane B is where most of the daily value sits, and where the quiet errors live. Take the 2pm check-out request from the table. The first draft an assistant produced (illustrative):

DRAFT FOR APPROVAL: Hi [first name], of course, 2pm check-out is no problem! We'll have everything ready for you. See you on Sunday.

It sounds right and it is wrong twice. The written policy says check-out is 11am, with late check-out until 1pm at $[amount] when no arrival is booked into that room, so "no problem" promises something the policy does not. And "we'll have everything ready" implies a service nobody agreed to. The approver's job is a 60-second check: open the room list, see whether the room has an arrival on Sunday, and rewrite the promise to match what is true.

Hi [first name], thanks for asking. Late check-out until 1pm is possible when the room is not needed that afternoon, at $[amount]. The room list shows [it is free: I can confirm 1pm / it is needed: the latest I can offer is 11am, and we can store your bags until you leave]. Let me know and I will note it on your booking.

Teach approvers three checks, in this order. Does every number and policy in the draft come from the fact sheet? Is every promise backed by a system or a person who has looked? Does the tone sound like your place, not like a brochure? Most bad drafts fail the second.

Allergens: the line that never moves

An allergen answer needs three things an assistant lacks: today's recipe, today's supplier, and the cross-contact risk in your kitchen. A confident, tidy and wrong reply is the worst outcome, because it reads as reliable. Food businesses carry legal duties around allergen information in many places, so ask the food-safety inspector who covers you, or an adviser, what applies to you, and write the answer process into the sheet.

The rule is blunt: any allergy or intolerance mention goes to the chef on duty, who answers from the kitchen's current allergen record and says so to the guest. The assistant may do two small things around it. If your printed menu lists allergens dish by dish, it can say so, and add that a member of the kitchen team will confirm before the guest orders. And it can help you format the allergen matrix, the table of dishes against the major allergens, from the chef's own source. The chef signs it off, and the date goes on it. For the longer discussion, should an AI chatbot answer allergen questions goes through the failure modes.

Do not let dietary preference and allergy blur. "Is the curry vegan?" can be answered from a menu document. "Is the curry safe for my dairy allergy?" cannot. Test your setup with both phrased as a single message and make sure the second wins.

Complaints and refunds: what the AI may prepare for staff

A complaint is a relationship problem with a deadline attached. The guest wants to see that a person has read it. What an assistant can do, behind the scenes, is take the first few minutes of admin off the manager: summarise the complaint, pull the booking details, and lay out the options the policy allows. The manager chooses and writes, or approves, the reply.

CASE NOTE (drafted by the assistant, for the manager)
Guest message, in one line: Room 6, stained carpet, no apology at check-in.
Booking: 2 nights, rate $[amount], paid in full, anniversary request noted.
What the policy says: [paste relevant lines]
Options inside policy: apology and a drink on arrival (already past); partial refund up to $[amount]; a voucher for a future stay.
Not decided by the assistant: which option, and the wording of the reply.
Reply due by: [time].

Refunds are the second half of the same rule. An assistant can explain your refund policy in plain words, copied from the written policy, but it must never say that a refund will be paid. The payment may involve a card processor, a booking site's own rules or a chargeback already under way, and only the person who can see all of that should promise anything. Scripts for the conversation itself are in the tutorial on answering restaurant complaints with AI without escalating. For a hotel, the service recovery scripts tutorial at the end of this page covers the same ground.

Bookings: the AI takes details, the system makes the promise

Bookings are the best job to delegate and the easiest to get slightly wrong. The assistant collects five fields: name, date, time, party size and a phone number or email. It passes them to your booking system, and the system answers yes, no or "need a person". Only the system's answer becomes a booking. Until then, the assistant uses wording such as "I have sent your request for a table for six on Saturday at 7pm; you will get a confirmation message when it is in the book."

The staff-only exceptions are predictable: parties above a size you set, anything needing a deposit, any special requirement beyond a high chair, and any date that sits near a local event or a private hire. The assistant flags those with "BOOKING NEEDS A PERSON" and the reason, and puts them in Lane B or C.

If you want an assistant to answer the phone as well, prices vary widely. As of October 2026, Smith.ai lists its AI receptionist as free for 25 calls a month, then $3 a call, with a Pro plan from $150 a month, and Yelp Host starts at $249 a month after a trial of up to 30 days. Check both pages before you rely on those figures. Whichever you choose, the rule sheet comes first: the vendor's setup screens will ask what to do with complaints, allergies and refunds, and you want the answers written already. The tutorial on restaurant chatbots for hours, bookings and hand-offs covers the technical side. If you use WhatsApp for guest messages, its business messaging policy asks for a clear route to a person when you automate, which your Lane C holding line provides.

Sorting a real week: 120 messages in a 12-room hotel

The quickest way to see your own split is to sort a week of real messages. Print or export last week's guest messages and mark each A, B or C using the four questions. It takes about half an hour. Take a 12-room hotel with a small restaurant. In the illustrative week, 120 guest messages arrive by email, web form, social messages and phone notes:

That is 249 minutes, about four hours, back in a week. Lane A accounts for roughly 180 of those minutes and Lane B for roughly 80, less 10 for fact-sheet upkeep. Lane A also answers at 11pm when nobody is on shift, which the minutes do not capture. Lane C saves nothing, and that is correct: the messages that most affect reputation and safety are exactly the ones handled by a person.

The split changes with the business. A cafe's inbox is mostly Lane A: opening hours, whether dogs are welcome, whether there are high chairs, whether it takes bookings for eight. A restaurant has more Lane C from allergy questions and deposit disputes. A hotel has a long tail of Lane B special requests. If your own sort puts more than a third of messages in Lane C, look at what is producing so many complaints, allergy questions or refund requests before you look at any tool.

A first week to set up, then a two-week audit of each lane

A sensible first week runs in this order:

1. Day 1. Sort last week's messages into lanes, as above, and note your own split.
2. Day 2. Write the fact sheet: hours, menu, prices, policies, directions, with a date on it.
3. Day 3. Fill in the rule sheet and the contacts, and give the assistant both.
4. Day 4. Write ten mixed messages and see which lane wins.
5. Days 5 to 7. Run everything as drafts only, with a person reading each one before it goes.

Mixed messages are the real test, because real guests do not write one topic at a time. "Booking for six on Friday, and one of us is coeliac" is a booking with an allergy inside it, so it belongs in Lane C. "Lovely stay, but breakfast was cold, and can we book again for May?" is a complaint and a rebooking, so the complaint decides the lane and the rebooking waits for the person. "Do you do gluten-free cake for my dad's birthday?" sits on the edge between preference and allergy, so it goes to the chef until they say otherwise. If any of those slips into Lane A, add the missing words to the trigger list and run the ten again.

After two weeks, audit. Pull 20 messages from each lane and look for one specific failure each:

• Lane A: wrong facts. The standard is zero. One wrong fact means correcting the fact sheet that day and re-reading the last week's replies on that topic.
• Lane B: heavy rewrites. Count the drafts a person had to rewrite rather than tweak. More than 5 in 20 means the prompt or the fact sheet needs work, not the approver.
• Lane C: leaks. Count messages that should have been Lane C where the assistant said anything beyond the holding line. The target is zero, and each leak points to a missing trigger word.

Then tell guests what they are talking to. A first line such as "This is our automated assistant. A member of the team reads every request about allergies, complaints and refunds" costs nothing and sets expectations. If you have customers in the EU, the AI Act's chatbot transparency rules have applied since August 2026, and the duty sits mainly with the provider of the chatbot, so check what your supplier already shows.

Finally, tell your staff, in plain terms, which jobs are leaving their desk and which are not. People who fear that a chatbot is replacing them stop reporting its mistakes. The tutorial on training front-of-house staff to work alongside AI suggests how to run that conversation. The rule sheet helps: it shows that the jobs needing judgement and warmth were never on the list.

Questions owners ask when drawing the line

Can AI answer guests' dietary questions at all?

Preferences, yes, from a written menu document: vegetarian, vegan, halal or dairy-free dishes as you have labelled them. Anything with the words allergy, intolerance, coeliac or 'can't eat' goes straight to staff, who answer from the kitchen's current record. The assistant's only job there is to say that someone will reply, and when.

Should I tell guests a reply was written by AI?

If a bot or auto-reply answers on its own, say so in its first line, because guests who ask a bot a question deserve to know. A Lane B draft that a member of staff has read, edited and sent is their reply, so there is nothing to label, but they must stand behind every fact and promise in it. When in doubt about which a message is, label it.

How often should the rule sheet be reviewed?

Review it after the first two weeks, then every quarter, and any day your menu, policies, suppliers or booking system change. Date the fact sheet and the rule sheet, and name one person who owns them. A sheet nobody owns drifts out of date, and a stale fact sheet is how an assistant ends up quoting last season's opening hours.

Who in the team should approve Lane B drafts?

The person who would have written the reply anyway: the front-desk lead for guest requests, the manager for reviews, the chef for menu copy. Approval is fastest when the same person does it each time and has the fact sheet open. Do not route every draft to the owner, or the saving disappears.

Further reads

• AI Voice Agent for a Hotel Front Desk: Calls to Hand Over and Keep — Phone calls to hand over and calls to keep.
• Hotel Guest Complaints Handling: Service Recovery Scripts With AI — Service recovery scripts for the Lane C complaints.
• AI Chatbot Online for Hospitality: Questions a Café Can Safely Automate — Cafe questions a chatbot can safely take.
• What AI Can and Cannot Do for an Independent Café — What AI can and cannot do for an independent cafe.
• How to Reply to Hotel Reviews With AI Without Sounding Canned — Reply to hotel reviews without sounding canned.
• Can AI Answer the Phone and Take Restaurant Reservations? — Whether AI can answer the phone and take reservations.
• AI for Hotel Revenue Management: Weekly Reports for Small Hotels — A twenty-minute Monday routine: a PMS export, a sheet for occupancy, ADR, RevPAR and pace, and an AI summary that ends in questions, not rate orders.
• ChatGPT Prompts for Coffee Shop Owners: 10 for Menus, Posts, Rotas — Ten ChatGPT prompts for a coffee shop, from cost per cup to rotas and posts, each with example inputs and a check before you use it.
• Hotel SOPs Written With AI: Check-In, Housekeeping, Night Audit — An interview-style prompt turns what your receptionists, room attendants and night auditor already do into signed-off SOPs with version control.
• How to Use ChatGPT for Restaurants: Five Jobs to Hand It Each Week — A Monday-to-Sunday routine: ordering, prep and specials, rota, weekend promotion and the weekly review, with allergen answers kept human-checked.
• AI Agents for Hotels: What They Can Do Today and Where They Fail — What a hotel AI agent can handle now, where it fails on bookings, guest messages and housekeeping, and the tests and hand-off rules to set first.
• AI for Coworking Spaces: Enquiries, Tours and Room Bookings — A job-by-job sort for a coworking space: enquiry replies, tour booking, member FAQs and room requests AI can take, and the jobs that stay with staff.
• Can AI Draft a Restaurant Staff Handbook? What an Adviser Must Check — Sort each handbook section into AI-safe, owner-facts or adviser-only, then see the tips, leave, shift and version-control checks to finish before issuing.
• ChatGPT for Hotels: What It Can and Cannot Do at the Front Desk — What a hotel front desk can hand to ChatGPT, what stays with staff, the data-protection rule, a translation check and a worked Friday shift.
• How Can AI Help Hotels? Housekeeping Checks and Maintenance Requests — Room-turn checklists, inspection patterns, fault tickets and rota drafts for a small hotel: what AI produces, how to check it and what stays with a person.
• Restaurant Tip Pooling: Can AI Draft a Fair Policy Under the Rules? — AI can draft the wording of a tip pooling policy, but your rules decide who may share. A worked split, the clauses to include and the checks.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: WhatsApp Business Messaging Policy (escalation paths); Smith.ai and Yelp Host pricing pages (October 2026 list prices); EU AI Act Article 50. All message counts and minutes are illustrative.