AI in Hotels: What Adoption Research Shows and Where to Start

A hotel considering AI needs a more useful question than whether other properties have adopted it: which task can the team define, supervise and evaluate without making guest information less reliable? The HOTREC / HES-SO European Hotel Distribution Study 2026 offers context for that decision. It reports AI use among some surveyed hotels, but it does not establish that adoption improves profitability, reduces staffing needs or suits every operation.

Read the adoption figures with their denominator in mind

The study, published in September 2026, uses 2025 as its reference year. Its survey took place from February to May 2026. The supplied research reports AI adoption at 26%, no use at 59% and plans to adopt at 15%. These are findings from the study’s respondents and analysis, not a count of every European hotel and not a measure of how well deployed tools perform. A plan to adopt is not a completed installation or a measured result.

The study covers 2,713 hotels: 1,882 individual responses from 28 countries and aggregate data from 831 chain properties. Independent-hotel responses are weighted towards smaller and medium-sized businesses, with a median of 39 rooms. Chain information is supplied in aggregate, so it should not be treated as though each chain property completed the same individual questionnaire. Larger, urban, higher-category and chain-affiliated hotels are overrepresented among AI adopters. That association can inform questions about resources and processes; it cannot show that size caused adoption or that adoption produced a financial benefit.

There is another denominator worth keeping separate. The study’s weighted European distribution estimates concern shares of overnights and exclude the chain aggregates. They are not shares of transactions, AI users or all hotels. They should not be substituted for the reported AI figures. Here, the distribution research provides context for the hotel sample; the operational guidance below is editorial reasoning, not an additional finding of the study.

Define the job before choosing the tool

For an owner or general manager, the first decision is whether a proposed use case addresses a recurring task that the hotel can supervise. For a technology director or integrator, the corresponding question is where approved information resides, who can access it and how an error reaches a person able to correct it. A polished answer is not necessarily a correct one: breakfast arrangements, access instructions and service availability can change while an assistant continues to draw on an older document.

A useful starting exercise is to follow a real information flow without putting an AI tool into it. Where does reception obtain an answer to a guest question? Who approves a change? How does a colleague working another shift learn that it changed? If the team relies on conflicting files or informal messages, connecting an assistant to those materials is unlikely to resolve the conflict. Establishing ownership and a maintained source is a prerequisite for a dependable pilot, not a benefit to assume the pilot will create.

Organisational readiness also differs by task. A hotel may be able to prepare draft replies from approved public information while lacking a safe process for giving a tool access to reservations or internal records. A group may have central technology support yet still need each property to maintain its own opening hours and service details. Neither scale automatically makes a particular use case appropriate.

Compare use cases by what staff must verify

Approved FAQ assistance. A limited assistant could draft answers about information the property has authorised for guests, such as breakfast hours or directions. Staff should be able to see the approved source and its review date. The principal risks are stale information, ambiguous questions and an answer that presents a conditional service as guaranteed. Begin with review before sending rather than assuming an unsupervised response is acceptable.

Translation drafts. A tool could prepare a draft in a guest’s requested language from a message written or approved by staff. A colleague should check that times, conditions and the intended tone have survived translation before sending it. Translation is not a substitute for deciding what the hotel can actually offer. If nobody available can confidently review a particular language, the team needs another verification or handoff route rather than treating fluent-looking text as proof of accuracy.

Staff knowledge lookup. Searching approved internal guidance may help a colleague find a relevant procedure, provided the result identifies its source and does not conceal uncertainty. Internal material needs tighter access boundaries than a public FAQ. Instructions involving safety, payment or exceptions to policy should direct staff to the responsible person or current procedure, not invite them to rely on an unverified summary.

Human approval and handoff. Approval is a workflow decision across these uses, not a separate capability demonstrated by the adoption figures. Specify who reviews a draft, what they must check and when a query is passed to a manager or specialist. A request about a disputed charge, an accessibility need or a time-sensitive problem may require direct human handling even if a draft answer seems plausible. If the designated reviewer is unavailable, the workflow needs a fallback.

These comparisons are recommendations for evaluating possible deployments. The study does not establish that these use cases are prevalent, effective or supported by any named product. A purchaser should verify proposed functions, information access and approval controls with the supplier and integrator before treating them as part of a project plan.

Prepare information and access boundaries

A practical information source distinguishes property-wide facts from temporary notices and internal instructions. Each approved item should have an owner, a means of updating it and a way to withdraw it when it becomes obsolete. For a group, central wording might cover a common policy while each hotel owns its local details. When the two conflict, the team needs an agreed resolution process rather than an assistant that chooses between them silently.

The technology team should document which sources a proposed tool may read, which users may edit them and whether guest or staff information is included. Start with the least information necessary for the chosen task. Keep confidential terms and individual guest details out of a public-answer source. Before any connection to operational systems is considered, establish who authorises access, what the tool is expected to do with the data and how staff can detect and correct an inappropriate answer. Those are planning checks, not claims about an existing integration or a specific legal duty.

Maintenance is part of the operating model. Assign someone to update guest-facing details when a service changes, and make the update visible to reviewers. Preserve an escalation route for questions the approved material does not answer. An assistant that admits a gap and hands it to staff is preferable to one that fills the gap with a confident guess. Procurement discussions should therefore cover not only the interface but also source ownership, access restrictions and the work required to keep information current.

Run a hypothetical pilot that can fail usefully

Consider a hypothetical property testing staff-reviewed drafts for routine guest questions. This is an evaluation design, not an observed COTT test or an estimate of savings. The property would first define the eligible questions and record how staff handle comparable queries without the tool. It would prepare approved answers, identify the reviewer on each shift and decide which subjects must bypass drafting entirely.

During the pilot, reviewers would record whether a draft was correct against the current source, whether it needed editing, whether the right language and conditions were preserved, and whether an uncertain or sensitive question reached the right person. They would also record staff effort: reading, correcting and maintaining drafts consumes time even when generation is quick. A guest complaint or a staff correction should be investigated as an information or workflow problem, not merely counted against a model.

Before expansion, the team should agree what level of error warrants a pause, who can stop use and what evidence would justify trying another language or task. Compare the pilot with the initial workflow only where the tasks are genuinely comparable. Query mix, seasonal service changes and differing reviewer availability can distort a simple before-and-after comparison. Without measured results, it would be misleading to report time saved, improved satisfaction or a financial return.

Set different responsibilities for a property and a group

At a single property, a manager can own the approved information, designate reviewers and make escalation routes clear to each shift. A narrow pilot is easier to inspect than several simultaneous uses. After reviewing errors and maintenance effort, the manager can decide to revise the source, change the workflow, extend the pilot or stop it. Staff should know that correcting a draft is part of the evaluation, not a failure to use the tool.

A group needs an additional distinction between central standards and local authority. A central team can set rules for access, review and reporting; each property still needs a named owner for facts that vary locally. Technology directors and integrators should check whether proposed reporting distinguishes properties and tasks clearly enough to locate recurring errors. Expansion should depend on reviewed evidence from the pilot and a workable process for each additional property, rather than on the original deployment schedule alone.

Ellia is COTT.TV’s multilingual guest concierge, built around hotel information and the guest-service experience. Explore the guest-TV demo, then use the evaluation checklist to review answer accuracy, approved content and the route to hotel staff. Keep the hotel team responsible for operational decisions and for maintaining the information guests rely on.

Decision reference

Practical questions

Does AI adoption demonstrate a return on investment?

No. Adoption is a survey measure. A hotel must evaluate correctness, staff handling, guest outcomes and its actual costs in a defined workflow.

Which information should an assistant use?

Use current hotel-approved material with clear owners and escalation routes. Sensitive or uncertain decisions should go to the responsible staff member.

Sources and methodology

The sole external source for the adoption figures and sample description is the HOTREC / HES-SO European Hotel Distribution Study 2026, reference year 2025. Evidence was reviewed on 23 October 2026. The report was published in September 2026 and surveyed hotels from February to May 2026; its reference year is not a 2026 observation. The individual responses and aggregate chain data have different forms, the independent responses are SME-heavy, and weighted European distribution estimates exclude chain aggregates and measure shares of overnights. The reported AI figures should not be read as a census or as proof of causation, profitability or job savings. Use-case comparisons, safeguards and the hypothetical pilot are editorial recommendations, not findings from a customer study, product test or manufacturer certification.

Editorial visual: Inn. TRENDS.