How to set up a multilingual chatbot without losing the details in translation
A multilingual chatbot handles everyday questions in your customers' languages, as long as you write the prices, names and policies yourself and test each language.
Key takeaways
- In a 2020 CSA Research survey, 75% of consumers were likelier to buy again with support in their language.
- Start with one knowledge base, then write native-language answers for your most common questions.
- Write prices, product names and policies yourself, since automatic translation checks are weakest on names.
- Test every language with a fluent speaker, and hand payments, complaints and complex chats to a person.
- How does a multilingual chatbot choose which language to reply in?
- Which content setup should you use for each language?
- What should you translate yourself instead of trusting the AI?
- How do you test each language before launch?
- When should a chatbot hand the chat to a person who speaks the language?
- Which multilingual setup fits your business?
A multilingual chatbot can answer customers in their own language without hiring a speaker for every one. In a 2020 CSA Research survey across 29 countries, 75% of consumers said they’re more likely to buy again when customer care uses their language. Here’s how to set it up.
The work comes down to five jobs: choose how the bot picks a language, decide what content it answers from, write the risky details yourself, test each language with someone who speaks it, and plan the handoff, the moment the bot passes a chat to a person.
How does a multilingual chatbot choose which language to reply in?
It detects the language of each message the customer sends and writes its reply in that language, falling back to a default language when it can’t tell. Language detection, the step where the bot works out which language a message is written in, is the first thing to check, because a wrong guess sends the whole reply off in the wrong language.
Don’t assume customers who read English are happy to be answered in it. In CSA Research’s 2020 survey, 60% of respondents with the strongest English still wanted customer care in their native language.
Four settings decide how well detection works for your customers:
- A default language, used when a message is too short to read, such as “ok”, “hi” or a single emoji.
- An allowed list of the languages you can stand behind, so anything else gets a reply in your default language with a short explanation.
- Detection on every message rather than once per chat, so a customer who switches from French to English mid-conversation gets English back.
- The widget’s interface language for buttons and placeholders, which may be a separate setting from the reply language.
If your tool has an allowed list, keep it short at first. A bot that answers in every language it detects will also answer in languages nobody on your team can check. There’s a help center walkthrough of default and allowed reply languages if you want to see these settings in one tool.
Which content setup should you use for each language?
Use one knowledge base for every language, add answers written in the customer’s language for your most common questions, and bring in a fluent speaker for languages where errors would cost you sales. A knowledge base is the set of help articles, policies and question-and-answer pairs the bot answers from.
The bar is higher than “readable”. Liferay’s 2026 Website Localization Report surveyed 800 US multilingual consumers about websites and apps. Of those, 82% were sure they could spot text a machine translated instead of a fluent person, and 56% had backed out of buying because a company’s site or app didn’t fit their language well. The survey didn’t ask about chat, but it shows how closely these customers read.
Option 1: One knowledge base, AI replies in every language
You write your content once, usually in English, and the AI finds the right answer and replies in the customer’s language. It’s the fastest way to start, and there’s only one version to update when a policy changes. LenoChat’s AI agent for customer conversations, for example, composes each answer in the customer’s own language rather than converting an English version, and every answer links to the source it came from.
The weak spot is smaller languages. Researchers who tested smaller AI models on 200 languages using FLORES-200, a standard translation test set, found that low-resource languages, the ones with little text online for models to learn from, got persistently lower quality than widely spoken ones, according to a 2025 paper from the University of Luxembourg.
Option 1 also depends on clean source content. If your knowledge base is thin or vague, the bot has nothing good to translate, so start with knowledge base answers written so a bot can repeat them accurately.
Option 2: Native-language answers for your top questions
You keep one main knowledge base but write the answers to your most asked questions directly in each language, so the bot starts from fluent text where it matters most. A McGill University study of more than half a million chats at a large North American retailer found that six topics accounted for most questions: orders, coupons, products, shipping, payments and account problems.
That gives you a short writing list. Cover those topics in each language and most everyday chats start from text a fluent person wrote, while the AI handles the long tail. You don’t need every page in every language, just the answers customers ask for again and again.
Those answers are mostly product and order information, and that’s the content shoppers weigh when they choose. In CSA Research’s 2020 survey of consumers in 29 countries, 66% said that when two products are alike, they’d go for the one described in their own language. Among people with the least English, about 85% said so.
Option 3: A fluent speaker reviews or takes over
A person who speaks the language checks the bot’s key answers before launch and answers the chats it can’t handle. It costs the most time, and it’s the approach customers give the most credit to. Localization means adapting content to a language and culture rather than only swapping words. In Liferay’s survey, which asked US multilingual consumers about websites and apps rather than chat, only 20% gave AI by itself the credit when localization was good, against 43% who credited people being involved.
Good translation alone doesn’t close the gap. Even among people who judged brands’ website and app translations good or excellent, 54% had abandoned a purchase because the localization fell short, Liferay reported. Treat that as a sign of what multilingual shoppers notice, not as a chat benchmark. A fluent reviewer catches what reads wrong for the country, not just what’s grammatically wrong.
The reviewer doesn’t have to be a full-time hire. A bilingual employee or a freelancer can check your top answers once, then take that language’s handoffs. If you do hire a bilingual agent, a first-week training plan for chat agents covers what to teach them first.
Here’s how the three options compare on the work each one asks of you.
| Approach | Content sets you maintain | Fluent-speaker time | Who answers hard chats |
|---|---|---|---|
| Option 1: One knowledge base | One | None | Your default-language team |
| Option 2: Native top answers | One, plus top answers per language | Once, to write top answers | Your default-language team |
| Option 3: Fluent review | One, plus reviewed answers | Ongoing, for review and handoffs | A fluent agent |
Option 1 asks the least of your team, and option 3 gives the strongest result in a smaller language. Option 2 needs less fluent-speaker time than option 3 and still covers the topics most chats are about, which makes it the usual next step after option 1.
What should you translate yourself instead of trusting the AI?
Write prices, product names, policies and anything a customer might hold you to yourself, because a small slip in those details changes what you’ve promised, and automated checks are least reliable at spotting it. A January 2025 study of automatic translation quality scoring, software that rates how good a translation is, looked at low-resource language pairs and found the scoring went wrong most often on names and other named entities, on transliteration and on tokenization. Transliteration is writing a word in another alphabet, and tokenization is how a model splits text into pieces it can read.
In other words, the tools meant to flag a bad translation are weakest on exactly the details a store can’t afford to get wrong. The same study found that asking a general AI model to grade translation quality did worse than purpose-built scoring tools. So don’t ask your chatbot to check its own translations. Ask a person.
Here’s what to write or check by hand in each language:
- Product and brand names, kept exactly as you sell them, so a reviewer can confirm none were translated.
- Prices and currency, with a clear line on which currency you charge in and whether the bot converts it.
- Return, refund and warranty policies, translated by a fluent speaker because customers will quote them back to you.
- Shipping times, order cutoffs and delivery days, where one wrong word changes the promise.
- Sizes and units, including any size chart that differs by country.
- Your greeting, handoff and closing messages, which set the tone of every conversation.
Prices need extra care in online stores. Some chatbots show your store’s base price as written rather than converting it, so a customer abroad may see a currency they didn’t expect. Decide what the bot should say about currency and write that answer yourself in each language.
See how LenoChat handles this for small teams.
How do you test each language before launch?
Send the bot the questions customers really ask, in each language, and have someone fluent read every reply before you switch it on for real visitors. Use your tool’s private test or preview mode if it has one, so practice chats never reach a customer. For the checks that apply to any bot, work through a checklist for testing chatbot replies before launch first, then add the language steps below.
Step 1: List your languages and top questions
Pick the languages you’ll support from your actual inbox and order data, not from a wish list. For each one, note the ten questions you get most, starting with orders, shipping, returns and payments.
Step 2: Write test messages the way customers type
Include typos, slang, missing accents and messages that mix two languages. Add very short ones like “ok” or “thanks” to see which language the bot falls back to.
Step 3: Check every name, price and policy
Compare each reply against your source content, line by line. Product names should match exactly, prices should show the right currency and policy terms should say what your policy says.
Step 4: Have a fluent speaker read the replies
Ask them to mark anything that sounds translated, too formal or wrong for the country. Multilingual customers are quick to notice these flaws: in Liferay’s 2026 survey, 90% of US multilingual consumers had run into a language or localization flaw on some website or app.
Step 5: Switch languages mid-conversation
Start a chat in one language and continue in another. The reply should follow the customer’s latest message, not the first one.
Step 6: Ask for a person in each language
Type a request for a human in every supported language, including casual phrasings. Confirm the handoff fires and that the message the customer sees is in their language.
Repeat the name, price and policy checks whenever you change a policy or a price. A fix in one language doesn’t carry over to the others unless you check it.
When should a chatbot hand the chat to a person who speaks the language?
Hand it over when the customer asks for a person, when the bot can’t find a confident answer, and when the chat is about a payment, a complaint or a problem with several moving parts. Those are the chats where a wrong reply in a second language costs you the most.
Customers often need that escape route. In an April 2025 Kinsta survey reported by No Jitter, 55% of US consumers said they often had to ask an AI support tool for a human, 71% said it had trouble following a complicated problem and 61% had been given wrong information by it.
A person who speaks the language also keeps the conversation moving. In the McGill study, people on the support team mirrored how customers wrote better than the chatbot did, and when an agent mirrored them, customers answered sooner and more often. Shoppers there seldom took shipping or payment questions to the bot, which shows where people already prefer a person.
McGill marketing professor Vivek Astvansh suggests figuring out what a customer wants before deciding whether a bot or a person should take the chat. He also says to coach your bot and your agents alike to mirror the way each customer writes, as he argued in The Conversation.
Check that the handoff itself works in every language. A good setup spots a request for a human in any language the bot supports and passes the whole conversation to the agent, so the customer doesn’t have to explain twice.
If nobody on your team speaks the language, say so plainly and offer email, where a reply can be written carefully and checked before it goes out. Some questions shouldn’t go to a bot in any language, and this guide to the questions a chatbot shouldn’t answer without a person covers which ones.
Which multilingual setup fits your business?
Start with one knowledge base plus native-language answers for your top questions, and add a fluent reviewer for any language that brings in real revenue. That covers most small stores and service businesses without a translator on staff.
If your customers write in a few widely spoken languages, begin with option 1 and add option 2 once you see which questions come in each language. If a language earns you real money and AI handles it poorly, or if most chats in it are complaints and payments, go straight to option 3 for that language.
If you have one bilingual person, use them where they count most: reviewing the top answers in their language and taking that language’s handoffs. Everything else can stay with the bot and your main team.
Best for: stores and service businesses that get the same everyday questions in two to five widely spoken languages and can have each language checked by a fluent speaker.
Not for: businesses whose customers mostly write in a smaller language nobody on the team can check, or whose chats are mostly payment disputes and complaints.
Ready to answer every customer from one place? Start free with LenoChat.
Frequently asked questions
Do I have to translate my whole knowledge base for a multilingual chatbot?
No. Many AI chatbots can answer from content written in one language and reply in the customer's language. Write native-language answers for your most common questions and your policies, because that's where accuracy matters most.
Can customers tell when a chat reply was machine-translated?
Many believe they can. In Liferay's 2026 survey of 800 US multilingual consumers, 82% were sure they could spot text a machine produced instead of a fluent person.
Which languages do AI chatbots translate worst?
Low-resource languages, the ones with little text online for models to learn from. A 2025 study testing 200 languages found they got persistently lower translation quality than widely spoken languages, so have a fluent speaker check those replies.
Is a multilingual chatbot worth it if most customers read English?
Usually. In CSA Research's 2020 survey, 60% of respondents with the strongest English still wanted customer care in their native language.
Sources
Image credits
Related Articles
MarketingAI chatbot instructions that keep your bot on brand and out of trouble
Five short blocks of AI chatbot instructions keep your bot on brand and away from promises you can’t keep. In Cyara research reported by Agile Brand Guide, 56% of US consumers trusted a company less after its chatbot gave a wrong answer. Each block below is ready to paste. Paste each block into your chatbot’s […]
MarketingHow to plan live chat staffing for a small team
How many chats will your business get? How many chats can one person handle at once? How do you turn chat volume into a headcount? How do you cover lunch, nights and busy days? When should a chatbot cover for a person? Which setup should you choose? You can work out live chat staffing from […]
MarketingHow to handle an order cancellation request chat without losing the customer
Most order changes and cancellations take one reply, once you know whether the parcel has shipped. Wrong addresses are costly: in Sendcloud’s 2025 survey of UK shoppers, 64.3% said a failed delivery would stop them reordering. Here’s how to run each order cancellation request chat. These chats come in three shapes: fix the address, swap […]
MarketingHow to run agency live chat for several clients
Table of contents How do you keep each client separate in one inbox? Who should see each client’s conversations? How do agents write in each client’s brand voice? How do you keep service consistent across clients? Which inbox setup should your agency choose? Agency live chat can cover several client brands without mixing conversations or […]
MarketingMake Your WhatsApp Click to Chat Link Easy to Find
A WhatsApp click to chat link lets customers contact you without saving your number. In a Meta-commissioned Kantar survey fielded from April to September 2025 across 22 markets, 73.3% preferred messaging businesses. Put the link, QR code and button where customers already need an answer. How do you make a WhatsApp click-to-chat link? Photo: Amina […]
MarketingApology message templates for wrong items, late orders, double charges and missed replies
A good apology message template lets you say sorry fast without sounding like a form letter. Only 40% of US complainants came away fully satisfied with how their complaint was handled, according to a 2025 survey by Customer Care Measurement & Consulting. Below are ready-to-paste apologies for four common mistakes. What should every apology message […]