AI receptionists: what the phone actually handles, what email does, what messaging does

The switchboard is probably the least measured channel in an independent hotel. The property management system records bookings, the channel manager records where they came from, revenue management records prices: none of these systems records the calls nobody answered. A request that arrived at eleven at night, or during the five o’clock check-in peak, simply does not exist anywhere. It appears in no profit and loss account, it feeds into no indicator, and so it is never discussed.

It is into this measurement gap that conversational agents applied to the front desk have moved. The subject has stopped being experimental and has taken on real relevance for small properties too, for two reasons worth keeping separate. The first is technical: language models have made possible a telephone conversation that does not run through a menu tree. The second is about the labour market, and in Italy it weighs more heavily. According to Unioncamere’s Excelsior survey, front-desk staff account for around 78% of the roles sought in accommodation, together with night porters, customer service staff and reservations managers, and recruitment difficulties affect around 41% of positions. Research by Confcommercio with Roma Tre University estimates a shortfall of 275,000 workers across Italian services in 2026, more than 200,000 of them in services and tourism, attributing 70% of unfilled positions to the simple absence of candidates.

It is worth taking the subject channel by channel, though, because phone, email and messaging differ completely in maturity, risk and return, and treating them as a single block is the fastest way to make the wrong decision.

A note on the case studies in circulation

Before going further, a word of methodological caution. Most of the cases cited as evidence of the state of the art do not describe current technology.

The most recurrent is Hyatt: eight contact centres, around seven million calls a year, $4.4 million saved, a third less cost per call. The primary source is a marketing document from the vendor that built the system, and it dates from 2017. It was not a generative model but assisted speech recognition, supported upstream by human operators processing thousands of utterances per shift to bring accuracy to acceptable levels. It is a legitimate case, but it describes another technology and another era. Likewise, the virtual concierge launched by a Marriott group brand, often presented as a current example, is a pilot that began in December 2023 at three properties in the United States. And the single most quoted figure in this sector — the 58% reduction in front-desk workload attributed to a voice agent with chain approval — cannot be verified outside the vendor’s own material.

This does not mean the technology does not work: it does, and better than those dated cases suggest. It means that the basis for the decision has to be found in your own numbers, not in those published by whoever is selling the system. That premise holds for everything that follows and will not be repeated.

The phone

This is the channel the market has concentrated on, and it is also the one where the gap between what is technically possible and what is prudent is widest.

Technically, a voice agent connected to the property management system now answers in under a second, recognises the caller’s language, checks real availability, quotes a rate and closes a booking. It is no longer the automated switchboard asking you to press a number. The implementations that work best, though, are not the ones replacing the front desk: they are the ones that pick up what the front desk was not handling in the first place. Nights, peak days, the second call while the first is in progress, the request in a language nobody on shift speaks. That is a substantial difference, because it moves the economic assessment away from staff savings — which in a thirty-room property are often illusory, given that the staff are not there anyway — and towards recovering requests that are lost today.

On guest acceptance, independent evidence is thin but consistent. A study published in 2026 in the Journal of Hospitality and Tourism Technology, by researchers at the University of South Florida and West Virginia University, examined the operator’s and the guest’s point of view together across a sample of 145 people. The clearest finding is that the main obstacle is neither language comprehension nor technical accuracy, but what the authors call emotional authenticity: guests continue to prefer human interaction, and reservations about privacy and trust are strongest in the voice context, where the voice is recorded. Industry operators proved significantly more enthusiastic than guests, and the explanation offered is simple: they see the relief in workload, while the guest sees only who is on the other end. The authors conclude in favour of a hybrid model in which automation opens the conversation and carries it as far as is appropriate, then hands over to a person.

The finding is useful because it says where to draw the line: the phone conversation is the point at which a property has the least margin for error, because it is synchronous, cannot be re-read before it is spoken, and leaves the hotelier no opportunity to correct anything in between.

Email

This is the channel least talked about and, for an independent property, the one with the most favourable ratio of benefit to risk. The reason is structural and is rarely spelled out.

Email is asynchronous. It does not require sub-second latency, it does not force a real-time decision, and above all it allows a step that is impossible on the phone: human review before sending. A system that reads the reservations inbox, extracts dates and intent, queries the property management system and prepares a pre-filled draft does not make the mistake of promising a policy that does not exist, because that draft still passes under the eyes of whoever signs it. Handling time for a request drops from several minutes to a few dozen seconds without any decision being delegated to an automated system. In a property where the same person alternates between the front desk and the correspondence, it is the most immediate gain available.

The percentages circulating on this channel are encouraging and consistent with one another: between seventy and ninety per cent of requests resolved without human intervention, response times cut by more than eighty per cent, double-digit increases in acceptance of upgrade offers. These are plausible orders of magnitude for anyone who knows the volume of repetitive correspondence at an average property, where the same question about parking, check-in times and cot availability arrives dozens of times a week. Verification, in this case, is within anyone’s reach: average response time and the share of requests resolved same-day are measurable before and after, on your own inbox, without having to take anyone’s word for it.

Messaging and the OTA inboxes

The third channel is the most fragmented. It covers WhatsApp, messages arriving from booking portals and enquiries from the website, and it has one characteristic that sets it apart from the other two: the expectation of immediacy is as high as on the phone, but text allows the same control as email.

It is also the channel where the constraints are contractual rather than technological. Portal rules on what may be written to a customer before the stay, and on which contact details are actually passed to the property, limit from the outset what an automated system can do. A guest writing from a portal is not yet a direct contact of the hotel, and treating them as one is a mistake that precedes any discussion of artificial intelligence.

One further data point applies to all three channels but weighs most here. According to a widely cited industry analysis, 59% of guests are not recognised as returning customers at check-in, and about a fifth of those choose a different property on subsequent stays. It is a phenomenon anyone can observe in their own house: the memory of the returning customer lives in the head of whoever is on the desk, and it leaves with them at the end of the season.

Where the reasoning stops

The real limits are not the technical ones, which shrink every quarter, but the ones about liability.

The first is that what the agent says binds the property. The most cited precedent is a 2024 decision by a British Columbia tribunal, which held an airline liable for incorrect information given by the chatbot on its own website, rejecting the defence that the chatbot was a separate entity from the company.¹ The sum at stake was trivial and the decision, taken in a small-claims context outside Europe, has no binding force in Italy. The principle, however, is the one any lawyer would apply here as well: a tool used by a business in its dealings with customers has no will of its own, and its statements fall on whoever uses it. The operational consequence is that the agent must never be able to formulate a contractual condition itself — cancellation policy, deposit, non-refundable rate — but must only be able to read it from a controlled source.

The second limit is regulatory, and it became live only a few weeks ago. Since 2 August 2026 the transparency obligations of the AI Act apply to systems intended to interact directly with people: the guest must be able to know, at the latest at first contact, that they are not speaking to a human operator, and this information must surface during the interaction, not be buried in the website’s terms.² Domestically, Italy’s first law on the subject adds a requirement for businesses and professionals to declare their use of artificial intelligence tools in the services they provide.³ For a hotel, compliance is modest: a clear opening sentence and the immediate availability of a handover to a person. But it is a compliance obligation, not a matter of style, and the temptation to pass the agent off as a person — which some vendors present as a product feature — is now the fastest way to turn an investment into regulatory exposure.

The third limit is the one the academic evidence flags and that no technical update resolves: the guest in difficulty, the one calling because something has gone wrong, is looking for a person. Designing escalation as a secondary function means having misunderstood the problem.

What is coming: agentic booking

There is a second movement under way, slower than the noise suggests, and it is the exact reverse of the subject treated so far: not the hotel automating the reply, but the customer delegating the request to an assistant of their own.

Through 2025 and 2026 booking portals opened applications inside general-purpose conversational assistants, Google announced booking functions within its conversational search mode, and some chains connected their own applications to the same ecosystem. Then, in March 2026, the movement slowed: OpenAI withdrew the direct purchase button from its main interface, stopped processing travel transactions and repositioned the assistant on search and discovery, with payment handed off to third-party applications. The share prices of the major portals rose on the news, and analysts read it as an admission that travel is too complex to be sold inside a chat window, at least for now.

Research conducted by one of the largest online distribution groups in 2026 describes the same phenomenon from the demand side, and calls it the trust gap: travellers readily adopt artificial intelligence to plan and compare, but almost 70% still prefer to complete the booking through a brand they trust. That reading is consistent with what emerges from the academic research: automation is accepted as long as it takes work away from the user, not when it makes decisions on their behalf.

For an independent property the practical consequence is less dramatic than the telling suggests, but it is not nothing. There is no need to chase integration with every assistant that appears on the market. What is needed is that the property’s information be consistent and legible to an automated system as well — availability, conditions, services, policies — which is exactly the same work an internal voice agent needs in order not to invent answers. It is the rare case in which the investment required today and the one required for the future scenario coincide.

What could already be done, and almost nobody does

Looking at what is technically available today and goes unused, the list is more interesting than the list of future promises.

The first item is measurement. An agent answering the phone produces, as a side effect, the first complete record of inbound requests the property has ever had: how many calls, at what hour, in which language, asking what, how many went unanswered before. This is data no independent hotel possesses, and it has value independently of the automation: it makes it possible to know whether the front-desk staffing shortage really costs something, and how much. The right question to ask a vendor is not what percentage of calls the agent handles, but whether it returns that data in exportable form.

The second is recognising the returning customer. Connecting the agent to the stay history makes it possible to know, at the first word, that the caller has already been a guest three times. It is a trivial function technically and almost never switched on, because it requires the history to be clean, and at most properties it is not. It is worth noting that the problem blocking automation here is the same one that blocks any loyalty activity: it is not an artificial intelligence problem.

The third is pre-arrival upselling, which on the written channel has a very low risk profile and an immediate return. An early check-in offer sent to someone who has flagged a morning arrival, or a superior room to someone who has booked for an anniversary, requires no decision-making autonomy from the system: it requires only that someone read the data at the right moment, which is precisely what never happens at the front desk on full days.

The fourth is language, and it is the point at which the advantage for Italian seasonal properties is clearest. A coastal hotel receiving enquiries in German, Dutch and Polish in July does not hire three people to cover them: it loses them, or handles them badly. A system that replies in thirty languages does not replace a native-speaking receptionist in the relationship, but it covers the informational phase entirely, which is the part where the enquiry is lost.

The fifth, and the least considered, is feeding this information back into management accounting. Requests turned down because the room was unavailable are an indicator of unmet demand; the hourly distribution of calls is a staffing input; the recurrence of questions about services tells you which information on the website is not working. None of these readings requires sophisticated technology. It requires that conversations stop being an event and become data.

Conclusions

Treating the AI receptionist as a single product leads to wrong decisions in both directions: those who adopt it everywhere discover they have put an automated system exactly where the guest wanted a person, and those who reject it outright carry on losing requests nobody is counting.

The channel-by-channel reading suggests an order. Written correspondence is the most sensible entry point, because it allows human review and delegates no decision. Messaging comes next, with the portals’ contractual constraints to be checked first. The phone comes last, and is best confined to what is not being handled today anyway — nights, peaks, uncovered languages — while declaring the nature of the system clearly to the caller, as the regulation now requires.

As for the scenario in which it is the customers’ assistants that go looking for hotels, the preparatory work is the same as the work needed today to stop an internal system from saying foolish things: consistent information, conditions written in one place, an orderly history. It is an unspectacular observation, but it is the one that holds: no form of automated guest relations works better than the data it rests on.

Frequently asked questions

Can a voice agent replace the receptionist at an independent hotel? No, and the implementations pursuing that goal are the ones with the worst results. The available academic evidence indicates that guests prefer human interaction and that reservations increase precisely in the voice context. The sensible perimeter is the requests that go unhandled today: night calls, overlaps at peak times, languages not covered by the staff on shift.

Which channel is the right place to start? Written correspondence. It is asynchronous, so it tolerates human review before sending, and this removes the main risk, namely that the system formulates a contractual condition on its own. The time saved is immediate and the step is reversible at no cost.

Is the hotel liable for what an automated system tells a customer? The general principle is that a tool used by a business in its dealings with customers has no will of its own and its statements fall on whoever uses it. The operational consequence is that rates, penalties and cancellation policies must be read from a controlled source and not generated by the system.

Is there an obligation to tell the guest they are speaking to an artificial intelligence system? Yes. Since 2 August 2026 the AI Act’s transparency obligations apply to systems intended to interact directly with people, and the information must surface during the interaction rather than sit in the website’s general terms. Italian national law provides for a comparable disclosure duty for the use of artificial intelligence tools in professional and business services.

What does a property need to have in order before switching on a conversational agent? Three things, all of which precede the technology: terms of sale and cancellation policies written in one place and kept current, a record of past stays free of duplicates, and service information consistent across the website, the booking engine and the portals. An automated system does not correct inconsistencies upstream, it propagates them faster.

Notes

Footnotes

1. Moffatt v. Air Canada, 2024 BCCRT 149, Civil Resolution Tribunal of British Columbia. A small-claims decision, with no binding force in the Italian legal system. ↩

2. Article 50 of Regulation (EU) 2024/1689 (the AI Act), applicable from 2 August 2026 including to systems already placed on the market. The European Commission’s guidelines on the implementation of the article were approved on 20 July 2026. The obligation to machine-readably mark generated content runs instead from 2 December 2026. ↩

3. Italian Law no. 132 of 23 September 2025, in force since 10 October 2025. ↩