95+ languages

Write it once. Answer it in ninety-five languages.

The assistant detects the language of every message and replies in it, reading from content you only maintain in one language. No translated help centre, no second bot, no language selector for the visitor to find.

Germande

Kann ich meine Bestellung noch ändern?

Bestellungen können bis zum Versand geändert werden — in der Regel innerhalb von zwei Stunden nach der Bestellung.

retrieved from: shipping-policy (English)

Japaneseja

返品にはどれくらい時間がかかりますか?

返品が届いてから5営業日以内に返金処理を行います。

retrieved from: returns-page (English)

Portuguesept

Vocês entregam no Brasil?

Sim, enviamos para o Brasil. O prazo estimado é de 7 a 12 dias úteis.

retrieved from: delivery-zones (English)

Arabicar

هل يمكنني الدفع عند الاستلام؟

الدفع عند الاستلام متاح في مناطق محددة فقط، ويظهر الخيار عند إتمام الطلب.

retrieved from: payment-methods (English)

Every answer above came from the same English source pages.

The translation treadmill, and the way off it

Maintaining parallel help centres is how policies quietly go out of sync. One source of truth is not just cheaper — it is more accurate.

The usual approach

  1. 1Translate the help centre into four languages
  2. 2Keep four versions of every policy in sync
  3. 3Discover in six months that the Spanish refund page still says 14 days
  4. 4Give up on the fifth language entirely

With retrieval

  1. 1Maintain one authoritative set of pages
  2. 2The assistant reads them and answers in the visitor’s language
  3. 3Update once; every language is corrected at the same moment
  4. 4Languages you never planned for are covered anyway

Three things worth understanding

Detection happens per message

Not per session and not per visitor. Someone who opens in English and switches to French mid-conversation gets French from that point, which is exactly what happens when a person is more comfortable in their own language once the conversation gets technical.

Retrieval stays in your source language

The passages retrieved are your own English ones. The model reads them and composes the answer in the visitor’s language — so the fact being conveyed is always the fact you wrote, not a fact from a translated copy that drifted.

Tone survives the crossing better than you expect

Modern models carry register across languages reasonably well. A support voice that is warm and direct in English does not come out stiff in German. It is worth testing with a native speaker before launch, in the two or three languages that matter most to you.

Quality is not uniform, and pretending otherwise helps nobody

Answers in widely spoken languages are consistently strong. Further down the list, phrasing gets more literal and idiom is less reliable. If a language is commercially important to you, spend twenty minutes testing it with a native speaker before you launch — and remember you can always pin the assistant to a single language in the system prompt if that is the safer choice.

Language questions

Do I need to translate my website first?

No. That is the point. The assistant reads your existing content in whatever language it is written and produces the answer in the visitor’s language. Translating your site is still worth doing for SEO and for the reading experience, but it is not a prerequisite for multilingual support.

How does it know which language to use?

From the message itself. Each message is assessed independently, so a mid-conversation switch is handled naturally and no visitor has to find a language selector.

Which languages are supported?

More than 95, covering every widely spoken language and a long tail of smaller ones. The practical limit is the underlying model, and the models available in CustomerBot span the major frontier and open families — several of which are notably strong on non-European languages.

Is quality the same across all of them?

Honestly, no. Quality is best in high-resource languages — English, Spanish, French, German, Portuguese, Japanese, Chinese — and gets more variable further down the list. If a specific language is commercially important to you, test it properly with a speaker rather than assuming parity.

Can I force it to always answer in one language?

Yes. The system prompt is yours to write, so an instruction to always reply in a particular language regardless of input is respected. Some regulated businesses need exactly that.

Does it cost more to answer in another language?

No. Pricing is per conversation and per model tier, not per language. A conversation in Korean costs what the same conversation would cost in English.

Ask it something in a language you speak

Train it on your English pages, then test it the way your customers will. It is the fastest way to judge this properly.

Start free