2026-05-18

Setting up a multilingual WhatsApp chatbot: an SEA-focused guide

How to build a WhatsApp chatbot that handles English, Bahasa Malaysia, Mandarin, and mixed-language messages the way SEA customers actually write them.

If you run an SMB in Malaysia, Singapore, Indonesia, or the Philippines, your WhatsApp customer messages don't come in one language. They come in all the languages, mixed inside the same message. "Hi do you have stock for the ungu one? berapa harga?" is a real customer message from a Malaysian fashion brand. Half English, half Bahasa, one product question, sent at 11pm.

Most chatbot tools handle this badly. They either pick a language at the start and stick with it, or auto-translate every message and lose the casual register that customers actually use. Here's how to set up a multilingual WhatsApp chatbot that handles SEA-style messaging well — what to look for, what to avoid, and how to test it.

What "multilingual" actually means in SEA context

There are three different problems hidden inside "multilingual":

Language detection — figuring out what language a message is in
Code-switching — handling messages that mix languages mid-sentence
Cultural register — answering in the right level of formality / casualness for that language

Most tools handle (1) okay, struggle with (2), and ignore (3) entirely.

A bot that auto-translates "berapa harga" to "what price" and replies in formal English misses the entire vibe of a Malaysian Whatsapp conversation. The customer wrote casually in mixed languages — they're not expecting a translated AT&T-style reply.

The four things to look for in a tool

1. Per-message language detection (not session-level)

Most tools detect language once per session and lock to that for the rest of the conversation. That breaks SEA messaging because customers freely switch. A customer might message in English at first, then switch to Bahasa when they're frustrated, then back to English with a question.

What to ask: "Does your tool detect language per message, or per conversation?" If it's per conversation, ask if there's a way to reset detection mid-thread.

2. Native handling of mixed-language input

A message like "got the floral one in size M? Tak ada?" should not be:

(Bad) auto-translated to "do you have the floral one in size M? Don't have?" (loses register)
(Bad) replied to entirely in English just because the first word was English
(Bad) refused to answer because "language could not be confidently detected"

It should be: answered in the same mixed register, or in whichever language carries the actual question. The reply could be "yes the floral one in M is in stock! RM 89, can ship today if you order before 3pm."

What to ask: "Show me an example of how your bot answers a mixed-language message — copy/paste a real SEA customer message and show me the reply."

3. Voice-note transcription in non-English languages

A surprisingly large chunk of SEA WhatsApp messages are voice notes. Especially from older customers, customers who type slowly, or anyone driving / cooking / multitasking. The voice notes are usually in the customer's most fluent language — typically Bahasa, Tamil, or Mandarin for Malaysian customers; Bahasa Indonesia for Indonesian; Tagalog for Filipino.

A chatbot that can transcribe English voice notes but not Bahasa ones loses 30%+ of the value.

What to ask: "Does your transcription work for Bahasa Malaysia voice notes? Mandarin? Tagalog?" If they say "we support voice notes" without naming languages, push for specifics.

4. Honest fallback when uncertain

The worst thing a multilingual chatbot can do is hallucinate confidently in the wrong language. Better: "sorry, can you tell me again — were you asking about the floral M or another colour?"

What to ask: "What does your bot do when it's not 100% sure what the customer asked? Show me an example."

Setting it up: a 10-minute playbook

If you're using a tool that handles all four things above (Replai is built around this, but other tools may also work), here's how to set it up for a typical Malaysian SMB.

Step 1: Decide your supported languages

Don't try to cover everything. Pick 3-4 you actually have customers in. For most Malaysian SMBs: English, Bahasa Malaysia, Mandarin (simplified). Add Tamil if you have a Tamil customer base; add Cantonese if you're in KL or Penang.

For Singaporean SMBs: English, Mandarin, Malay, Tamil. For Indonesian: Bahasa Indonesia, English, Mandarin. For Filipino: Filipino (Tagalog), English, Cebuano.

Step 2: Add your business knowledge in your primary language

Drop your menu, services, pricing, hours, location into a Google Doc — in whichever language you'd describe them most naturally. You don't need to translate your knowledge into every supported language. A good multilingual bot reads English-only source material and answers in whatever language the customer asked in.

This is counter-intuitive — many tools require you to maintain parallel versions of every doc. Don't go down that path; it's a maintenance nightmare.

Step 3: Add language-specific phrasing tweaks where it matters

Where you DO want language-specific control is in politeness markers and common cultural phrases. Examples:

For Bahasa Malaysia replies, the bot should know to use "boleh" / "ok" casually, not always default to formal "silahkan"
For Mandarin replies, the bot should default to traditional or simplified depending on your audience (most KL audiences read both; Singapore tends simplified; older customers may prefer traditional)
For mixed-language replies, the bot should pick up local terms like "tapau" (Malaysian for takeaway) and respond appropriately

Most tools let you add a "tone + style" instruction in the system prompt. Use it.

Step 4: Test with real messages

This is the most important step and the one most people skip. Take 10 real customer messages from your inbox (anonymise them if you want), paste them into the bot, and read the replies.

Specifically test:

A pure English message
A pure Bahasa / Mandarin / Tagalog message
A code-switched message
A voice note in the non-English language
A message with a typo / abbreviated text-speak ("got stock 4 the floral 1?")
A photo of a product with "ada yang ini?"

If 2/10 replies are wrong, that's normal — fix the source material and retest. If 5/10 are wrong, the tool's multilingual handling isn't ready for SEA traffic.

Step 5: Watch the first week of real traffic

Even after testing, watch your first week of real customer interactions actively. Reply to anything the bot gets wrong, and update your source docs based on the patterns. Most multilingual issues surface in the first 50-100 real conversations.

Common multilingual mistakes (avoid these)

Translating every product description. Wastes time, creates maintenance debt. A good bot reads English source material and answers in the customer's language.

Hard-coding language detection by phone number prefix. A Malaysian number can message in any language. Don't assume.

Using Google Translate as a fallback layer. It's better than nothing but loses register badly. Modern LLMs do this natively much better.

Forcing customers to "press 1 for English, press 2 for Bahasa". Adds friction, defeats the point of WhatsApp's conversational UX.

Treating "we support 50+ languages" as a quality signal. What matters is whether the tool handles your specific languages well — including code-switching and voice. Quantity is irrelevant.

A quick reality check

Test any tool's multilingual handling with this exact message — paste it into their demo or trial:

Hi, got time today ah? Mau check what time you open. Saya nak datang around 3pm with my mom, she sit on wheelchair, can or not? Got parking ke?

A good multilingual bot replies with something like:

Hi! Yes we're open today till 7pm. 3pm is fine — we have wheelchair-accessible parking right at the front entrance, and our shop is on the ground floor so no stairs. See you and your mom then 🙂

A bad multilingual bot replies in stilted full English, ignores the wheelchair question, or asks the customer to "please write in one language".

Try this message on your current bot (or any one you're evaluating). The reply tells you more than any pricing page.