The languages an AI agent handles on WhatsApp: how it detects and mirrors the customer's language, which languages are strong, where retrieval gets weaker, and how to set it up for a worldwide audience
A business that serves the world gets messages in a dozen languages, and hiring for each is impossible. The honest question is not whether an AI agent speaks Spanish or Arabic (it does) but how well, in which situations, and what the owner has to do to make it reliable.
Quick answer
An AI agent on WhatsApp reads the customer's message in any major language and replies in the same language, from a knowledge base written in one language, with no setup beyond a line in the instructions. Major world languages are strong, including those with non-Latin scripts. Mixed-language messages and informal spelling work. The weaker spot is retrieval: the agent finds the relevant lines in your knowledge by matching terms, and a question in Thai or Arabic matches English knowledge less precisely, so a glossary of key terms in the languages you expect improves accuracy. ReplyKit handles the language work on every plan.
How it works
The language model that writes the replies was trained on text in many languages, so understanding Spanish or writing Japanese is not a feature switched on per language; it is how the model works. When a message arrives, the agent reads it, finds the relevant parts of your knowledge base, and composes a reply in the language the message was written in. Your knowledge stays in English (or your language). The instruction is one line: "Reply in the customer's language." More on the setup is in enquiries in a language you do not speak.
What is strong
| Area | Languages | Notes |
|---|---|---|
| Major European | English, Spanish, Portuguese, French, German, Italian, Dutch, Polish, Russian, Turkish | Reliable, including informal register |
| Middle East and South Asia | Arabic, Hebrew, Hindi, Urdu, Bengali, Tamil | Reliable; dialectal Arabic is handled but formality may need setting |
| South-East Asia | Indonesian, Malay, Thai, Vietnamese, Tagalog | Reliable; mixed English (Manglish, Taglish) works |
| East Asia | Chinese (simplified and traditional), Japanese, Korean | Reliable; formality defaults matter and are settable |
| Mixed and informal | Any combination, typos, slang, emoji | Handled; see the slang article below |
The list is not exhaustive; dozens more languages work. Very small languages and heavy regional dialect are where quality drops, and the agent should say it is unsure rather than guess. Slang and code-switching are covered in slang, typos and mixed languages.
Where it is weaker: retrieval
Before writing a reply, the agent selects the relevant lines from your knowledge base by matching the question against the text. That matching works on terms, and a question in a non-Latin script matches English lines less precisely than an English question does. In practice the agent may say "I'm not sure" for a Thai question your English FAQ answers. The fix is a glossary: your product and service names, the words customers use for them in their languages, place names and payment methods, in the knowledge base. Retrieval is explained in how retrieval works, and the glossary method in knowledge for a multilingual agent. We tell customers this plainly because a confident claim of perfect multilingual accuracy would be untrue.
Formality and tone
- Languages with formal and informal address (German, French, Japanese, Korean, Spanish, Portuguese) need a default: set it in the instructions per language.
- The persona (short, warm, one question at a time) carries across languages; the phrases to use and avoid are language-specific, so list them in the languages that matter.
- Emoji conventions vary; a rule like "at most one emoji" is safe everywhere.
- Dates, times and currencies should be stated with zones and codes so the agent does not have to guess conventions; see multiple time zones.
Setup for a worldwide audience
- Instruction: "Reply in the language the customer writes in; if unclear, use English."
- Formality defaults for the languages you expect.
- A glossary of untranslatable terms and customer vocabulary.
- Test with ten questions in each of your top three languages in the test chat.
- Review the unanswered-questions list weekly for foreign-language gaps; each is a glossary line.
Frequently asked questions
Which languages can the agent reply in?
All major world languages, including those with non-Latin scripts. Very small languages and heavy dialect are less reliable.
Do I need to translate my knowledge base?
No. One language is enough. A glossary of key terms in other languages improves accuracy.
Why would the agent say it is unsure for a question my FAQ answers?
Retrieval matches terms, and a question in another script matches English lines less precisely. Add the customer's words for the topic to the glossary.
Does it handle mixed-language messages?
Yes, including code-switching, slang and typos.
Can I set formality per language?
Yes, with a line in the instructions for each language that needs a default.
Is there an extra charge for languages?
No. It is included on every plan.
One knowledge base, every customer's language
Replies mirror the customer's language from day one. 7-day free trial, no card.
Sources and further reading
- Harvard Business Review, The Short Life of Online Sales Leads (Oldroyd, McElheran, Elkington, 2011) · firms that responded within an hour were about seven times more likely to qualify the lead
- WhatsApp Business Messaging Policy · opt-in, opt-out and prohibited-use rules every business must follow
ReplyKit is an independent product and is not affiliated with or endorsed by Meta or WhatsApp. This article is general guidance for running a small business, not legal, financial or regulatory advice.