ChatGPT can draft guest replies, translate messages, tidy directions and turn shift notes into handovers, provided a receptionist reads everything before it goes out. Payments, ID checks, complaints, refunds and safety incidents stay with staff, and guest names, card numbers and ID details are never typed in. Anything that promises something is a person's decision.
A front desk runs on two things a language model doesn't have: facts only the property holds (tonight's availability, the rate, the policy) and the authority to make promises (a late check-out, a refund, a room move). ChatGPT writes in a warm, confident voice, so a draft can commit the hotel to something nobody approved. It also has no way to confirm who a guest is. The workable set-up is a receptionist who drafts with it in a separate tab, with the guest's details left out, and who stays the author of every message.
The desk, moment by moment
Most of a front desk's written work falls into ten moments. The table shows where ChatGPT can take the first draft and what a person still does. The right-hand column is never optional.
| Moment | ChatGPT's part | Staff's part |
|---|---|---|
| Pre-arrival message: directions, check-in time, parking | Drafts from your fact sheet and tidies the tone | Checks times, rates and policies against the booking system, then sends |
| Guest writes in another language | Translates the message and drafts a reply in that language | Reads a back-translation; has a bilingual colleague check anything about money, allergies or safety |
| Early check-in or late arrival request | Drafts a polite holding reply | Checks room status and night cover, then makes or refuses the promise |
| Directions and suggestions for places to eat | Writes from a list you maintain | Confirms opening hours, which change |
| Problem during the stay | Drafts apology wording once staff have chosen the fix | Decides the remedy and any compensation |
| Question about a charge on the bill | Explains how bills are built, in plain words | Finds the cause in the property system and decides on any correction |
| Lost property after check-out | Drafts a reply asking what the item is and where it was left | Checks the lost-property log, confirms the person asking is the guest before releasing anything, and arranges collection or postage |
| Group or event enquiry | Drafts a reply asking for dates, numbers and needs | Quotes rates and holds rooms |
| Review reply | Writes a first version | The manager edits and posts |
| Shift handover | Turns an anonymised log into a one-page note | Adds what must be said aloud, such as a guest who needs a call |
For ready-made prompts for the guest emails in the first rows, ChatGPT prompts for hotels: guest emails and rate copy builds each one on a single fact sheet. The handover row is covered in hotel shift handover notes with ChatGPT.
Take the bill row. A guest emails the morning after check-out: "There's a charge of $38 from you on my card that I don't recognise. I want it refunded today." With no standing rules, a typical draft says: "We're so sorry for the confusion! We've refunded the $38, and you should see it back on your card within 3 to 5 working days." That is a refund nobody authorised, a cause nobody looked up and a bank timescale the model made up. Under the rules block further down, the draft is a holding reply:
Hi [GUEST], thanks for letting us know, and sorry for the worry. I'm checking the $38 charge against your bill now and will reply by [time] with what it was for. If it was charged in error, [manager] will arrange the correction and confirm it to you in writing. [receptionist]
It repeats the amount the guest gave, says what is being checked and puts a time on the answer. The receptionist then opens the guest's folio (the running bill) in the property system. In this illustration the charge is a minibar item posted to the wrong room, so the manager approves the refund and the receptionist writes the confirmation personally, quoting the correction as it appears in the system.
What stays with staff, and why
The red list is short, and each entry has a reason that makes it easy to remember.
- Payments and card details. ChatGPT can't take or refund a payment, and card numbers, expiry dates and security codes never go into a chat. Your payment provider's agreement and card-industry security rules say how card data may be handled, and a chatbot isn't part of that arrangement.
- ID and registration checks. Checking a passport or driving licence is a person looking at a document and at the person holding it. Never upload an ID image to ChatGPT. Where guest registration is required by law, it remains a staff process.
- Complaints and compensation. A complaint needs a decision about fault and remedy. ChatGPT can prepare the wording of an apology after staff decide, but it can't decide, and a draft that offers a free night has made an offer. The scripts in hotel guest complaints and service recovery with AI show the order: acknowledge, fix, then compensate within limits staff can quote.
- Refunds, discounts and rate changes. Authority sits with named staff and the booking system, neither of which the model can see.
- Safety, medical and security incidents. Emergency services first, then the manager, then the incident log. Any wording sent to a guest or an insurer afterwards belongs to the manager.
- Allergies and dietary promises. Never from ChatGPT. The kitchen's written allergen information is the only source, and a person passes it on.
- Anything involving a child, a vulnerable guest or a dispute between guests. These need judgement and records, not a drafted paragraph.
- Availability and rates. The property system is the source. A model asked "do you have a room on Friday?" will say something fluent, and none of it will come from your calendar.
The broader sorting of hospitality jobs into delegate, draft and keep is in AI for hospitality: which jobs to delegate and which stay with staff. The desk adds limits that sorting doesn't cover: proving who a guest is, handling card data and keeping the registration record.
The data-protection rule: what never gets typed in
Guest messages carry personal data: names, contact details, dates of stay and often more. Data-protection law such as the GDPR treats some details, including health and religious belief, as especially sensitive, and a dietary or mobility request can reveal both. So the rule is simple: guest details are not typed into ChatGPT. The table shows the swaps.
| What the message contains | Never paste | Paste this instead |
|---|---|---|
| Name | The guest's name | [GUEST] |
| Contact details | Phone, email, address | Leave out |
| Booking and room | The real reference and room number | [BOOKING], [ROOM] |
| Dates | Exact dates alongside a name | "Friday arrival, two nights" |
| Card details | Any digits | Nothing; there is no safe version |
| ID | Passport or licence numbers, scans, photos | Nothing; there is no safe version |
| Health, allergy, mobility or religious needs | The detail | "a dietary requirement", "a mobility need" |
| Complaints that name people | Other guests' or staff names | [GUEST B], [STAFF MEMBER] |
| Children | Names, ages, school details | Leave out |
A ten-second test before pressing enter: if someone found this chat, could they identify a guest or learn something sensitive about them? If yes, it isn't ready.
Settings are a backstop, not a licence. On personal plans, the switch labelled "Improve the model for everyone" under Settings, then Data controls, stops chats being used to improve OpenAI's models, but it doesn't delete chats already saved. Temporary Chat isn't used for training either and may be kept for up to 30 days for safety, according to OpenAI's data controls help page. ChatGPT Business doesn't train on business content by default. None of this makes pasting identifiable guest data a good idea, because the text is still stored by a third party. Redaction is the main control. Whether your use needs a line in your privacy notice or a processing agreement is a question for your data-protection adviser, and none of this is legal advice.
Translation: useful, with a check built in
Guests write in languages the desk doesn't speak, and ChatGPT translates fluently in the widely used ones. The risk is a fluent sentence that says something slightly different from what you meant. The routine that catches it takes two minutes:
- Translate the guest's message into English, and read it before replying.
- Write your reply in English, from your fact sheet, and have ChatGPT translate it into the guest's language.
- Open a new chat, paste only the translated reply, and ask for a literal translation back into English.
- Compare the back-translation with your English original. Differences in money words, times, allergies and safety are the ones that matter.
- Keep the English original in the booking record, so a colleague can see what you meant.
The data rule applies to step 1. Strip the guest's name, booking reference and any health or religious detail before pasting. If the sentence you most need translated is the allergy or mobility line, give that sentence to a bilingual colleague or to a translation service covered by your own data agreement, and keep it out of a personal ChatGPT chat.
Here is the kind of slip it catches. Your English says: "We place a temporary hold of one night's rate on your card; it is released at check-out." Two illustrative translations into the guest's language, each with a literal back-translation:
| Version | Translated reply | Back-translation |
|---|---|---|
| Faithful | Retenemos temporalmente el importe de una noche en su tarjeta; se libera al terminar su estancia. | We temporarily hold the amount of one night on your card; it is released when your stay ends. |
| Slipped | Hacemos un cargo temporal del importe de una noche en su tarjeta; se devuelve al terminar su estancia. | We make a temporary charge of one night's amount to your card; it is returned when your stay ends. |
The second version reads perfectly well in the guest's language, and nothing in it looks wrong unless you can read it. The back-translation shows "charge" and "returned" where the original said "hold" and "released". A guest who reads that may call their bank. The words to compare every time are deposit, hold, charge, refund, fee, cancellation, and any clock time, particularly where the guest's habit is a 12-hour or a 24-hour clock.
Quality drops in less widely used languages, so test with a native speaker before relying on it. Never translate allergen statements, safety instructions or legal terms without a human who reads the language.
Three properties, three desks
The same rule fits every property, but the set-up differs.
A nine-room guesthouse with one person on the desk. The owner answers messages in the evening. A free or Plus account in a browser profile used only for drafting is enough, and a Project holding the fact sheet saves retyping policies. The risk is habit: pasting a guest's email "just this once". A sticky note by the monitor is the control, and it works better than a policy document nobody opens.
A 40-room city hotel with three shifts, group bookings and several languages. Several people use it, so each needs a login, which points to a ChatGPT Business workspace: individual seats, centralised administration, and no training on business content by default. The translation check needs a bilingual colleague for money and safety wording, and group enquiries are drafted from a rate sheet you control, never from the model's idea of a group rate.
A 14-room boutique hotel with no night staff. Guests message through WhatsApp and booking platforms overnight, and the temptation is an auto-reply bot. A drafting tool doesn't fix overnight cover; it helps staff clear the morning queue faster. Anything that answers by itself is a chatbot decision with its own rules, covered in WhatsApp for small hotels: Meta's AI rules, templates and handover.
When does it pay? As a rough guide, and it's my judgement rather than a measured figure: if staff spend more than an hour a day composing written replies, drafting help repays the checking time. Below that, saved replies inside the booking system are cheaper and safer.
A Friday evening shift at a 30-room hotel
Here is one busy shift, invented but realistic, sorted by what happened and who did what. The minutes saved are the drafting time ChatGPT took off the receptionist, before checking.
| Time | Moment | ChatGPT's part | Minutes saved |
|---|---|---|---|
| 17:05 | Early check-in request by email | Holding reply; staff check housekeeping | 3 |
| 17:40 | Directions for a guest arriving by car | Drafted from the fact sheet | 2 |
| 18:10 | Late arrival message in another language | Translation, draft reply, back-translation | 6 |
| 19:00 | Room too cold | None; staff fix it | 0 |
| 19:30 | Question about a charge on the bill | Polished the explanation after staff found the cause | 3 |
| 20:15 | Guest reports noise from the room above | Apology wording after staff chose the remedy | 4 |
| 21:00 | Card declined at check-in | None; money | 0 |
| 21:30 | Enquiry for 12 rooms in June | Holding reply asking for dates and numbers | 5 |
| 22:00 | Passport copy needed for registration | None; ID | 0 |
| 22:30 | Guest with a nut allergy asks about breakfast | None; staff answer from the kitchen's written sheet | 0 |
| 23:00 | Shift handover | One-page note from an anonymised log | 8 |
ChatGPT touched 7 of the 11 moments and decided none of them. The four it never touched were the cold room, the declined card, the passport copy and the allergy question. Drafting saved 31 minutes. Reading and editing each of the seven drafts took about a minute, so 7 minutes, and the net saving is about 24 minutes.
A quick sum with assumed figures: five shifts like this a week is 2 hours, which at an assumed $18 an hour for desk time is $36 a week. Two ChatGPT Business seats on annual billing cost $40 a month. The comparison says money isn't the barrier. The checking is, and on a quiet shift the saving shrinks while the checking doesn't, which is why the rule of thumb above matters.
A Project that keeps the rules in place on a busy shift
The rule fails when it depends on memory at 22:30 on a Friday. Build it into the tool. Use a ChatGPT Project, the practical home for shared instructions now that custom GPTs are being retired. It keeps standing instructions and reference files together, and sharing is available on every plan. Put this in the Project instructions and the fact sheet in its files:
You help front-desk staff at a small hotel draft replies. A receptionist
reads and sends everything.
Use only the facts in the fact sheet. If a fact is missing, write
[CHECK: what is missing].
Never promise a room, a rate, a refund, compensation, a late check-out
or an upgrade. Never state allergen or medical information.
Guests are written as [GUEST], bookings as [BOOKING], rooms as [ROOM].
If a message mentions payment, ID, injury, safety, an allergy or a
complaint about staff, say it needs a manager and draft only a short
holding reply.
Keep replies under 120 words, warm and plain.The fact sheet is the other half. If a fact isn't on it, the model can't state it from your hotel, so the sheet decides what drafts can say. Ten lines cover most desk replies: check-in and check-out times, breakfast hours, parking, pets, the cancellation policy, how the card hold works, quiet hours, luggage storage, the step-free access you can confirm, and the after-hours contact. Write each as one plain sentence:
HOTEL FACT SHEET (checked by the manager on [date])
Check-in: from [time]. Check-out: by [time].
Breakfast: [hours], [place]. Dietary information: ask staff; the kitchen
holds a written allergen sheet.
Parking: [number] spaces, [free or paid], first come first served.
Pets: [policy].
Cancellation: [policy exactly as on the booking confirmation].
Card hold on arrival: one night's rate, released at check-out.
Quiet hours: [times].
Luggage storage: available before check-in and after check-out.
Step-free access: [what you can confirm; "ask us" for anything else].
After-hours contact: [method and hours].The date matters. A sheet marked "checked by the manager on [date]" tells staff when to re-check it, and a policy change means the sheet changes first, before any reply uses the new rule. Facts that are rarely needed but costly to get wrong, such as the exact cancellation terms, are pasted word for word from the booking confirmation instead of paraphrased, because a paraphrase is where invention starts.
Then the surrounding habits:
- Each person has their own login. Sharing one breaks OpenAI's terms, which bar passing account credentials around, and a shared desk login also hides who pasted what.
- A card by the monitor reads "No names. No card numbers. No ID. No allergies."
- New staff get a 20-minute briefing using five of the moments in the table above, including one that must be refused.
- Once a week, a manager reads five drafts for promises that slipped through, and adds a line to the instructions for each one found.
- If something identifying is pasted by mistake: tell the manager, delete the chat, note what and when, and ask your data-protection adviser whether anything further is needed. What to do in the first day after staff paste client data into ChatGPT walks through it.
Three questions to ask before sending
When a message isn't on any list, these three catch most of what matters:
- Does it promise anything? A room, a rate, a time, a refund, a favour. If yes, the person making the promise reads and owns the sentence, from the system and not from the draft.
- Does it touch money, identity, health or safety? If yes, staff write it, or write the key facts into it from the booking system, the document in front of them or the kitchen's written information.
- Does it say anything the guest hasn't told us or we haven't confirmed? If yes, delete it. Models fill gaps with pleasant detail, and pleasant detail is how a hotel ends up promising a sea view it doesn't have.
Everything else, the greeting, the directions from your own sheet, the rewrite of a stiff sentence, can be handed over with a clear conscience, and a receptionist who has stopped composing the same twelve emails every week has more time for the guests at the desk.
Questions hotel owners ask about ChatGPT at the desk
Can ChatGPT answer guests directly on our website or WhatsApp?
It can sit behind a chatbot, but that is a separate decision with its own rules. The bot should say it is automated (for guests in the EU, the AI Act's chatbot disclosure duty has applied since August 2026), stay away from payments and complaints, and hand over to a person fast. Check your provider's and Meta's rules before connecting WhatsApp.
Is it OK to paste a guest's booking email into ChatGPT to summarise?
Not as it arrives. A booking email holds a name, contact details, dates and often health or religious details. Swap those for placeholders such as [GUEST], [BOOKING] and [ROOM] first, or summarise inside your own booking system. Turning off model training doesn't make pasting identifiable details safe, because the text is still stored.
Which ChatGPT plan suits a front desk?
Not a shared personal login: OpenAI's terms say you may not share account credentials, and a desk computer is shared by shifts. ChatGPT Business gives each person a seat, lists at $20 a user a month on annual billing ($25 monthly), needs at least two seats and doesn't train on business content by default.
Can staff use ChatGPT on their own phones at the desk?
Only for wording that contains no guest details. On a personal account the hotel can't see what was pasted, can't check the training setting and loses the chats when the person leaves. If drafting is part of the job, give staff seats the hotel controls, keep the fact sheet in a shared Project, and treat a guest detail typed into a personal phone as a data slip to report.
Can ChatGPT translate our policies for the website?
Yes, as a first draft. Have a fluent reader check cancellation, deposit, allergen and safety wording before publication, and keep the English text as the master so changes flow one way. Machine translation of money and safety terms is where small slips become disputes.
Further reads
- How Small Hotels Use AI to Answer Guest Messages Around the Clock — Answering guest messages around the clock without losing the personal touch.
- AI Voice Agent for a Hotel Front Desk: Calls to Hand Over and Keep — Which phone calls a voice agent should hand over to staff.
- Hotel Front Desk Scripts: Can AI Role-Play Difficult Guests? — Role-play difficult guests to train desk staff.
- How to Reply to Hotel Reviews With AI Without Sounding Canned — Review replies that don't sound canned.
- Hostel Software: Can AI Handle Guest Messages in Several Languages? — Handling guest messages in several languages.
- Where Should a Small Hotel Start With AI? — Where a small hotel should start with AI.
- How to Use AI in the Hospitality Industry: Small Hotel and Café Jobs — A 12-room hotel with a café ranks four AI jobs by hours back and risk, with prompts and fixes for guest messages, review replies, rotas and menus.
- 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 Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI help articles (Data controls FAQ, Projects in ChatGPT), OpenAI Terms of Use, OpenAI ChatGPT Business pricing page. Hotel scenarios, figures and sample messages are invented for illustration.