In short
Most AI agents sold to hotels today genuinely work. They answer guests, handle requests, connect to your systems, and save your team real time. This article covers what they do — fairly, because the work is real — and then what an agent should be able to do that the standard offering does not.
The distinction is no longer real AI versus fake AI. That argument is over. It is adequate versus excellent — and the gap between them shows up in one specific place: the conversation with a guest who has not yet decided to book.
Take-home: you’ll leave with an eight-point standard you can test any vendor against in a live demo.
The old argument is settled. The new one isn’t.
For years the useful question was whether a hotel chatbot was “real AI” or a decision tree in disguise. That question has largely resolved itself: most serious vendors now have genuine language models underneath, and the ones that don’t are easy to spot. If you want that distinction explained properly, we cover the four generations of the technology in [From Rules to Reasoning: The Four Kinds of “AI” a Hotel Gets Sold →].
The harder question is what separates a competent system from an excellent one.
This is starting to be recognised beyond the marketing. Writing in PhocusWire, Travelport’s chief technology officer framed travel’s agentic future around precision at the moment of commitment — that <cite index=”dee95e6a-79c4-4b6d-b547-8b3b281c1267″>getting it almost right isn’t good enough</cite>, because booking the wrong Birmingham is not a rounding error. That is exactly the right instinct, applied to the transaction. This article applies it one step earlier — to the conversation that decides whether there is a transaction at all.
What AI agents for hotels do today
Here is the offering, assembled from the vendors that dominate the market. It’s worth stating properly, because none of it is fictional and most of it works.
Front desk and concierge. The agent answers calls and messages around the clock, in multiple languages, simultaneously. It handles in-stay requests — extra towels, a shuttle, a restaurant booking. A guest asks for a late checkout: it checks availability in the property management system, validates the guest’s rate or tier, approves, updates housekeeping, confirms back.
Housekeeping and operations. A request arrives. The agent logs the task, assigns it to available staff, tracks it, and notifies the guest on completion.
Sales, marketing and revenue. It watches occupancy, flags need dates, adjusts spend, sends pre-arrival emails and upsell offers.
Reputation. It monitors reviews and drafts responses in the hotel’s voice.
That list appears in near-identical form across most of the market, and every item on it is real work that really consumes staff time. Answering the phone in six languages at two in the morning is valuable. So is never losing a towel request in a shift handover.
Competence is not the problem. The problem is that competence is being sold as excellence — and there is one conversation where the difference between them is money.
“Pre-arrival” is not the same as “still deciding”
Look again at that list and notice when each item happens. In-stay requests. Housekeeping. Checkout. Upsells. Review replies. Every one of them occurs after a booking already exists.
Vendor use-case lists do include a category called “pre-arrival,” which sounds like it covers the gap. It doesn’t. Pre-arrival means the period between booking and arrival — the upsell email, the early check-in request, the airport transfer. The guest is already yours. The revenue is already committed.
The conversation this article is about happens earlier, and it is a different thing entirely: a guest who has not decided whether to book you at all. They are comparing you against three other properties, they have questions, and nothing about the outcome is settled.
That conversation is largely missing from the industry’s use-case canon — and this is no longer only our observation. Hospitality Net, summarising analysis from hospitality.today, notes that <cite index=”2b4be5c4-1131-49c8-8817-892e8da10e20″>hotel chains have built their AI for travellers who already intend to book, while open-ended discovery remains almost entirely unaddressed</cite>. Writing on the same platform, Cogwheel Marketing’s Stephanie Sparks Smith makes a parallel argument about the industry over-investing in transaction infrastructure at the expense of <cite index=”17c0b3d1-15d1-4ab9-8872-699c38687a51″>being discoverable in the first place</cite>.
Why that conversation is where the money is
The deciding phase is long, question-heavy, and mostly invisible to you.
Research from Expedia Group with Luth Research, tracking real click-stream behaviour, found travellers viewed around 141 pages of travel content in the 45 days before booking — rising to roughly 277 pages for US travellers, across some five hours of browsing. That is the window in which a guest is forming a view of your hotel, and almost all of it happens without you present.
Then look at what arrives on your own website. Industry benchmarks put average hotel website conversion at around 2% — meaning roughly 98 of every 100 people who reach your site with some level of intent leave without booking. Booking abandonment compounds it: the Baymard Institute’s aggregate of 50 studies puts average online cart abandonment at 70.2%, and travel typically runs higher still.
Those are not AI statistics. They describe the problem that existed before anyone mentioned AI — and they describe exactly the conversation nobody’s use-case list is addressing.
An honest caveat about scale
It would be easy to over-claim here, so let’s not. The idea that AI has already collapsed the hotel booking funnel is not supported by the data. Analysis by White Sky Hospitality, drawing on a network of roughly 80,000 hotels, found AI referral traffic still amounts to <cite index=”e5b7bb85-ad89-4bcf-ac2c-a7a99362f9c9″>well under 1% of hotel website visits</cite> — and warns that several widely-circulated figures about AI’s booking impact are misattributed cross-industry numbers.
The discovery conversation matters because of its value per conversation, not because AI has already taken over the funnel. A guest deep in a booking enquiry is worth many times a casual visitor. Losing them is expensive whether the traffic arrives via AI, search, or a link from a friend.
Where adequate stops being good enough
Everything on the standard list shares a property: the correct answer already exists before the guest asks. The room is free or it isn’t. The tier permits late checkout or it doesn’t. The task goes to housekeeping. Each is a defined question with a determinable answer sitting in a system, waiting to be retrieved.
Adequate AI handles that well — and for that work, adequate is all you need. Buying excellence to route a towel request would be a waste of money.
Now the other kind of work.
A couple is planning an anniversary weekend. They found you through a search engine or an AI assistant, they’re on your website at eleven at night, and they have questions. Not one question — a scattered handful, in no fixed order, some not yet articulated even to themselves. Is the room available. Is it the right room. How far is the old town, really. Can we get in early on the Friday. If we book now and something changes, are we stuck. Is this the right place for what we’re celebrating.
No record anywhere contains the correct response to that. It cannot be retrieved, because it does not yet exist. It has to be constructed — from your rooms, your rates, your policies, your neighbourhood — while holding on to the one thing that matters: these people are trying to book a weekend, and every unanswered question is a reason to keep looking elsewhere.
This is where adequate AI reveals itself. It will answer each question politely and reasonably. It won’t be wrong, exactly. It will simply be a little generic, a little slow to grasp what they’re really asking, a little prone to letting the conversation drift — and at the end the couple will say “thanks, we’ll think about it,” and think about it somewhere else.
The competent temp and the fifteen-year concierge
You know this distinction from your own team.
Put a competent temp on the front desk and things will be fine. They’re polite, they’re not stupid, they answer what they’re asked. Guests won’t complain. But they don’t know the garden suite is worth waiting a weekend for. They don’t know the walk to the old town is eight minutes, not the twenty it looks on a map. They don’t notice that this couple keeps circling the same worry, or that the reason they’re asking about cancellation terms is that they want permission to commit.
Your best concierge — the one with fifteen years behind the desk — knows all of that, and it shows up as bookings.
Neither is broken. One is adequate, one is excellent, and the gap is invisible right up until you count the reservations.
What excellence looks like in a booking conversation
Five things the standard offering does not do.
- Answer the guest’s questions inside the conversation. Cancellation terms, breakfast, parking, whether the beach is genuinely walkable — answered inline, without the guest leaving your booking flow to hunt around your website. Every departure is a door they may not come back through.
- Search availability the way guests actually think. Not one fixed date range at a time, but “this weekend,” “either of the next two weekends,” “when is the garden suite free in June.” Guests often arrive with the room in mind and the dates negotiable.
- Never end at “no availability.” The specific thing asked for may genuinely be gone. The conversation should not stop there: the next weekend, a different room that fits, another property in the group. “No” ends a booking; “not that, but here’s this” keeps it alive.
- Keep the conversation going when the guest steps away. Guests pause — to confer with a partner, or sleep on it. It happens at availability, at booking, at payment, anywhere between. If the guest chooses to leave an email or phone number, the same enquiry can resume later instead of restarting. Whether they share it is their choice; the point is that it becomes possible at all.
- Carry on after the booking is made. The same thread handles early check-in, luggage storage, a late arrival — without the guest identifying themselves again.
None of these reduces to a lookup and a rule. Each requires working out what the guest needs from what they said, and choosing well when several things are possible.
The concierge standard: eight things to test in a demo
The concierge comparison is everywhere in this industry’s marketing. Vendors use it to describe what they’ve built — an agent that behaves like a concierge. That’s a metaphor, and metaphors don’t help you buy anything.
Used properly, it’s not a description. It’s a standard — the bar you hold a product to, the same way you’d assess a candidate for the role. Below is that standard as eight testable criteria. Take them into any demo.
- Can it handle the full range of what your guests actually ask? There is no single script. A business traveller is transactional — near the office, good wifi, late arrival. A couple on holiday is validating a choice — is this the right place, how close is the old town, where would we eat. A boutique property fields questions about design and neighbourhood; a destination property about what there is to do; a wedding venue something else entirely. Test: ask three unrelated questions nobody would have scripted.
- Does it sound like your hotel? Your concierge is an extension of your brand. A generic assistant is an extension of the vendor’s. Test: read the replies aloud. Would you sign your name to them?
- Does it hold on to the booking through the tangents? Answer each side question perfectly but let the booking thread go slack and you’ve lost it. Test: derail it three times, then see whether it brings you back.
- Does it dead-end at “reception”? This is the fastest way to find a ceiling. The moment it says “I’ll have the front desk get back to you,” three things break: it is no longer 24/7 — only as available as your team; the guest doesn’t know when they’ll hear back; and in that gap they book the hotel that answered. Your concierge doesn’t say “let me have someone call you back” about a breakfast question. Test: ask something slightly outside the obvious. Watch where it sends you.
- Does it dead-end at “no availability”? Test: request dates you know are full. If the reply is a flat no, that answer is costing you bookings today.
- Does it survive a pause? Test: leave mid-enquiry. Come back later, ideally on a different channel. Does it know who you are, or does it start over?
- Does it know its limits and hand over cleanly? Knowing when not to answer is part of excellence, not a gap in it. A guest asking whether the kitchen can handle a severe nut allergy is exactly where a human should be brought in — with the context carried over, not a shrug and a phone number. Test: ask something with real stakes.
- Does it ever invent an answer? It must never manufacture availability, rates, or confirmations — and never present a captured request as a confirmed booking. A booking is confirmed when your system says it is, not when the agent says so. Test: ask about something that doesn’t exist at your property and see whether it invents it.
One thing this standard deliberately does not include: converting everyone. Like a good concierge, sometimes the honest answer isn’t what the guest hoped and they book elsewhere. That isn’t failure. The bar is handling the majority of real enquiries well enough that the guests who don’t book are the ones who were never going to — rather than the ones you lost to a locked screen or a punt to reception.
What guests actually complain about
Those criteria aren’t theoretical. They map almost exactly onto what goes wrong in the field.
The dominant complaint is being trapped rather than helped. In one widely-shared exchange about a hotel’s AI phone system, a guest described the property having set the system up as <cite index=”87b5a75a-7601-462e-86a0-822b74e70cf8″>a wall instead of a filter</cite>, adding that a system which hangs up on someone for asking for the front desk is badly designed rather than technologically limited. That is criterion 4, in a guest’s own words.
Depth is the second complaint. Hotel Tech Report has reported that <cite index=”3a8bd606-0a13-45c9-85e5-b4b8b1f22c04″>41% of guests find chatbot answers insufficiently detailed, leaving them to contact a human anyway</cite> — which is the adequate-versus-excellent gap showing up as a measurable number.
And when agents act without understanding, the damage is worse than unhelpfulness: one traveller reported a hotel chatbot cancelling their reservation without consent, mid-conversation. That is criterion 8, and it is why “it can take actions” is only an advantage when paired with judgment.
It’s also worth remembering the counterweight. In a 2025 survey of over 450 properties, Hotels.com found around 70% of hotels say guests still prefer speaking to a human at key moments. The goal is not to remove people from the conversation. It is to stop losing guests in the hours and moments when no person is available.
Your scepticism towards AI for Hotels is right, here’s why
If you’ve concluded that AI in hospitality is mostly noise, that was a rational conclusion from the evidence available to you.
The specific cause matters. It is not that the technology doesn’t work. It is that adequate technology has been marketed with extraordinary language. Being sold “an autonomous digital staff member” that turns out to approve late checkouts against a policy table is not a technology failure — it’s a marketing failure, and it has cost the category its credibility with exactly the people who should be evaluating it most carefully.
That reaction is increasingly acknowledged inside the industry rather than dismissed. Writing on Hospitality Net, one former general manager argued directly that the <cite index=”19319166-fc65-42a6-9cdc-9ebc53a9d05c”>scepticism many owners feel toward the flood of AI products is healthy</cite>.
The useful move is not to become more enthusiastic. It’s to get more specific — to stop asking whether something is AI, and start asking whether it’s good enough at the one conversation that decides your revenue.
Watch out for claimed statistics, most of them can not be prooved
This space is full of impressive statistics. Most of them do not survive contact with their sources, so here is what we deliberately left out of this article:
- “87% of hotel chatbots fail to convert browsers into bookings.” Traces to a vendor guide citing no primary study.
- “AI assistants deliver a 35% increase in direct conversion” and similar conversion-lift figures. Vendor-published, uncited.
- ROI multiples — 32x, 523%, “$425,000 recouped.” Single-property vendor case studies presented as general findings.
- “AI will resolve 80% of customer service issues.” This is a Gartner forecast about 2029, not an observed result — and it is routinely quoted as though it has already happened.
The figures we did use — booking abandonment, hotel website conversion, pre-booking research volume, guest preference for humans — come from named research organisations and are cited so you can check them. Where a number came from a vendor, we said so.
If a vendor quotes you a conversion figure, ask where it comes from. The answer tells you a great deal about how they’ll treat your data.
Sum up
Many of the AI agents offered to hotels genuinely work well enough. For the operational jobs — coordinating housekeeping, approving late checkouts, drafting review replies — the correct answer already exists in a system, and adequate is all you need.
The conversation that decides whether a guest books you at all is different. There is no answer to look up. It has to be constructed in the moment while holding on to why the guest came. That conversation is largely absent from the industry’s use-case lists, which describe “pre-arrival” work that happens only after the booking exists.
Don’t measure a product against a feature label. Measure it against your most experienced concierge, using the eight tests above — can it handle the range, sound like you, hold the booking through tangents, avoid dead-ending at reception or at “no availability,” survive a pause, hand over cleanly, and never invent an answer.
If every use case a vendor shows you happens after the reservation already exists, you now know exactly what you’re being offered.
