Multilingual AI customer support

Every AI support tool advertises a language count. The count is the least useful thing about it. What decides whether your Arabic, French and English customers get served properly is what happens around the translation, and this page is about those four things: detection, completeness, handoff and rendering.

The four places multilingual support actually breaks

The underlying models all speak roughly the same set of languages, so "100+ languages" tells you almost nothing about how a bot will behave on your account. Here is what actually goes wrong, in the order you will hit it:

• Detection. Guessing the language from a two-word opener, or from a perfectly ordinary English sentence full of foreign place names.
• Completeness. The whole answer arriving in one language, rather than an Arabic greeting followed by English prices.
• Handoff. What happens the moment a human takes over and does not read the language.
• Rendering. Whether the Arabic you sent survives WhatsApp's layout rules with the phone number still readable.

Below is how Replio handles each one, including the parts it deliberately does not do.

Detection reads what they wrote, not where they are

The obvious signals are the wrong ones. A phone's reported locale says English on most handsets in Doha and Lagos no matter what the owner types, and a country code tells you a market, not a language. Replio treats both as weak evidence. The language comes from the words in the message itself. That sounds trivial until you meet the two cases that break it.

Place names are not language

"I want to report a problem with my trip from Lekki to Ikeja" is an ordinary English sentence. Read loosely it looks Yoruba, because Lekki and Ikeja are Lagos districts. That misread happened on a real account, and it matters more than it sounds: the detected language goes on to drive what the human agent sees and what language their reply is translated into, so an English speaker would have been answered in Yoruba. Replio's detector is told to judge spelling and grammar only, and told outright that cities, countries, brands and people's names carry no information about language. A message that mentions Lekki, Doha or Kerala in English is English.

A greeting is not evidence

"Slm" is shorthand for salaam almost everywhere. It is also Turkish SMS shorthand for selam. A Nigerian school's customer opened with it and received a full welcome in Turkish. The fix was not a cleverer guess, it was refusing to guess: while everything a contact has ever written adds up to fewer than eight letters, Replio answers the message in front of it but will not save that reading as the contact's language. A wrong call on a greeting costs one message instead of poisoning the whole thread. Where the text is genuinely short and ambiguous, the customer's country breaks the tie, and only breaks the tie: a Qatari number writing English still gets English.

People switch mid-conversation

Replio re-checks on each new inbound message instead of deciding once at the start, so a customer who opens in French and continues in Arabic is followed rather than corrected. It is also what keeps things sane when someone switches to English to paste an error message and then goes back.

The whole reply, or none of it

The most common failure in a multilingual bot is a translated greeting wrapped around an untranslated answer: a warm مرحبا, and then the prices, opening hours and steps still sitting there in English because they came out of the knowledge base and the model treated them as quotations. Replio's language instruction is explicit that everything lifted from the knowledge base gets translated too: prices, durations, product and class names, requirements, steps. A reply is not finished while the substance is still in the source language.

It also defers to the message in front of it. If a contact's stored language says Arabic but their newest message is plainly English, the newest message wins. History tells you what someone has been writing in; it does not outrank what they just sent.

When plain English beats broken Hausa

This is the section most vendors leave out. Translation quality is not uniform across languages, and a model straining in a lower-resource language produces output a native speaker finds worse than English. On one real account, Hausa replies came back with the word for yesterday used as filler for "anytime", feminine verb forms for a male customer, and a closing line that read roughly as "we put anger to help you". That is not something you fix with more enthusiasm.

So the instruction carries a quality floor. Write it natively, with real words, correct grammar, correct gender agreement and the respectful form of "you", or write that part in clear simple English instead. Never word by word, never two languages inside one sentence, never an invented word you are unsure of. A business that answers in plain English looks fine. A business that answers in broken Hausa looks careless, and the customer can tell immediately.

One knowledge base, not one per language

Maintenance is what actually stops teams going multilingual: a knowledge base per language, each drifting out of date the moment a price changes, and nobody on staff able to proofread half of them. Replio removes the copies. You write your prices, hours, policies and FAQs once, in whichever language you think in, and the AI answers out of that single source in whatever the customer wrote. Change a policy once and it is correct in every language the same minute.

There is also no language list to choose from for customer replies, so there is nothing to add when a new one turns up. The first person who writes to you in Amharic gets an answer without you having known they were coming. The only two settings that use a list are your team's language and the optional fixed reply language below.

Your team keeps working in one language

Set a Team language in Settings and the shared inbox translates in both directions around it. Open a conversation in a language your team does not read and a language bar appears at the top of the thread: a pill naming the customer's language, and a plain sentence reading "Shown in English, replies sent in Arabic". Each incoming message and each of the bot's own replies carries a translated line underneath the original, so an agent can see both rather than taking the translation on trust.

The agent then types in their own language and Replio translates on send. The customer receives fluent Arabic; the agent's original stays in the thread for the agent to read back later. The composer's placeholder changes to an arrow and the customer's language, which is a small thing that stops somebody forgetting which way the conversation points, and if an agent types a reply in a language that does not match the one set, the inbox says so before it goes out and offers to switch. Nobody pastes anything into a translation tab. It is the same shared helpdesk inbox your team already works in.

Worth being straight about the packaging: the AI answering customers in their own language is on for every plan including Free, with nothing to enable. The team-side inbox translation is the Live Translate add-on, included free on Pro and Business and $19 a month on Starter, and it is not part of the Free plan. Trials get a preview so you can watch it work before deciding.

Arabic that renders like Arabic

Nobody demos this part. WhatsApp decides each line's direction from its first strong character, so an Arabic line that happens to begin with an emoji, a digit or a Latin brand name lays itself out left to right and reads wrong. Replio prefixes a right-to-left mark to every line containing Arabic so it renders the way it was written.

Worse, and far more common: grouped digits inside right-to-left text run backwards. A number sent as +974 6659 3558 arrives looking like 3558 6659 974+, which is a phone number a customer will copy and misdial. Replio wraps international and grouped numbers in a left-to-right isolate so they always read forwards, while leaving prices like 30 ريال and times like 9 صباح untouched. Your public help centre flips to right-to-left as well, for Arabic, Persian, Hebrew and Urdu, and can be published in English, Arabic, French and Spanish.

Voice notes in the customer's language

A large share of Arabic-speaking and West African customers send voice notes rather than typing. Automatic language detection on speech is good on a paragraph and poor on the eight-second clip people actually send: one Arabic note came back as Latin gibberish and the bot answered that it could not understand the language. Replio hints the transcriber with the language that customer has been writing in, which is the one per-customer signal worth trusting, and the note comes back right.

The honest limit: speech-to-text does not cover Yoruba, Hausa or Igbo. Chat and translation work normally in all three, but voice notes in them come back rough or empty, and the dashboard warns you at the moment you pick the language rather than letting a customer find it for you.

When you want one language, not mirroring

Mirroring the customer is the default and it suits most businesses, but it is wrong for a shop that genuinely cannot follow up in another language. Answering a French customer fluently in French, when nobody in the building speaks French and the follow-up call will be in Arabic, just moves the disappointment two messages down the line.

So there is a setting for the other behaviour: always reply in one language. The bot then answers in that language whatever the customer wrote, and opens with a single line in their language explaining that this is the language you reply in, said plainly, without apologising at length and without pretending it did not understand them. It replaces the mirroring rule rather than joining it, because the two instruct opposite things and a bot handed both follows whichever it read last.

Arabic, English and French on one number

This is the specific combination Gulf and West African businesses ask for, and it is why this page exists. A Doha clinic fields Arabic and English on one WhatsApp number, often mixed inside a single message. A Casablanca or Abidjan business adds French to the same queue. None of it needs routing or a primary language: each message is detected on its own and answered from the same knowledge base, so the Arabic customer, the French customer and the English customer all read correct prices without anyone maintaining three copies of them.

In practice: a customer writes السلام عليكم، بكم الكشف؟ and gets the consultation fee in Arabic with the clinic's phone number rendering forwards. The next writes "Vous ouvrez samedi ?" and gets Saturday hours in French. The third writes in English and gets English. Your receptionist sees all three in English in the inbox, replies to all three in English, and each customer reads the answer in the language they wrote in. Add a fourth language next month and nothing changes.

What is automatic and what is not

And the caveat that belongs on any page like this: detection and translation are best-effort and vary by language pair. If a language matters commercially to you, the right move is to read the first week of real conversations in it rather than assume, and to keep anything legally binding or clinically precise in front of a human.

Frequently asked questions

Related

Last updated 2026-09-04. Language detection and translation are best-effort and may vary by language pair; please verify critical replies where accuracy is essential.