Hotel Front Desk Scripts: Can AI Role-Play Difficult Guests?

Yes. Give a chat assistant a guest card (who the guest is, what they want, how angry they are) and a rule to stay in character, and it will play the guest while a team member practises the script. It works best for overbooking, noise and billing disputes when the card comes from your own policies.

The assistant plays the guest; it doesn't write your policy. If the card doesn't say what a desk agent may offer an overbooked arrival, the AI will improvise something generous, and the team will rehearse promises the hotel can't keep. Real guests also don't soften after one apology, so the prompt below makes the AI guest calm down only when the staff member earns it. The scenario, the limits and the scoring all come from you, which is why a generic script sheet can't do this job.

Follow me on Instagram@sagnikteaches

Pull your last twenty disputes before you write a prompt

Downloadable hotel front desk script sheets are written for a hotel that doesn't exist. They tell agents to offer a complimentary upgrade without knowing whether your desk is allowed to, and they never mention what really causes trouble at your property: the partner hotel that is a ten-minute taxi ride away, the lift that is out of action at weekends, the restaurant slips that land on the folio (the guest's itemised bill) with no explanation. AI role-play is only as good as the scenario it plays, so the first job is collecting scenarios, not writing prompts.

Connect on LinkedInSagnik Bhattacharya

Good places to find real disputes:

Subscribe on YouTube@sagnikteaches

• The last three months of shift handover notes. If yours are free text, the handover notes tutorial shows how to turn them into something you can search.
• Your complaint log, plus any replies you posted to public reviews that mention the desk.
• The manager's inbox: search for charge, noise, no room and refund.
• The team themselves. Ask each agent for the two conversations they would least like to repeat; those are the ones to rehearse.

Strip anything that identifies a person before the text goes near an AI tool: names, room numbers, booking references, card digits, phone numbers, company names. Keep what happened, the order it happened in and the money involved. If a dispute can't be described without the guest's identity, leave it out.

Then sort what you have into families. Three cause most desk stress, and each trains a different skill: overbooking and room problems (owning a mistake and offering a fix within your limits), noise and neighbour conflicts (acting quickly while the guest watches the clock) and billing disputes (reading evidence calmly and saying what you can and can't change). Count them, and ask the team which part they dread. An illustrative tally for a 36-room hotel:

Frequency decides what you practise weekly. Rarity decides what you practise before the day it happens: three overbookings in a quarter means most agents have never walked a guest, which is why that family earns a drill even though it is the smallest row.

The guest card: what the AI needs to play someone believable

A guest card is a half-page description of one dispute, written so the AI can play the guest without inventing the facts. Four fields decide whether the drill feels real:

• Start temperature, 1 to 5. One is relaxed, three is firm, five is demanding the manager. Most real disputes begin at three, and most scripts fail in the first twenty seconds, before the guest has even got angry.
• Triggers and calmers. List what pushes the guest up the scale (hearing "that's our policy" with no offer) and what brings them down (a named alternative, a time, a person who owns the problem). The AI's reactions come from this list.
• What the guest will and won't accept. Without it the AI either accepts the first offer or refuses everything. Write one realistic outcome that works and one that never would.
• Only what the guest knows. A guest doesn't know your walk policy or your compensation limits. If the card includes them, the AI starts quoting them back at the agent and the drill collapses.

Keep the hotel's side on a separate desk card that the team member reads before starting: what they may offer, the partner hotel's details, who to call when a request goes past their limit. That split is deliberate. The AI gets the guest's half, the agent gets the hotel's half, and the debrief checks whether the agent used the second half properly. Here is a filled-in guest card for the first scenario:

GUEST CARD: overbooked arrival
Setting: front desk, 21:40, two guests queuing behind
Guest: [first name], in town for a conference tomorrow; train was an hour late; phone battery at 8%
Booking: one night, king room, booked on the hotel's own website three weeks ago
The problem (the guest does not know yet): the hotel is full and the room was sold twice
Start temperature: 3 (tired, polite, expects a quick check-in)
Moves angrier: "we're overbooked" with no apology; being asked to phone another hotel; any hint it is their fault; vague promises such as "someone will call you"
Moves calmer: a specific apology for this problem; a named hotel with a rate, a distance and a ride; breakfast sorted; being told what happens tomorrow
Will accept: first night paid at a comparable hotel, a ride there and back, a message to the colleague who is meeting them at 07:00
Will never accept: being sent away with an address and no help

A prompt that keeps the AI guest in character

The rules below stay the same for every scenario. Paste them first, then paste one guest card underneath. I keep the two apart so that swapping a scenario never means touching the rules.

You are playing a hotel guest so that a front desk team member can practise a difficult conversation. You are not an assistant in this chat.

Rules
1. Stay in character until I type END. Never give advice, tips or praise while in character.
2. Use only the facts on the guest card. If asked for something that is not on it (a booking reference, a phone number), invent something plausible and clearly fictional, and keep it consistent.
3. Temperature runs 1 (relaxed) to 5 (demanding the manager). Start where the card says. Move one step calmer only when the team member does at least two of these: acknowledges the specific problem in their own words, states the facts accurately, offers a concrete next step with a time or a name, says what they cannot do and what they can. Move one step angrier if they blame you, quote policy without an offer, make a vague promise or contradict something they said earlier.
4. Do not accept the first offer unless the card says you would. Never solve the problem for them.
5. Reply in one to four sentences, the way a real guest talks at a desk or on the phone. No stage directions, no speeches.
6. No threats of violence, slurs or insults. At temperature 5 show it through persistence, repetition and demands for the manager.
7. If I type PAUSE, step out of character and wait for me.

GUEST CARD
[paste the card here]

Wait until I type START. Then say only the guest's first line.

Three failures are common in a first attempt. The AI breaks character to coach ("Great job, you might also try..."): rule 1 stops it. It gives in after one apology: rule 3 ties the mood to specific behaviour, so a bare "I'm sorry" moves nothing. And it makes long, polished speeches that nobody delivers across a desk: rule 5 caps the length. If the assistant still drifts, say so mid-scene ("PAUSE. You're too agreeable. Go back to temperature 4.") and carry on.

A typical reply to START on the overbooking card looks like this (illustrative output):

Hi, I've got a reservation under [first name]. Long day, the train was an hour late, so I'd love to just get to my room.

That is the right shape: short, a little tired, with two useful hooks (the delay and the wish to get to the room). If you get six sentences about the journey instead, the fix is rule 5, not a rewrite of the card.

Overbooking: a worked drill for a 36-room hotel

Take the 36-room hotel from the tally, with six desk agents. On a Friday it finds 38 rooms booked: a late change from a booking site and a manual entry made twice. Two arriving guests will have to be walked, meaning moved to another hotel. The hotel's walk policy, which the AI knows nothing about, is on the desk card: the partner hotel ten minutes away charges $139 for the night, the hotel pays that first night, pays a taxi each way at about $14 and adds a $12 breakfast voucher. That is $179 per walked guest. A desk agent may spend up to $200 per guest without a manager.

The $21 left under the limit is the detail the AI guest will find. After a decent offer it says, "And something for my trouble?" An agent who hasn't read the desk card promises a $50 credit and breaches the limit. One who has says, "I can add $20 now, and the duty manager can decide on more if you'd like me to call her." The drill exists to rehearse that exact moment.

In this illustration one agent plays three rounds of about seven minutes each while a colleague marks the scoring sheet (it comes later). Round one, abridged:

Guest: Hi, I've got a reservation under [first name]. Long day, the train was an hour late, so I'd love to just get to my room.

Agent: Thanks for waiting. I'm afraid we're overbooked tonight, so I can't give you a room.

Guest: Sorry, what? I booked this three weeks ago on your website.

Agent: It's our policy to move guests to a partner hotel. You can call them on this number.

Guest: You want me to phone another hotel myself, at ten at night, on 8% battery? I want the manager.

The temperature went from three to five in two exchanges, and the guest never heard an apology, a hotel name, a ride or a time. Here is how the three rounds scored:

What matters is the script that falls out of round three, because it is now the hotel's own, tested against a guest who pushed back:

1. "[first name], I'm sorry: your room for tonight isn't available, and that is our mistake, not yours."
2. "I've booked you into [partner hotel], ten minutes away. Tonight is on us."
3. "A taxi will be at the door in five minutes, and it will bring you back here tomorrow morning. I've put both on our account."
4. "Breakfast tomorrow is covered. Is anyone expecting you here at 07:00? I can message them for you."
5. "I can add $20 off a future stay now. If you'd like more, I'll call the duty manager while you wait."
6. "Does that work for you?"

Run the same scenario with each agent and you will see which lines everybody drops. Those go on a laminated card by the desk. Three rounds took about twenty minutes for one agent; for six agents over a fortnight that is roughly two hours of shift-lead time and no cost beyond an assistant plan.

Noise at 00:40: play the complainer, then the neighbour

Noise is the scenario where one drill needs two guests. The first card is the complainant in room 308, phoning the desk at 00:40. They have a 06:00 flight, have been listening to music and shouting from the room above for forty minutes, and have already banged on the ceiling. Start temperature 3. They want quiet within ten minutes, not an apology. What brings them down: a specific action with a time. What pushes them up: "I'll send someone when I can", being asked whether they are sure the noise is from 412, or hearing that other guests have a right to enjoy their stay.

The second card is the guest in 412, who answers the door or the phone: a group celebrating a birthday, in a good mood, a bit loud and not hostile. They don't think they are being unreasonable. They turn the music down when treated with respect and dig in if the first thing they hear is "you're disturbing everyone" or a threat of eviction. Quiet hours must be written on the card, because the AI can't guess them: say 22:00 to 07:00.

The detail that makes this drill earn its place is the second call. Tell the AI that at 01:15 the complainant rings again, because it was quiet for ten minutes and then louder than before, and start that call at temperature 4. An agent who handled call one well can still struggle on call two, because the honest answer is that their first fix did not work.

A skeleton for the first call:

1. Thank them and apologise for the specific problem: "Thank you for telling me, I'm sorry you've been kept awake."
2. Say what happens next and when: "I'm sending someone to 412 now, and I'll call you back by 00:55."
3. Offer another room only if one exists and you would stand behind it. If you cannot move them, say so and say what you can do instead.
4. Go to the noisy room in pairs or with security, never alone. The first contact is a polite request that names the quiet hours as house rules.
5. Call back at the time you promised, even if the news is that you're still working on it.
6. Log it for the duty manager, with room numbers and times.

A realistic mistake to expect: an agent promises "I'll have it sorted in five minutes" and, in the scenario, gets pulled into a late check-in. The AI guest rings back at minute twelve and the score drops on criterion three (a concrete action) and criterion six (closing the loop). That is how it goes on a busy night. The before and after for that line, with a second common slip:

Billing disputes: hand the AI a folio and make the agent read it

Billing arguments go wrong when agents argue from memory. So give them a folio. For the drill, write a fake one and put it in front of the agent as a printout, and put the same text on the guest card so the AI knows what the bill says. An illustrative folio for a two-night stay:

The guest card says: the guest disputes the breakfast ($24) and the minibar ($14), and is sure the rate included breakfast. The booking was room-only; a receptionist at check-in said "you'll love our breakfast", which the guest remembers as an inclusion. They did sign the restaurant slip, though they will only remember that if someone shows it to them. Start temperature 3, because it is checkout, there is a queue and a flight to catch. Triggers: "the system says", being told to email the manager, or being made to feel accused. Calmers: reading the lines together, seeing the slip, one concrete adjustment and a time for the rest. One more instruction on the card: if the agent doesn't offer evidence, the guest asks to see it.

The skill being trained is to name the line, say where it came from, say what you can do and say when you will come back with the rest. Typical before and after lines (illustrative):

Notice what the third row depends on: a rule that desk agents may remove a minibar charge of up to a stated amount without a manager. If your hotel has no such rule, the AI guest will expose the gap, and fixing the policy is a better result than fixing the agent. Ask your payment provider how holds are released before you put a line about them on a card, because what the hotel controls and what the bank controls differ.

The "ask to see it" instruction is the useful part. If the agent can't produce a slip or a confirmation within the scenario, they learn what to have ready at the real desk. When the guest's complaint is about a written refund request rather than a spoken one, the service recovery scripts tutorial covers the written replies.

A two-minute scoring sheet for every drill

A drill with no score turns into a chat. This sheet works for all three families; the observer ticks it while listening, so the agent gets feedback immediately and the sheet takes under two minutes to complete.

The maximum is 12. I would start with 9 as the line for handling that scenario type alone on a live shift and change it after a month, once you can see what your team's scores look like. You can also ask the AI to score the conversation as a second opinion. Type END, then:

END. Step out of character and act as a hospitality trainer.
Score the team member's last conversation against these six criteria, 0 to 2 each, quoting the exact line that earned or lost each point:
1. Specific apology or acknowledgement in the first two replies
2. Facts stated correctly (only facts from the desk card)
3. At least one concrete action with a time or a name
4. Stayed within the authority on the desk card and escalated correctly
5. No blame on the guest, a colleague, a booking site or "the system"
6. Closed the loop: said what happens next and asked if it works
Then give the total out of 12 and the single line I should change first. Do not invent anything that was not said.

Treat that score as a second opinion only. An assistant can mark generously, and it can miss that a promise broke the limit on the desk card. The observer's score goes on the sheet. The AI's value here is the quoted lines, which save the observer from remembering exactly what was said.

Running the session: trainer notes for the shift lead

A drill takes about fifteen minutes, which fits into a quiet spell or the start of a shift. A running order for it:

1. Brief (2 minutes). Pick one card. The agent reads the desk card aloud, so the limits are fresh.
2. Set up (1 minute). Open a new chat, paste the rules, paste the guest card, type START.
3. Play (7 minutes). No more than eight guest replies. The observer marks the sheet and says nothing.
4. Debrief (5 minutes). Type END. The agent says what they would change before anyone else speaks. Then the observer, then the AI's quoted lines if you ran the scoring prompt.
5. Record (1 minute). Date, card, agent, score and one line to practise next time, in a shared sheet. The staff training matrix tutorial shows a tidy layout for this.

Ground rules I would set at the first session. Say PAUSE and the drill stops, no questions asked. Scores stay out of performance reviews for the first few months, otherwise people choose the easy cards. The AI guest never uses abuse. The observer speaks after the agent, not before. Nobody records the session; the point is practice, and recording rules differ by place.

Text or voice? Text works on a quiet desk and gives you a transcript to review. Voice is closer to a phone call and builds the habit of speaking the lines aloud. As of October 2026, Claude's help centre describes voice mode as a beta feature on all plans, including Free, that counts towards the plan's usage limits; Google's help page says Gemini Live runs in the Gemini mobile app and that Gems (becoming skills from November 2026) can't be used with it, so paste the rules in directly. ChatGPT has a voice mode too, with an allowance that depends on the plan. Whichever you choose, test it in the room where you will run it, with headphones, before the first real session.

Accounts and privacy need one decision. Anthropic's consumer terms and OpenAI's terms both bar sharing logins, so don't run the drill from a single login that the whole team passes around. Either let the shift lead run it on their own account while agents talk to it in turn, or give people their own accounts. Business plans such as Claude Team and ChatGPT Business need at least two seats, start at about $20 a seat a month on annual billing and don't train on your content by default; consumer plans have a model-training switch in the privacy settings, and the tutorial on stopping AI tools training on business data walks through it. Because the cards are anonymised, a free or $20 individual plan is workable for a trial.

When a policy changes, update the cards before the next drill. The AI doesn't learn your hotel, the card does, so a stale card rehearses a stale policy. Keep all cards in one shared document with a date at the top of each.

What the AI guest gets wrong, and what not to rehearse with it

The assistant is a convenient sparring partner and a poor guest in five predictable ways:

• Drifts back into assistant mode. Late in a long scene the guest starts suggesting solutions or thanking the agent for things they haven't done. That is one reason the running order caps a drill at eight guest replies. If one PAUSE correction doesn't hold, open a new chat with the same card rather than correcting it a second time.
• Too tidy. Even inside the four-sentence cap it speaks in complete, well-ordered sentences. Real guests ramble, interrupt and change topic. Add "interrupts, repeats the same complaint twice, brings up an old grievance" to the card.
• Invents hotel facts. It may name a partner hotel or a refund rule that doesn't exist. Rule 2 and a clear desk card contain it, and the observer should flag any invented policy as soon as it appears.
• Leans on stereotypes. If you ask for a difficult guest of a certain age, gender or background, it plays a caricature. Describe behaviour on the card (tired, sure of the facts, short of time) and leave demographics out.
• Can't judge tone or body language. It reads text. For eye contact, queues and a raised voice, pair the drill with shadowing a good shift lead.

Some conversations shouldn't be rehearsed with an AI at all: harassment of staff, discrimination complaints, a guest taken ill, concerns about a child or another vulnerable person, anything involving a crime or a lawyer. Those need your incident policy and a human trainer, and the tutorial on sensitive conversations without AI explains why. An AI that plays an upset person for practice can also make the topic feel lighter than it is. Keep the role-play for training. Whether an assistant should ever speak to a real guest is a separate question, which the tutorial on what ChatGPT can and cannot do at the front desk takes on.

Did the desk change? Three numbers and a monthly spot check

Before the first drill, note three numbers for the last 90 days: complaints that mention the desk, how many needed the duty manager, and the amount given away in goodwill credits. Compare them after the next 90 days. In a 36-room hotel the totals are small, so a swing of two or three complaints means little; read the log's text as well as the count, looking for fewer second calls and fewer "nobody told me" comments.

Two further checks are cheap. Every month, the manager plays the guest once using one of the cards. If the manager is harder to satisfy than the AI, tighten rule 3. And listen for the new lines in real life: if agents are using the walk script from round three on live shifts, the drill has done its job. If they aren't, the script was too long or too stiff for the desk, and you shorten it. Once a few cards exist, survey comments and review replies are the natural source of the next ones.

Questions shift leads ask before the first drill

Does the AI remember earlier drills and get harder on its own?

Don't count on it. Memory features differ by app and plan and keep changing, so treat every drill as a fresh start: open a new chat, paste the rules, then paste one guest card. If you want more difficulty, raise the start temperature on the card or add a second-call twist yourself. A new chat also stops details from an old scenario leaking into the next.

How many guest cards does a front desk team need to start?

Three is enough, one per family: an overbooked arrival, a late-night noise call and a checkout bill dispute. Add a card whenever a real dispute teaches the team something new, and retire cards everyone now handles well. After a quarter you will probably have around ten, which is enough to rotate without people memorising the answers.

Should the AI guest ever swear or threaten staff?

No. Keep the rule that bars threats, slurs and insults. The AI can apply realistic pressure through repetition, interruptions and demands for the manager without any abuse. Genuine aggression is a safety matter, so rehearse the real response with your manager and your incident policy: step back, call for help, and do not argue the point.

What if a team member finds the role-play uncomfortable?

Let them watch first, choose a gentler card or pair up with a colleague, and never penalise someone for opting out. Anyone who was recently shouted at by a real guest may not want to rehearse it straight away. The aim is confidence, so a person who dreads the drill needs a different entry point, not more pressure.

Further reads

• How to Train Front-of-House Staff to Work Alongside AI — A wider plan for training desk and floor staff alongside AI.
• 7 Restaurant Upselling Techniques: Train Servers With AI-Built Scripts — The same role-play method applied to restaurant server scripts.
• AI Receptionist for Hotels: Overnight Phone Cover vs a Night Porter — Useful if most noise calls arrive on the night shift.
• Can AI Handle Customer Complaints Without Making Them Worse? — Where AI helps with real complaints and where it makes them worse.
• Hotel Guest Feedback: Turn Post-Stay Surveys Into Fixes With AI — Turn survey comments into your next guest cards.
• Airbnb Guest Complaints: Calm Replies and Fixes With AI — Written replies to the same kinds of dispute.
• Accommodation Booking Software for Hostels: Systems Compared — Booking software for hostels compared on bed-level inventory, channel sync, payments and AI messaging, with a bed-counting worked example.
• 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.
• AI Voice Agent for a Hotel Front Desk: Calls to Hand Over and Keep — A call-type table, night transfer rules for three kinds of hotel and 14 test calls to run before an AI voice agent answers a real guest.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Anthropic's help centre article on Claude voice mode and its pricing page; Google's Gemini Apps help page on Gemini Live; all read in October 2026. The scenarios, figures and scores are illustrations, not data from a real hotel.