Skip to content

Multilingual AI concierge: what works and where it stumbles

AI conciergeHOHEIT Redaktion, 7 min readAuf Deutsch lesen

Blick durch eine offene Holztür in einen Salon mit Fischgrätparkett, Lederstühlen, Holzvitrine und hohem Fenster im Morgenlicht

The short answer

A multilingual AI concierge works well when it answers from a checked knowledge base for your property and the reply language is fixed in advance. The weak spots are place names, house rules with exceptions, and prices. Fixed wording, live data from your PMS and a handover to your team help there, along with the AI disclosure in every language.

Key takeaways

  • According to Statistik Austria's preliminary figures, 74.3 percent of overnight stays in 2025 were by guests from abroad, with large differences between the federal states.
  • The knowledge base is maintained in one language; the model replies in the guest's language. Spot checks in the main languages do not replace a trial phase.
  • Errors mainly arise with proper names, house rules with exceptions, and prices and levies. These should be stored as fixed entries or as live data.
  • The AI disclosure under Art. 50 of the EU AI Act (KI-Verordnung) has applied since 2 August 2026 and must appear no later than the first interaction, ideally in the guest's language.

An AI concierge answers guest questions reliably in many languages when it answers from a checked knowledge base for your property and not from general knowledge. It becomes difficult with place names, house rules with exceptions, and prices: there it needs fixed wording, clear boundaries and a handover to your team.

Language is not a side issue for Austrian properties. According to Statistik Austria's preliminary figures, 74.3 percent of overnight stays in 2025 were by guests from abroad [Q1]. How a concierge works in general and what it costs is described on our page on the AI concierge for hotels and holiday apartments.

Which languages do guests in Austria speak?

German remains the most important language, but not the only one. According to Statistik Austria, in 2025 there were around 58.55 million overnight stays by guests from Germany, 40.46 million from within Austria, 11.26 million from the Netherlands and 4.34 million from Switzerland and Liechtenstein (preliminary, published on 30 January 2026) [Q1].

The differences between the federal states are large. In Tyrol, according to the same analysis, around 45.98 million of 50.02 million overnight stays were by guests from abroad; in Styria, around 6.33 of 14.08 million [Q1]. A property in Tyrol therefore needs different languages from one in East Styria, and the guest statistics from your own PMS tell you more than any average.

Bild: Aufgeschlagenes Notizbuch mit handschriftlichen Notizen auf einem Holztisch in einer Hotelrezeption, daneben eine Kaffeetasse, im Hintergrund ein Gast an der Rezeption und warmgelbes Licht.

How does a multilingual AI concierge work?

A language model can understand a question in Dutch and answer in Dutch, even when the knowledge base exists only in German. The facts come from your entries, the language from the model.

How well this works depends on the language and the model. Anthropic reports, for Claude Sonnet 4.5 with extended thinking, results in a knowledge test relative to English: for example 97.0 percent for German, 97.9 percent for Italian and 97.5 percent for French [Q2]. This is a translated multiple-choice test, not a conversation at reception; Dutch is not in the table, and Anthropic recommends that you test the languages relevant to your own use yourself [Q2]. For operation, the documentation also advises that the reply language be set explicitly rather than left for the model to guess [Q2].

Anyone running a multilingual AI concierge should therefore maintain the knowledge base in one language and spot-check the answers in the others. There is no need to maintain a hand translation of every answer.

Geöffnetes Sprossenfenster mit abblätterndem weißem Rahmen, dahinter eine grüne Flusslandschaft unter bewölktem Himmel

Where do multilingual answers go wrong?

Most errors do not arise in grammar, but with terms that cannot be translated. Three areas are particularly risky in practice.

  • Place names: a model translates "Wien" correctly, but for the names of alpine huts or lift systems it may invent a translation that nobody on site would recognise. Set in the knowledge base that proper names remain unchanged.
  • House rules with exceptions: "Dogs welcome, except in the breakfast room and in rooms 12 to 14" must arrive complete in every language. Short, unambiguous rules are safer than running text with subordinate clauses.
  • Prices and levies: the tourist tax, final cleaning and surcharges belong in the knowledge base as fixed values or come live from the PMS. A concierge that converts or estimates prices is a risk.

The tone is also a factor. The formal German "Sie" form of address has no direct equivalent in English, and other languages have their own levels of formality. Decide how formal the concierge should be in each language.

Calculation: how many questions arrive in which language

The following figures are example values for a property with 24 rooms. In the example we take the distribution of origin from the Austria-wide average in Statistik Austria's figures [Q1]; your own distribution from the PMS replaces it.

The example assumes:

  • Occupancy: 60 percent, so around 5,256 room nights a year
  • Length of stay: 2.5 nights, so around 2,102 stays
  • Questions per stay before and during the stay: 3, so around 6,306 questions a year
  • Share of guests from Austria and Germany: around 63 percent, so around 3,970 questions in German
  • Share from the Netherlands: around 7 percent, so around 440 questions
  • Other countries of origin: around 30 percent, so around 1,890 questions, mostly in English

In our example, then, more than a third of the questions would not be in German. Many come in the evening, when nobody is at reception; what then goes unanswered is described in the article enquiries after reception closes.

And for holiday apartments?

Holiday apartments generate fewer questions, but they are more specific. An example with example values: six apartments in Tyrol, mostly foreign guests, self check-in with a key safe.

Here the questions concern arrival, the key safe, rubbish, heating and parking, often shortly before arrival and in English or Dutch. Each answer has to work without a follow-up, because nobody is on site. For the knowledge base this means steps rather than descriptions, with codes and routes that you check individually for each apartment.

With several apartments, assignment matters more than the number of languages. The concierge must know which apartment the guest is staying in before it gives a code or a route, because both differ from unit to unit. This assignment should come from the booking in your channel manager, not from the guest's memory of the apartment number.

When does the concierge hand over to a person?

A multilingual concierge always hands over when there is no reliable answer, and it passes on a summary to your team while the guest is told in their own language that a person will get back to them. Typical cases are complaints, group enquiries, special requests and anything that requires a commitment.

How much an AI concierge actually takes off your hands in daily work, and where the limit lies, is set out in the article what an AI concierge takes off your hands and what it does not. To answer correctly there, it needs the same clean property data that search engines need; the article which data a hotel needs so that AI answers are right explains what that data is.

Must the concierge identify itself as AI in every language?

Yes, the disclosure applies regardless of language. According to the European Commission, Article 50 of the EU AI Act (KI-Verordnung) applies from 2 August 2026 and requires that people are told when they are interacting with an AI system, unless this is obvious (as of 24 July 2026) [Q3]. According to the regulation's text, the information must be clearly recognisable no later than at the first interaction and must meet accessibility requirements [Q4].

In practice this means the notice appears in the language the guest is writing in, not only in German.

How we implement multilingual answers

With us, the concierge answers only from the knowledge base that you approve and from live data in your PMS. It introduces itself as an AI at the start of each conversation and hands over to your team when it has no reliable answer. No AI can completely rule out incorrect information, so before launch your team asks questions for two weeks and corrects the answers.

Setup takes three to five weeks with us, mainly for the knowledge base, and starts at EUR 5,000. Once a month you receive a report with the most frequent questions and the gaps that occur particularly often in one language.

When a multilingual concierge does not pay off

If almost all your guests come from German-speaking countries and rarely write in, a concierge brings little. A well-maintained area with frequently asked questions on your website does almost the same job. And if house rules and prices are not yet written down anywhere, that work comes first, not the concierge.

Take the origin statistics for your property from the PMS, decide from them which languages to cover, and have the concierge set up with clean AI disclosure under the EU AI Act. Book a call.

Sources and further reading

[Q1] Statistik Austria: Record high again, 157 million overnight stays in 2025 (press release, preliminary results, 30.01.2026)

[Q2] Anthropic: Multilingual support, Claude documentation (retrieved on 26.09.2026)

[Q3] European Commission: Transparency obligations under Article 50 of the AI Act (as of 24.07.2026)

[Q4] AI Act Service Desk of the European Commission: Article 50 AI Act (retrieved on 26.09.2026)

Häufige Fragen

Do I have to translate the knowledge base into every language?

No. The model reads the German entries and answers in the guest's language. Only terms that need a different name in another language should be translated, for example the name of a room category in English.

Does the concierge understand dialect or typing errors?

A language model understands colloquial language and typing errors in most cases. With strong dialect or mixed languages, it may need to ask a follow-up question. Test such cases during the trial phase with real messages from your guests.

Can the concierge also translate confirmations and property information?

Technically yes, but legally relevant texts such as cancellation terms should be approved by you in the most important languages. The concierge then refers to the approved version instead of translating it freely.

Which languages should I test first?

The three most common countries of origin in your own guest statistics, not the national average. For the other languages, a smaller sample in the first month is usually enough.

Related service

AI concierge

A digital concierge that answers guest questions on your website, WhatsApp and email, around the clock and in the guest’s language. It only uses your house data, discloses that it is an AI and hands over to your team when unsure.

Scope, process and prices

Talk to us

Direct booking, without the commission

We build the booking site, connect it to your channel manager and keep it running.

Book a first call

How a guest asks on your website.

The HOHEIT Concierge answers from what your house has published, in the guest’s language. What it does not know, it passes on to your team.

  • Labelled as an AI assistant
  • Answers only from approved knowledge
  • Unknown questions become open questions in the portal
Test it with your website

Haus Steinbach is an invented demo house; photos partly AI-generated.

What would this look like for your house?

More reading

How much commission do you pay today?

Tell us about your house. You get an honest assessment and, after a first call, a concept with a fixed price.

Talk to usCall