When your chatbot answers in another language: what gets lost in translation

One of the quiet advantages of an AI assistant is language. A store with an English website can answer a Spanish customer in Spanish, a German one in German, at any hour, without hiring anyone. Most of the major platforms advertise this: the assistant detects the customer’s language and replies in it, from content written in another.

It works better than most people expect. It also fails in a few predictable places, and because nobody on the team reads the Polish or Italian conversations, those failures stay invisible.

How it works

The assistant reads your content in its original language, finds the relevant pieces, and writes the answer in the customer’s language. The translation happens inside the answer, not in your content. That means:

• the facts come from your pages, so they’re as right or wrong as your pages are;
• the wording is the model’s, so it can drift, especially for numbers, dates and terms with a specific meaning;
• if you have pages in several languages, the assistant may read one language’s version and not another, and they may not say the same thing.

Where it goes wrong

Numbers and formats. “€1,200” in English is “1.200 €” in German and “1 200 €” in French. Usually the model handles it; sometimes a decimal comma becomes a thousands separator, and €1,50 becomes €150. Dates swap day and month between British and American formats.

Country-specific rules in a translated answer. A Belgian customer asks in French about delivery, and the assistant answers from the page for France, because that’s the French-language page. The answer is in the right language and for the wrong country.

Legal terms. “Right of withdrawal”, “statutory warranty”, “store credit”. Each has a precise equivalent in every EU language, and a loose translation can change the meaning. “Avoir” in French is store credit, not a refund; a model that translates “refund” as “avoir” has changed your policy.

The switch. The customer writes in Dutch, the assistant answers in English. Or the conversation starts in Italian and the hand-over message, written by you in English only, arrives in English. Small things, and they make the business look careless.

The pre-written parts. Greetings, consent messages, buttons and hand-over texts are usually written once, in one language. The model translates its own answers; your fixed texts stay as you wrote them.

A made-up example

A moving company in the Netherlands has an English and a Dutch website. The English page says: “Packing service: €45 per hour for a team of two, minimum 3 hours.” The Dutch page, updated in spring, says €49 per hour; nobody updated the English one.

A German customer asks, in German: “Was kostet der Packservice?” The assistant reads the English page, since there’s no German one, and answers “45 € pro Stunde”. A Dutch customer asking the same question gets €49. Both answers are fluent; one is out of date, and it reaches every customer who doesn’t write in Dutch.

How to test it

Pick the two or three languages your customers write in most, other than your site’s main one. For each, ask five questions that involve:

1. a price with decimals or thousands;
2. a date or deadline;
3. a rule that differs by country (delivery, returns, VAT);
4. a legal term (withdrawal, warranty, refund versus credit);
5. a hand-over, to see what language the fixed message is in.

Compare each answer with the page it should come from. If you don’t read the language, a machine translation of the answer back into yours is enough to check the facts. You’re looking for the number, the date and the condition, not the style.

How to fix what you find

• Keep translated pages in step with the original, or remove them from the assistant’s sources and let it translate from one source of truth.
• Write country rules per country, with the country in the first sentence: “Delivery to Belgium: 3–5 working days, €9.90.”
• Spell out legal terms in each language you sell in, especially refund versus store credit.
• Translate the fixed texts: greeting, hand-over, out-of-hours message, consent.
• Set the reply language rule in the assistant’s instructions: always answer in the customer’s language.

Language support is listed by each vendor; Intercom, for example, documents which languages Fin supports and how it detects them.¹ Zendesk recommends testing AI agents with realistic customer messages, which for an international store means in customers’ languages.²

If your store sells across borders, Assistant Setup tests the assistant in the languages your customers use, and Assistant Care keeps checking them. For the general method, see the pre-season checklist.

Footnotes

1. Intercom Help, “Fin AI Agent explained”, intercom.com/help. ↩

2. Zendesk, “Best practices for testing AI agents”, support.zendesk.com. ↩