AI in hospitality pays off on repeatable, transactional guest conversations: booking changes, cancellations, disruption rebooking and routine questions. It keeps a human one step away, and success is measured by containment and resolution, not by deflection. It falls short where a conversation needs judgment or empathy, such as complaints, emergencies and complicated billing disputes. Deployments should treat the assistant as the front of a routing system, not as a replacement for the desk.
Where AI in hospitality earns its keep, and where it stops
The clearest technical account of the pattern comes from researchers describing a hotel-booking chatbot deployed commercially and handling tens of thousands of hotel searches a day. Their bot handles the most common use cases, while human agents take the long tail of specialised and less common requests. Guests can ask for an agent at any point, and the bot hands over on its own when it does not understand a message. When the model’s confidence falls below a threshold, the message is classed as unknown rather than guessed at.
Guest surveys favour a hybrid of AI and human service. A Mews survey reported by Travolution found that 89% of guests want either a hybrid stay mixing AI and human service, or AI working quietly in the background on repetitive tasks, and almost none want an AI-led stay. Preference for a human rises with judgment: 72% preferred a human for bar and restaurant service and 66% for the front desk welcome. Room service orders came closer to an even split, at 50% human against 23% AI.
Even RaftLabs, a vendor selling these systems, says AI may struggle with personal emergencies, complicated billing disputes and accessibility needs that call for creative solutions. A hospitality question-answering prototype described on arXiv performed well on texts shorter than 3,000 words, but on longer texts it lost track of details and made significant mistakes. Hotel assistants answering from long policy documents or multi-property handbooks should have retrieval scoped tightly.
Which guest conversations should a hotel automate?
The deciding question for each conversation type is whether the right answer follows from policy and live data, or depends on reading the guest. The first kind should be automated. The second should stay with staff, with the assistant gathering context and routing.
| Guest conversation | Automate? | Why |
|---|---|---|
| Booking changes and cancellations | Yes, through tested policy logic | Agents can process cancellations and update reservations without a handoff when integrated with booking systems |
| Disruption updates and rebooking | Yes | In airlines, disruptions create spikes of simultaneous rebooking requests that phone systems cannot absorb |
| Routine FAQs (Wi-Fi, pool hours, amenities) | Yes | Routine queries are handled by chatbots today; personalised and complex requests still need staff |
| Room service and taxi requests | Only alongside an obvious human option | Even for taxis, the most automatable task Mews tested, 44% of guests chose a human over 28% choosing AI |
| Complaints and service failures | No: human-led | Upset guests need empathy, real-time judgment and de-escalation |
| Emergencies, accessibility needs, billing disputes | No: escalate fast with context | AI that does not recognise its limits creates terrible experiences in complex situations |
| Safety questions (illness, security, hazards) | Decline and route, never summarise | Bots that cannot cross-reference health warnings should decline safety questions rather than reassure |
AI that can only answer questions, not take action, has limited value. Scripted customer service chatbots follow decision trees, while next-gen chatbots interpret intent and act, for example by actually rebooking rather than explaining how to. The difference between conversational and generative AI matters here, and so does what makes an LLM agent more than a language model: an assistant that can change a booking can also change it wrongly.
Real examples of AI in hotels: Marriott, Hilton and others
RaftLabs, a conversational AI vendor, reports that Marriott International’s chatbot “MC” handles reservations, check-in and service requests across messaging platforms. RaftLabs says MC reduced support queries by over 60%, and credits integration with Marriott’s property management systems as the key success factor.
For Hilton, JobZoneRisk reports only that Marriott, Hilton and IHG have made mobile check-in and digital keys standard for loyalty members; it gives no assistant results.
Two more vendor-reported hotel cases are worth knowing. Quiq says its generative AI agent at Accor raised CSAT from 67% to 89%. RaftLabs reports that “Rose” at The Cosmopolitan reached a 70% deflection rate for routine inquiries and 85% guest satisfaction with AI interactions. The useful thread is not the percentages but the mechanism RaftLabs credits for MC: integration with live property systems.
Airlines and online travel agencies publish the clearer numbers
Airlines and online travel firms have published deployment figures, several through their technology vendors. Air India’s AI.g, according to Microsoft’s customer story, has resolved more than 13 million conversations with a 97% success rate, so only 3% go to a human agent. The same story notes that about half of Air India’s customers choose AI.g first and the other half still prefer human agents, a choice the developers built in from the outset.
Lufthansa’s case study with NiCE Cognigy reports about 16 million AI-handled conversations in a year, with peak days reaching 375,000 interactions. Those agents handle rebookings and refunds during strikes and other disruptions. Quiq reports an automated resolution rate above 40% for an unnamed airline, with conversations 16% faster. As reported by CX Dive, Expedia said AI helped it handle surging volumes when flights were cancelled across the Middle East, leaving human agents free for complex, time-sensitive issues. The same report notes Airbnb’s assistant resolved 40% of inquiries in its most recent quarter.
Rasa, another vendor, puts routine, predictable queries at 60-70% of typical airline call-centre volume.
How hospitality chatbots fail in public
The operator owns what the bot says
On 14 February 2024 the British Columbia Civil Resolution Tribunal found Air Canada liable, in Moffatt v. Air Canada, for misinformation its AI chatbot gave a customer. Air Canada argued that the chatbot was a separate legal entity responsible for its own actions. The Tribunal held that the airline was responsible for all information on its website, whether it came from a static page or a chatbot. The ABA’s analysis notes that this was not a court decision, but treats it as a clear reminder that companies remain liable for their AI tools and need internal policies on accuracy. The award itself was CAN$812.02.
Summaries that bury safety reports
Hallucination in travel means confidently invented flight times, amenities or policy details, which can cost guests a flight or a trip. A subtler failure is summarisation that averages away the warnings. According to a July 2026 Which? investigation reported by ManyFails, Tripadvisor’s AI review summaries consistently masked severe health and safety hazards. Hotels facing legal action over food poisoning outbreaks were described as spotless. At one Cape Verde resort with 102 mentions of food poisoning in reviews, the Ollie chatbot called food-poisoning risk “quite unlikely”. ManyFails’ diagnosis is that consensus-seeking summarisation privileges bland sentiment and smooths over rare but severe reports, and it argues for hard-coded safety tripwires that bypass generated prose. For anyone wondering why fluent models do this, it helps to understand how language models generate text: they produce the likely continuation, not the checked one.
Shipping below the bar of the feature it replaced
Expedia Group relaunched the Hotels.com app in May 2025 with an AI chatbot trip planner as the headline feature. In Skift’s hands-on testing it returned a Manhattan property for a Brooklyn search, suggested activities with no booking links and did not build itineraries despite being called a trip planner. Horwath HTL’s case study draws the right lesson: an AI feature should launch only if it meets at least the accuracy bar of the non-AI feature it replaces, and the failure was staging discipline, not capability.
Physical automation has failed in public too. Japan’s Henn-na Hotel opened in 2015 with 82 robots, and by January 2019 the chain had withdrawn many of them because they broke down often, cost too much to maintain and annoyed guests.
Disclosure under EU AI Act Article 50, and the human behind the label
Article 50 requires that guests be told that they are talking to an AI; it has applied since 2 August 2026, fines reach up to €15 million or 3% of worldwide annual turnover, whichever is higher (SMEs face the lower amount), and it covers non-EU companies serving EU customers.
Three points from the Commission guidelines, as summarised by CMSWire, matter specifically for hotels. AI chatbots embedded in online helpdesks are given as an example of what is not obviously AI, so disclosure is required. Complaints handling is named as a sensitive domain where periodic reminders are likely to be needed beyond the first notice. A business that commissions a bot and puts it in front of customers under its own brand may be the provider, with the disclosure duty that comes with that role, not merely a user of someone else’s product.
CMSWire also makes the point that ties compliance back to service design: the failure that hurts customers is usually not ignorance of who they are speaking to, but being unable to reach a human once the machine has failed, and Article 50 says nothing about whether anyone is behind the label. The practical answer is obvious “speak to a person” paths in chat and voice, escalation triggers on repeated clarification or negative sentiment, and a handoff that carries the full conversation so guests do not repeat themselves.
Beyond guest chat: housekeeping, pricing and upselling
We would not promise ROI figures for AI beyond guest conversations. The pattern is the same, though: AI takes the checking and routing, and people keep the decisions. RapidEye, an AI inspection vendor, argues that AI takes over the checking, scheduling and documentation in housekeeping, not the cleaning. Its case is coverage: supervisors can hand-inspect only a fraction of rooms, and AI can check every room at turnover for missed cleaning and damage while the supervisor still decides what happens next. In computer-vision terms, that is a defect-flagging task on photos the team already captures, which is a much safer job than autonomous judgment.
Mews lists pricing, housekeeping coordination and reporting among the behind-the-scenes work worth automating. On upselling, Canary Technologies describes conversational AI timing offers to context, such as offering late check-out to a guest asking about late-night entertainment. RaftLabs cites 10-25% increases in direct bookings and ancillary revenue. For a wider survey of the role of AI in the travel and hospitality industries, including recommendation and personalisation, see that overview.
On the front desk, the sources describe task displacement rather than replacement. RapidEye’s panel concludes that mobile and kiosk check-in absorb the transactional volume, so teams get leaner and the remaining role moves towards judgment and problem-solving. JobZoneRisk estimates that 70% of front-desk tasks are being displaced while complaint handling and concierge work stay human, though that figure is its own estimate.
How should a hotel measure whether its AI is paying off?
Measure whether guests’ problems got solved, not whether they went away. Quiq puts it plainly: track containment, cost per contact, CSAT and resolution time, not just deflection, because the goal is to solve problems. RaftLabs warns that deflection above 85% can frustrate guests and argues for routing guests to a human when they actually need one.
A measurement procedure that holds up, in order:
- Record the baseline before launch: call volume, cost per interaction and booking conversion.
- Set an accuracy bar per conversation type at least as high as the channel being replaced, and test it in a sandbox with real scenarios before guests see it.
- Track agent handoff rate and booking completion rate. These are the proportion of conversations involving an agent and the proportion ending in a completed booking. Bot failures push guests to request agents and abandon bookings.
- Read those rates against seasonality. Both metrics are noisy and move with factors unrelated to the model, such as seasonality and changes in hotel supply.
- Pair containment with resolution and CSAT, and keep reviewing escalation patterns, so a rising containment rate is not hiding unresolved guests.
- Pilot at one or two high-volume properties for 60-90 days before a phased rollout. RaftLabs says organisations skipping the pilot consistently report failures tied to training data and escalation logic, not the technology.
The feasibility check comes before any of this. Asking an assistant to do a repeatable, menial task instead of staff is a very different proposition from asking it to outthink them, and the first is far more likely to show measurable ROI. If most volume is policy lookups and booking changes, there is a business case to measure, and if it is mostly complaints and bespoke requests, there probably is not.
Engineering the assistant so the numbers hold
The Hotels.com failure was one of staging, not capability, and ManyFails traces Tripadvisor’s to naive summarisation design. The design we would favour lets the language model interpret intent while bookings, refunds and policy enforcement run through deterministic, testable business logic. Rasa makes the same argument for structured flows that can be inspected and audited. Below a confidence threshold, the system asks or hands off rather than guessing. Safety topics go to a tripwire, not to a summariser. Disclosure and the human route belong in the first message from the start, not retrofitted later. The design depends on integration with the booking engine and PMS, and on monitoring that watches resolution alongside containment.
That is the engineering work behind AI for travel and hospitality as TechnoLynx approaches it: start by establishing whether the outcome is plausible at current capability, then build for the conversations that clear that bar. The open question for each operator is not whether AI can talk to guests. Air India and Lufthansa have shown it can at scale. The question is which of its own conversations are policy and which are judgment, and whether the escalation path is good enough that a guest who reaches its limit still gets helped.
