Running WhatsApp support in Hindi and regional languages
Keep your knowledge base in one language and let the AI reply in whatever language the customer wrote in, including Hinglish. Maintaining separate translated bots is how content quietly drifts apart, and script-mixing in Roman-script Hindi is the specific case worth testing before you go live.
- Hinglish, not Hindi, is the real test. Most customers mix within a sentence.
- One knowledge base, many reply languages. Never maintain parallel copies.
- Templates do need separate approved language versions.
- Test transliterated Hindi written in Roman script explicitly.
How Indian customers actually write
Read a week of any Indian business's WhatsApp and the pattern is consistent: very little of it is clean, single-language text.
A customer writes "order kab aayega" in Roman script. Another writes the same thing in Devanagari. A third mixes: "bhai delivery kitne din mein ho jayegi, urgent hai". A fourth writes perfect English and then switches mid-conversation when they get annoyed.
This is why testing a bot only in English and Hindi misses the point. The dominant register is Hinglish written in Roman script, and it is the hardest of the three for a system to handle badly and the easiest to overlook when testing.
Regional languages add another layer. A Pune customer may write Marathi in Devanagari, a Chennai customer Tamil in Roman script, and both expect a reply they can read comfortably.
Voice notes complicate it further. A large share of Indian customers send audio rather than typing, often in a regional language, and a support setup that handles text beautifully can be caught out entirely by a thirty-second voice message.
One knowledge base, many replies
The instinct is to create a bot per language. It is the wrong instinct, for a practical reason rather than a technical one.
Parallel content drifts. You update the return policy in English, forget the Hindi copy, and three months later two of your bots are telling customers different things. Nobody notices until a customer quotes the wrong one back at you.
Keep one knowledge base in the language your team writes best, usually English, and let the model translate at the point of reply. Modern models handle this well for the everyday register support uses, and there is only ever one copy to keep current.
State it explicitly in the system prompt: reply in the language the customer used, including Hinglish, and do not switch unless they do.
Where translation must not be trusted
Some text should be reproduced exactly rather than paraphrased in another language.
- Legal and policy wording: cancellation terms, warranty conditions, refund eligibility.
- Medical instructions and dosage information.
- Prices, and anything a customer could reasonably treat as a commitment.
- Safety and allergen information.
Tell the agent to use the exact wording from the knowledge base for these, in the original language if necessary, rather than producing a natural-sounding translation that subtly changes meaning.
A paraphrased refund policy that sounds more generous than it is becomes a dispute you will lose, because the customer has it in writing from you.
Templates are different
Free-form replies can be generated in any language, but WhatsApp templates cannot.
Each language version of a template is a separate template that Meta reviews independently. A template approved in English can be rejected in Hindi if the translation reads differently, and it needs its own sample values.
Practically, this means keeping template languages deliberately few. Two or three covering most of your customer base is manageable. Ten is a maintenance burden that will quietly rot, and a stale template in a language nobody on your team reads is worse than not offering it.
Convo360 shows which language versions exist for each template, so it is at least visible when one has drifted or been paused.
Watch the approval outcomes per language as well. A template approved in English and rejected in Hindi usually means the translation reads more promotionally than the original, which is a wording problem rather than a policy one and is fixable in a single edit.
Testing it properly
Take thirty real customer messages, keeping the original spelling and mixing intact, and run them through your bot. Do not clean them up; the typos and the script-mixing are the test.
| Check | What failure looks like |
|---|---|
| Replies in the same language | English answer to a Marathi question |
| Handles Roman-script Hindi | Treats it as misspelled English |
| Keeps policy wording exact | A friendlier, looser translation |
| Product names untranslated | Item names rendered into Hindi |
The last row catches people out. Product names, courier names and place names should stay as they are, and a model translating "Blue Dart" into Hindi produces something nobody can act on.
When you still need a person who speaks the language
Automation covers the routine well. It does not cover everything.
An upset customer writing in their own language usually wants a human who genuinely shares it, and an obviously translated apology can make the situation worse. If a meaningful share of your customers are in one region, having somebody on the team who speaks that language is worth more than any configuration.
Voice is harder still. Accent and dialect vary enormously within a single language, and an AI voice agent that handles Mumbai Hindi well may struggle with Bhojpuri-inflected speech.
Be honest about this in your design. Automate the routine in every language, and route anything emotional to a person who can speak it properly.
Finally, listen to how your own team writes to customers in each language. The bot should sound like them rather than like a translation service, and the fastest way to achieve that is to put a few of their real replies into the prompt as examples of tone.
Example conversation
Answer customers in their own language
Start free in minutes — no code, no credit card. Bring your own AI keys and pay no per-message markup.
Frequently asked questions
Do I need separate bots for each language?
No, and you should avoid it. Keep one knowledge base and let the model reply in the customer's language, because parallel translated copies drift apart and nobody notices until a customer quotes the wrong one.
How well do models handle Hinglish?
Generally well for everyday support, because it is common in training data. Test it on your own messages rather than assuming, since performance varies more on mixed script than on either language alone.
What about templates in regional languages?
Each language version is a separate template reviewed independently by Meta. Keep the number small and deliberate, because a stale template in a language nobody on your team reads is worse than not offering it.
Should prices and policies be translated?
Reproduce them exactly rather than paraphrasing. A translated refund policy that sounds more generous than the original becomes a dispute you will lose, because the customer has it in writing from you.
Does this work for voice calls too?
Partly. Speech recognition and synthesis vary much more by accent and dialect than text does, so test with real callers from your actual customer base rather than with clean studio recordings.
Related reading
Available across India — see WhatsApp & AI calling automation by state and city.