Multilingual Customer Support With AI: How to Answer in Any Language

A customer in Lyon opens your chat and writes in French. Your help pages are in English, your two support people speak English and some Spanish, and the question is about a return. A few years ago the honest options were to hire a French speaker, bolt on a translation plugin, or reply in English and hope. Today an AI support agent can read your English returns page and answer in French within seconds.

That changes what "we support several languages" can mean for a small team, and it moves the hard work to what you translate, how you test, and what happens when a person takes over. This guide is for businesses with customers in several countries and a support team that speaks one or two languages: a store shipping across Europe, a SaaS product with users in Brazil and Germany, a clinic in a bilingual city.

It compares the three ways to offer support in several languages, shows where AI answers in other languages go wrong, and gives you a translation table, a handoff plan and a testing matrix. PepoChat details come from its documentation as of October 2026.

Can an AI chatbot reply in the customer's language?

Yes. A chatbot built on a large language model replies in the language the customer writes in without a separate bot per language, and it can answer from content written in a different language. Older rule-based bots needed a separate decision tree for every language.

Multilingual customer support is answering customers in their own language, before and after a sale, on the channels you offer. Cross-lingual answering is the narrower capability that makes it affordable with AI: the source content is in one language and the reply is in another. Your knowledge base is written in English, a visitor asks in Portuguese, and the AI support agent writes the answer in Portuguese from the English passage.

The capability is broad but not uniform. OpenAI's GPT-4 technical report tested a standard knowledge benchmark, machine-translated into 26 languages, and found the model beat the previous English-language state of the art in 24 of them, including lower-resource languages such as Latvian, Welsh and Swahili. That is one benchmark for one model, not a guarantee about your returns policy, and two languages did fall short. Quality depends on the language, the model and, above all, your content.

In PepoChat, the agent replies in the language the visitor writes in and can answer from content in any language it can read, as the product FAQ states. Replies come from OpenAI models and are grounded in your own knowledge base, never in general knowledge about your business. Two things are not translated for you: the texts you type into the widget settings, and your team's replies after a handoff. Both get their own sections below.

Why does answering in the customer's language matter?

Customers buy and stay in their own language, and support is part of staying. The most cited evidence is CSA Research's "Can't Read, Won't Buy" survey of 8,709 consumers in 29 countries, run with Kantar and published in 2020.

According to the CSA Research announcement of that survey:

• 76% of online shoppers prefer to buy products with information in their own language, and 40% will never buy from websites in other languages.
• 75% of respondents say they are more likely to buy from the same brand again if customer care is in their language. Even among the respondents most confident reading English, 60% favour customer care in their own language.
• 65% prefer content in their language even if it is poor quality, and 66% use online machine translation.

For support, two numbers stand out. The 75% is about customer care, so the language question does not end at checkout. The 65% says understandable beats polished: a clear answer in slightly imperfect Spanish beats a perfect one in English, which is exactly the gap an AI support agent fills well.

Language can also be a legal requirement rather than a preference. In Québec, the Charter of the French language recognizes the right of consumers to be informed and served in French, and the Office québécois de la langue française enforces it, as the province's consumer protection office summarizes. If you sell to consumers in Québec, French support is not an optional extra. Other markets have their own rules, so check the ones you sell into; this guide is practical advice, not legal advice.

Three ways to offer customer support in multiple languages

There are three approaches, and most teams end up combining two of them. The table compares them on cost, answer quality and the way each one fails.

Native speakers remain the best option for your largest non-English market once volume justifies the cost. They catch tone, know local habits and can make judgement calls your policy does not cover, but they are the most expensive option and the hardest to schedule: one Portuguese speaker covers one shift, five days a week.

A translation layer was the standard approach before language models. The per-character price is low: Google's Cloud Translation pricing lists the first 500,000 characters a month free and $20 per million characters after that for its default model, as of October 2026. The real cost is the plumbing, and the errors two hops add to short, informal messages.

An AI agent answering natively removes both hops: the model reads your content and writes the answer in the visitor's language, which keeps the meaning closer to the source. The risk moves elsewhere. The agent can be fluent and wrong in a language nobody on your team reads, so testing matters more than with the other approaches.

How does an AI agent answer in a language your content isn't written in?

The agent does two things for every message: it finds the passages in your knowledge base that match the question, then writes a reply from them in the visitor's language. Translation happens inside the writing step, one answer at a time, not as a pass over your content.

Think of a bilingual colleague who has read your English handbook: a customer calls in German, the colleague looks up the right page and answers in German. Nobody translated the handbook. That two-step pattern is retrieval-augmented generation, and our plain-English explainer on RAG walks through it.

Two consequences follow for multilingual support.

First, the search step has to find the right passage across languages. Retrieval built on embeddings matches meaning rather than exact words, which is what lets a German question land on an English page. For the top questions in your biggest languages, you can make the match easier with a passage phrased the way those customers ask. In PepoChat the lightest way is a Q&A pair: the question is stored with the answer, so retrieval matches the way visitors phrase it. Add "¿Hacen envíos a Canarias?" with its answer and you have covered that question without translating a page; Q&A pairs do not count as knowledge sources.

Second, the answer can only be as good as the passage: a vague English page produces a vague French answer, fluently. The rules in our guide to stopping chatbot hallucinations apply in every language: one version of the truth, clear statements instead of marketing copy, and an "I don't know" path that hands the conversation to a person. Check that the offer of a person appears in each of your languages too.

Where do AI answers in other languages go wrong?

AI answers in a second language fail in predictable places: legal wording, tone, names, numbers and the languages a model knows least. Knowing the list lets you test for each one instead of discovering it in a complaint.

Three of these deserve more than a table row.

Legal and policy text

A paraphrase of a legal clause is, in effect, a new clause. If your returns policy says "30 days from delivery" and an Italian answer says "30 days from purchase", you have made a promise in Italian that your English policy does not make. Where the policy must exist in the local language, publish an official translation and add it to the knowledge base, so the agent works from text you approved. Elsewhere, an instruction such as "When you explain a policy, keep every number, deadline and condition exactly as written" reduces the drift.

Tone and formality

Many languages encode politeness in grammar: French tu and vous, German du and Sie, Spanish tú and usted. The model picks a register from context, and its default may not match your brand. Agent instructions are free text, so state the register per language in plain sentences: "In German, address customers with Sie. In French, use vous. In Spanish, use tú." Our 12 instruction examples show how to phrase rules an agent follows reliably, and the agent instructions guide explains what the two fields can and cannot change.

Numbers, dates and money

Formats differ more than most teams expect. The W3C's internationalization article on date formats notes that MM/DD/YY is unique to the United States, most of Europe uses DD/MM/YY and Japan uses YY/MM/DD, and recommends month names and four-digit years wherever confusion is possible. So write "4 March 2026", not "03/04/26", on pages the agent will quote. If you sell in several currencies, publish the price for each market and add a Never do this rule such as "Never convert prices into another currency", so the agent never improvises an exchange rate.

What should you translate, and what can the agent handle?

Translate what carries legal or safety weight and what customers must read word for word, and let the agent handle everything it can explain from one accurate source. Most teams need to translate far less than they fear, and a few things far more carefully than they planned.

Two cautions. Every translated page is one more page to keep in sync, and a stale translation recreates two versions of the truth: an outdated German returns page and a current English one produce two different answers. And translated content must live where the agent can read it. A sitemap import or whole-site crawl counts as one knowledge source whatever languages its pages are in, and a crawl's include and exclude filters can target folders such as /fr/ or /de/. Our guide to training a chatbot on your website and PDFs covers each import option.

The greeting and suggested-questions problem

The texts you type into widget settings are shown exactly as written, so they do not switch language with the visitor. Choose a primary language for them, or write a short bilingual greeting, and let the agent switch as soon as the visitor types.

In PepoChat that covers the welcome headline and subline (up to 60 characters each), the greeting that opens every conversation, and up to three suggested questions. The agent's replies follow the visitor's language from their first message; the fixed texts around them do not.

Three patterns work:

• One primary language. Write everything in the language most visitors use. Visitors who prefer another language type in it and get answers in it. Best when one language clearly dominates.
• A bilingual greeting. Two short sentences, such as "Hi, I can help with orders, returns and plans. Écrivez-nous en français si vous préférez." Best when two languages each bring a large share of your traffic, as for a business in Montréal.
• An open invitation. "Ask in any language you like." Five words that tell visitors the agent will follow them. Pair it with an honest line elsewhere about which languages your team replies in.

A tapped suggestion becomes the visitor's first message, so an English suggestion opens the conversation in English. Keep suggestions in your primary language and phrase them as questions your content answers well; visitors can switch language at any point.

Human handoff when your team doesn't speak the language

When the AI support agent cannot answer, or the visitor asks for a person, the conversation reaches your team inbox in the visitor's language. Decide in advance which languages your team answers in, how they will reply in the others, and what the visitor is told.

In PepoChat the handoff works the same in every language: the conversation is marked escalated, every member gets an email with the last three messages, and the operator opens the full transcript in the team inbox, where the visitor panel shows the visitor's browser language and timezone. What the product does not do is translate; the Enhance button improves grammar and clarity in the language you wrote in. That leaves four practical options.

1. Assign it to a teammate who speaks the language. Use the assignee menu in the conversation header and leave an internal note with any context; notes never reach the visitor.
2. Translate your reply with a separate tool. Paste the customer's message into a translation service, write your answer, translate it back, and keep the sentences short and literal so the round trip survives.
3. Reply in plain English and say so. For a simple confirmation, short sentences in plain English are often enough, provided you say up front that your team replies in English.
4. Bring the speaker into the inbox. If a contractor or colleague covers a market, invite them to the workspace so they can reply directly. Starter includes three members and Growth ten, with extra seats at $15 a month.

Set expectations before anyone waits. Visitors are not emailed when your team replies; they see the answer when they reopen the chat in the same browser. If a translated reply will take a while, ask for an email address, or switch on lead capture before a handoff so the agent asks for contact details once before it escalates. With business hours on, make the away message (up to 300 characters) name your languages: "Our team replies in English and Spanish, Monday to Friday. The assistant can help in other languages at any time." Our guide to chatbot human handoff covers the rest of the escalation design.

How to test an AI agent in each language before launch

Test every language you expect traffic in, with real questions written the way native speakers write them, and check each answer against the source rather than for fluency. A fluent wrong answer is the failure that matters, and only a reader of the language or a careful back-translation catches it.

Build a small matrix: six question types down the side, one column per language, and one or two questions per cell. Here is the Spanish column of one.

Run the matrix in four steps:

1. Write the questions with a native speaker. Customers type with slang, typos and missing accents, so include some. A question copied from a textbook tests the wrong thing.
2. Use the test drawer. In PepoChat, Test the agent opens your live widget in a drawer; test chats do not count toward your monthly AI replies and stay out of analytics. Bookings and integration actions still run for real, so use a test order.
3. Check five things per answer. The reply language, every number and condition against the source page, the register, product names, and whether a handoff happens when it should.
4. Fix with content, then retest. Add a Q&A pair in that language for any question it missed, or correct the source page, and run the cell again.

After launch, read a sample of conversations in each language every week. PepoChat does not break analytics down by language, so this part is manual: the Topics card shows what visitors ask about most, and the inbox's visitor panel shows each visitor's language. Ten minutes a week per major language usually catches drift early.

What to do next

Start with the language data you already have: the top countries in your web analytics and the languages of the emails your team answers. Pick the two or three languages that matter, publish official translations of the policies those markets need, add Q&A pairs for each language's top questions, and run the matrix above. If your customers span time zones as well as languages, our guide to 24/7 support without a night shift covers business hours and away messages, and the ecommerce question list is a good source of test questions to translate.

PepoChat includes every feature on the free plan, so you can run the whole test on 200 AI replies a month before choosing a plan on the pricing page. Create a free workspace, import your site, and ask it something in the language your customers use.