Multilingual AI Chatbot: One Bot, Three Languages

Translation is the smallest part

The instinct is to treat multilingual support as a translation project: take the knowledge base, translate it, done. That approach produces a bot that is technically able to reply in three languages and practically wrong in two of them.

It is wrong because the questions differ. Customers writing in one language often arrive from a different channel, at a different stage, with different concerns — local delivery and payment methods in one language, international shipping and currency in another. The same translated answer serves both badly.

It is also wrong because a translated answer inherits the structure of the original. A policy written for one market, translated, still describes that market's process — and a customer reading it in their own language now has a fluent answer to somebody else's question.

Language detection, and what it should govern

Detection sounds trivial and determines most of the experience. Three decisions matter more than the accuracy of the detector itself.

1. 1Detect per message, not per conversationPeople switch. A customer may open in Russian, receive an answer, and continue in Azerbaijani because that is the language the product term exists in. A bot that locks to the first message will be wrong for the rest of the conversation.
2. 2Let the business restrict the listIf you serve three languages, the bot should reply in those three even when someone writes in a fourth. Answering in a language nobody on the team can support creates a conversation no operator can take over.
3. 3Decide what happens on mixed input'Salam, сколько стоит delivery?' is a normal sentence in many markets. The rule should be explicit: reply in the dominant language of the message, or in the language of the previous exchange, and never split an answer across two.
4. 4Handle the unknown case deliberatelyWhen detection is uncertain, the safest behaviour is to answer in the channel's default language and offer the others, rather than guessing.

Localised knowledge, not translated knowledge

This is the difference that decides whether the bot is useful. Some material genuinely is the same in every language. Other material is not, and treating the second kind as the first is the most common multilingual failure.

• Shared across languages: product capability, how something works, technical specifications, general policies.
• Different by language and market: payment methods, delivery times and carriers, applicable taxes, opening hours, which branches exist, legal and consumer-rights wording.
• Different in vocabulary even when identical in meaning: customers in one language may use the English product term, while the translated version of that term appears in no search anyone actually performs.
• Different in what is asked at all: in some languages the majority of questions will be operational, in others comparative. This shows in the transcripts within a month.

The practical structure is one knowledge base with per-language entries where the answer genuinely differs, rather than three parallel bases. Three bases drift, and the drift is invisible until a customer is told two different things.

Register and tone

Formality is not a translation setting. Languages carry different defaults, and getting this wrong reads as rudeness rather than as a technical fault.

• Some languages have a formal and informal second person, and the wrong choice is noticeable in the first sentence. Decide per language, not globally.
• Greeting conventions differ in length and warmth. A greeting that is correct in one language reads as abrupt or as excessive in another.
• Brand and product names should stay untranslated, and the material has to say so explicitly or the model will helpfully translate them.
• Apology conventions differ substantially. The wording that sounds sincere in one language sounds evasive translated directly into another.

Escalation, per language

This is the operational question that decides whether multilingual support is real. Automation can answer in a language the team cannot.

• Know which languages your operators actually cover, and in which hours. A perfect Russian conversation escalating to an Azerbaijani-only queue at midnight has produced a worse outcome than a polite refusal.
• Route by language before routing by topic where coverage differs, or conversations arrive at a queue that cannot read them.
• Write the fallback in every language. An untranslated 'I do not know' is the single most visible failure a multilingual bot can produce.
• State honestly when a language is only supported by automation. Customers accept that; they do not accept discovering it after waiting.
• Record the language on the record, so a callback is made by someone who can hold the conversation.

What to measure per language

1. 1Everything, split by languageVolume, containment, escalation rate, satisfaction. Aggregate figures hide a language that is performing badly, and the bad one is usually the smallest by volume — which is exactly why nobody notices.
2. 2Detection accuracySample conversations and check whether the reply language matched the customer's. Mixed-language input is where this fails and it is worth sampling specifically.
3. 3Knowledge coverage per languageThe same audit run three times. Coverage is almost never equal, and the gap is usually in the language the material was not originally written in.
4. 4Escalation outcomes by languageWhether escalated conversations actually reached someone who could answer. This is where a multilingual bot most often looks fine in the dashboard and fails in reality.

How Vexvon handles languages

Language is detected on every message rather than fixed at the start of a conversation, which is the behaviour mixed-language markets require. A company can restrict the list of languages the bot replies in, so it does not answer in a language nobody on the team can support.

The knowledge base is shared rather than duplicated: entries are typed into thirteen categories and targeted at the chatbot, the call agent or both, so an answer that differs by market can be held as its own entry while everything genuinely shared stays in one place. Answer transparency records which entry produced a reply, which makes a per-language coverage audit concrete.

On the voice side, scenarios define the language of a call alongside the direction, the voice and the fields to extract, and eight agent voices are available. Multiple DID numbers can sit behind the same system, so a different number can run a different language rather than asking the caller to choose.

Escalation carries the language with it. The conversation, the last fifteen messages of memory and the customer record all travel to the operator, and because every channel writes to one customer record, a person who wrote in Russian on Instagram and calls in Azerbaijani is one record rather than two.

Frequently asked questions

1. 1Is a multilingual chatbot just a translated one?No. Customers in different languages arrive with different questions, from different channels, with different expectations of formality. Translating the material produces fluent answers to the wrong questions.
2. 2Should language be detected per message or per conversation?Per message. People switch mid-conversation, particularly where a product term exists in only one language, and a bot locked to the opening message will be wrong for the rest of the exchange.
3. 3What should be localised rather than translated?Payment methods, delivery times and carriers, taxes, opening hours, branches, and legal or consumer-rights wording. Product capability and technical detail can usually be shared.
4. 4How do you handle mixed-language messages?Decide the rule explicitly: answer in the dominant language of the message or in the language of the previous exchange, and never split one answer across two languages.
5. 5What data and integrations are required?Per-message detection, a restricted language list, per-language entries where answers genuinely differ, translated fallbacks, and routing that knows which languages your operators cover in which hours.
6. 6When should a human take over?On the usual triggers, plus one more: whenever the conversation is in a language the available operators cannot hold. That case should be stated honestly rather than discovered after a wait.

Read twenty conversations in your smallest language

The fastest diagnostic is to read twenty real conversations in whichever language has the lowest volume. That is where coverage gaps, register mistakes and untranslated fallbacks concentrate, and it is the language nobody reviews precisely because it is small.