The question behind the question
Someone always asks whether the AI speaks French. It is the wrong question, and answering it literally gets teams into trouble. The useful question is narrower: when a visitor writes to us in French, what exactly should happen, and what will we do when the answer we give is wrong?
The distinction matters because a language model will almost always produce something in the language it was addressed in. Fluency is not the constraint. The constraint is whether what it says about your product, in that language, is true — and that depends on material you control.
Your content is the real limit
An AI assistant grounded in your help content is repeating your content back, reshaped for the question. When the question arrives in a language your content is not written in, something has to bridge the gap, and that bridge is the model paraphrasing your English article into French on the fly.
For a general explanation this often reads well. For the things that matter most it is exactly where errors appear: a policy term with a specific legal meaning, a plan name, a number with a currency attached, a phrase like business days that does not map cleanly between countries. The answer is fluent and subtly wrong, which is worse than an obvious failure because nobody questions it.
So the honest hierarchy is simple. Content written in the target language produces the most reliable answers. Content written in one language and paraphrased into another is usable for general help and risky for specifics. And nothing at all in a language produces confident invention.
Three options you can actually keep
Answer only in the languages you staff, and say so. The widget states which languages you support, and everyone else gets a clear route to email. Unglamorous, entirely honest, and it never produces a promise you cannot keep.
Write content in the languages that matter. Pick the one or two languages your analytics actually justify and write your top twenty articles in them. This is real work and it is the only option that makes AI answers dependable in that language, with the side benefit of ranking in it.
Let the AI answer generally, and be explicit about the boundary. Allow it to handle general questions in other languages while routing anything about pricing, orders, accounts or policy to a human. This is a reasonable middle path if — and only if — the boundary is configured rather than hoped for.
What is not an option is quietly assuming your English help centre is a multilingual one. That is a decision made by default, and it is the one that produces the confidently wrong answer.
Test before you promise anything
This is measurable in an afternoon, and the test is the same one you would run for any AI change: real questions, scored honestly.
Take ten questions your customers genuinely ask, translated the way a real customer would write them — informally, with typos, not in careful textbook phrasing. Ask them in the target language and score each answer right, wrong or dodged. Read the wrong ones closely rather than counting them, because the interesting failure is the answer that is ninety percent right with a number or a policy term mangled.
Then check the seams. What does the disclosure say in that language? What does the offline form say? If the visitor asks for a human, what do they see? A perfect answer wrapped in an untranslated interface still tells the visitor they are an afterthought.
The handoff is where it breaks
The failure mode nobody plans for is a successful bot conversation. The AI answers three questions in Portuguese, the visitor reasonably concludes they have found a Portuguese-speaking company, they ask something the bot cannot handle, and the conversation lands with an agent who reads only English.
That is a worse experience than never having answered in Portuguese at all, because you raised an expectation and then withdrew it mid-conversation. If you allow the AI to answer in a language you do not staff, the handoff message has to be honest about what comes next: that a colleague will reply in English, or by email, and roughly when.
If you genuinely do staff a language, route on it, and make sure the routing rule is based on something reliable rather than a guess from one short message.
Localise the frame, not just the answers
The reply text is the part everyone looks at, and it is only part of what the visitor reads. The greeting, the placeholder in the input, the offline notice, the pre-chat fields, the consent wording and the business hours all frame the conversation, and a visitor who reads a fluent answer inside an interface in another language notices the mismatch immediately.
Business hours deserve particular attention. Hours shown without a timezone are a small trap for a visitor several zones away, and an after-hours message that promises a reply shortly means something different depending on where the reader is standing.
How MyLiveChat fits
The widget interface can be presented in the visitor’s language, so the frame around the conversation need not be English by default. AI answers are grounded in the knowledge base you build, which is the lever that actually controls quality: articles you write in a language are the ones the assistant can answer from reliably in that language. The knowledge base is where that content lives, and training the AI on your content covers how it is used.
To be clear about what we do not do: MyLiveChat does not silently translate a live conversation between a visitor and an agent, and you should not plan around it doing so. If nobody on shift reads the language, that is a staffing decision, and our guide to multilingual support on a small team works through the realistic options.
What to measure
Start with the share of conversations by language, which most teams have never looked at and which frequently surprises them — either because a language they worried about barely appears, or because one they never considered is a tenth of their traffic.
Then track two quality numbers per language rather than in aggregate: the share of conversations that ended without an answer, and satisfaction where you collect it. An overall score that looks healthy can easily hide one language where every conversation ends badly, and averaging is precisely what stops you seeing it.