A customer in Luxembourg writes the first chat message in French. The next person in the same queue writes in German, English or Luxembourgish. Staff already switch languages in one afternoon. Since 2 August 2026, a bot that takes the first turn also has to say that it is a bot, and it still has to answer from material the company will stand behind.
This model is for an operations lead comparing conversational AI consultants Luxembourg firms can brief. One agent can cover French, German, English and Luxembourgish when language is a routing fact, every automatic sentence has an approved source, and a person can finish the chat. It is not a legal opinion or a client result. Dates below come from the AI Act and from Regulation (EU) 2026/1744, the Digital Omnibus on AI.
Why is a Luxembourg queue already multilingual?
Luxembourgish is the national language. The law of 24 February 1984, still cited by the government for administrative practice, allows French, German or Luxembourgish in administrative and judicial matters. Legislation is drafted in French. English is not an administrative language. It is a working language in finance, industry and the European institutions, but customers still arrive in more than one language.
STATEC's 2021 census remains the detailed official count of languages spoken at work by residents. People could name more than one. French led, at 69.2%, then Luxembourgish (54.4%), English (40.0%) and German (29.5%). Since 2011, English rose 11.5 percentage points, while Luxembourgish fell 5.1 and German 4.7. The count excludes cross-border workers and records speech, not written replies. Use it as a map of the queue, not as next month's log.
The jobs are cross-border. In Regards 02/26, published on 28 April 2026, STATEC reported nearly 494,000 salaried workers in Luxembourg at the end of 2025, 47% of them cross-border (about 232,000). Only one salaried worker in four had Luxembourg nationality. Commuters live in France, Belgium and Germany, so a city desk reads French and German from people who live in Belgium as well as from residents. Design the handoff so a colleague can finish that conversation.
One agent handles the four languages when the operating rules are explicit:
- Detect the language of the message, and reply in that language only when an approved source covers the point you are about to make.
- Keep French, German, English and Luxembourgish as separate test sets. An average score hides a weak Luxembourgish answer.
- Do not translate a contractual sentence into a language nobody has approved. Show the approved text, or hand the chat over.
- When the customer asks for a person, or the bot has no source, stop the bot. Pass the transcript, the detected language and the passages already used.
Give Luxembourgish the strictest gate. A fluent tone is not proof the wording is yours. Until you score replies against your own articles, send unsupported Luxembourgish questions to a person. If the shift cannot continue in Luxembourgish, say so and offer French or German.
What the chat window must show under Article 50
Article 50 of the AI Act has applied since 2 August 2026. For a support chat, do not fold two duties into one date.
Article 50(1) tells providers to design direct interactions so people know they are talking to an AI system, unless that is obvious to a reasonably well-informed person in the real context. Commission guidelines of 12 November 2026 (C/2026/6005) read the exception narrowly. A helpdesk chat that could be mistaken for a colleague does not qualify. Put a clear line on the first turn, in the language of the screen, and make a person easy to reach. Article 50(5) requires that notice to be distinguishable and accessible. One notice fits a routine question. Repeat it for a financial explanation, an insurance question, a complaint or a distressed customer. Annex III high-risk rules apply from 2 December 2027, and high-risk AI in regulated products from 2 August 2028. Those later dates do not move the chat notice.
Article 50(2) is separate. Providers of systems that generate synthetic text, audio, image or video must mark outputs so they can be detected as AI-generated. The mark is not the sentence on screen. Article 111(4), added by the Digital Omnibus, gives providers of generative systems placed on the market before 2 August 2026 until 2 December 2026 for that marking duty. The guidelines say a bot that both chats and generates receives the extra time only for the mark. A system placed on the market after 2 August 2026 does not receive the four months.
A firm that builds the bot and puts it into service under its own name is a provider, including an in-house build. A firm that buys a vendor bot is a deployer and has to leave the disclosure on. Codes of practice under the amended Article 50(7) concern marking and labelling in paragraphs 2 and 4. They do not postpone the chatbot notice.
Quoting an approved FAQ line is not the same as rewriting it. Article 50(2) may not apply where the output does not substantially change the input, but the guidelines treat that call as case by case. Article 50(4) covers deepfakes and AI text published to inform the public. Private correspondence is treated as unpublished, and a one-to-one support reply sits there. The guidelines put an Article 50 fine at up to EUR 15 million or 3% of worldwide turnover, whichever is higher, and at the lower figure for an SME. Log that the sentence in the window was shown.
A worked example on your own FAQ
Picture a Luxembourg firm whose chat repeats three questions: which form to send, whether a document has arrived, and the published hours. It has an owned FAQ and a ticket archive. The steps are a design, not a client result.
- Pick one intent family. Use questions the FAQ already answers. Leave complaints, refunds, account access and contract changes with a person.
- Fix the FAQ first. Each article needs an owner, a language, a product, an audience and a review date. If two articles clash, the bot must not choose.
- Mine tickets, do not quote them. Closed tickets show missing articles and the languages people use. They also hold names and one-off exceptions. Do not place another customer's ticket into a new reply.
- Disclose on the first bot turn, in the language of the screen, with a way to reach a person. Repeat it if the question becomes a complaint or a financial explanation.
- Answer only from a current passage. If retrieval finds nothing, say so and offer a person. Do not fill the gap from general knowledge.
- Auto-send an allow-list only. Hours and a checklist that quote the FAQ can go out. A draft about the customer's own file waits, unless a system of record returns a permitted status inside fixed words.
- Hand off a full package. The colleague gets the transcript, the language, the passages and the text already sent. The bot then stays silent. Teams in Belgium and the Netherlands hit the same defect when a retry talks over a person. The Belgian helpdesk workflow keeps classification, knowledge and routing apart so the handoff stays visible.
- Staff the languages. French, German, English and Luxembourgish belong on the rota. If Luxembourgish is uncovered, say which languages a person can take.
The chat holds personal data. Pass the model the question and the few passages it needs, not the CRM record. Name the controller, say why the conversation is kept, and keep a way to find that person's messages. The AI Act line does not replace that notice. In Luxembourg the authority for this processing is the CNPD.
The booking that matches this slice is a multilingual support-bot demo on your FAQ. Bring the articles you already trust. The session should show the disclosure, one grounded reply, and the package a colleague receives on handoff. Arrange it with the Brussels team from the contact page. The scope sits in our conversational AI work, and the same handoff pattern is how we design AI agents for Belgian businesses whose staff or customers also sit across the border.
Where the bot should stay silent
Use the bot for repeated questions, a maintained FAQ and a separate test in each language. Keep it off when articles clash, when raw tickets are the only knowledge, or when the first case is a complaint, a refund or an account change. Also stop if the plan treats 2 December 2026 as the day the bot may identify itself.
Measure each language before you widen the allow-list
Keep a two-week baseline before any customer sees a bot reply. Then review the same figures every week, split by language:
- the share of conversations in French, German, English and Luxembourgish;
- the share completed inside the allow-list, and the share handed to a person;
- unsupported answers in a labelled sample, including replies with no FAQ passage;
- minutes from a handoff request to the first human message, by language;
- the share of bot-started chats in which the disclosure was actually shown;
- the share of drafts a person sent unchanged, and the reason for each edit;
- FAQ articles created or corrected because the bot had no source.
Do not chase containment until unsupported answers and handoff time are acceptable in every language you offer. Strong English numbers, with Luxembourgish questions abandoned, are a failed pilot. Widen the allow-list only inside limits you wrote down first.
Key takeaways
- Say it is a bot on the first turn, since 2 August 2026.
- Reply in a language only when an approved source covers the point.
- Hand off the transcript, language, passages and text already sent.
