Running a WhatsApp Chatbot in Hindi and Regional Languages

Open any Indian business's WhatsApp inbox and the English is outnumbered. "Price kya hai", "क्या ये available है", "delivery kitne din me", Tamil typed in Roman letters, Marathi in Devanagari. A chatbot built and tested only in English will fail a large share of the people it was bought to serve.

This guide covers what it actually takes to run a multilingual WhatsApp chatbot in India — the parts Meta handles, the parts you must handle, and the testing that catches problems before customers do.

Three kinds of language input you will receive

• Native script — Devanagari, Tamil, Telugu, Bengali, Gujarati. Common among older customers and in tier-2 and tier-3 cities, and increasingly common everywhere thanks to voice typing.
• Transliterated — an Indian language typed in Roman letters: "kitna hoga", "ille", "hoy". This is the largest category by far and the one most chatbots handle worst.
• Mixed — "delivery kab tak hoga for Pune?" Two languages in one sentence, which defeats any system that first tries to detect a single language.

Design for the second and third. A bot that handles only clean Devanagari and clean English covers the minority of real messages.

Matching a language: menus versus free text

Button-based bots

The most reliable multilingual design is also the simplest: start the conversation with a language choice — English, हिंदी, मराठी — as buttons, store the answer on the contact record, and run the rest of the flow in that language. No detection, no ambiguity, and it works identically for every customer. The cost is one extra tap, and the maintenance cost of every flow existing in each language.

Keep the language list short. Three options covering most of your customers beats eight that nobody maintains. Stale Tamil translations are worse than honest English.

AI-based bots

A language model reads whatever the customer typed and replies in the same language, including transliterated Hinglish, without any menu. This is where AI genuinely outperforms a scripted bot for Indian traffic. Two cautions: instruct it explicitly to reply in the customer's language and script (models otherwise drift into English), and test regional-language accuracy yourself, because quality varies noticeably between Hindi and, say, Odia. Our guide to AI agents on WhatsApp covers the guardrails.

Templates: the part people get wrong

Message templates are approved per language. This means several concrete things worth knowing before you plan a campaign:

• You create one template name with a version for each language, and each version is reviewed separately. A perfect English template does not guarantee the Hindi one is approved.
• Rejections in regional languages are often about translation quality rather than policy — a machine-translated Marathi template reading awkwardly gets refused where the English original passed. Have a native speaker check before submission, not after.
• At send time your platform picks the version matching the contact's stored language, so the language field on the contact record must be populated. Contacts with no language set fall back to your default — usually English — which quietly undoes the whole exercise.
• Variables move. Word order differs between English and Hindi, so a placeholder sitting neatly at the end of an English sentence may belong in the middle of the Hindi one. Translate the sentence, do not translate word by word around fixed placeholders.

See our template approval guide for the general rules; everything there applies per language version.

Practical traps in Indian-language messaging

Character limits. Template and message limits are counted in characters, and Indian scripts consume them faster once conjuncts and matras are included. A message that fits comfortably in English can overflow in Hindi. Check the rendered length of every translated template.

Old handsets. Rendering of complex conjuncts is still imperfect on some budget devices. Test your key messages on an inexpensive phone, not only on the team's flagship handsets.

Numbers and dates. Keep digits in Latin numerals — customers read them faster, and they avoid rendering problems. Write dates unambiguously, and prefer "15 सितंबर" over a numeric format that can be read two ways.

Formal register. Machine translation defaults to a formal, textbook Hindi that sounds like a government notice. Customers respond better to the register your own staff use on the phone. Have a salesperson, not only a translator, read the drafts aloud.

Brand and product names. Leave them in English. Transliterating a product name breaks search, confuses returning customers and looks amateurish.

Keyword matching for transliterated text

If you run a scripted bot rather than an AI one, your keyword lists must cover the spellings people actually type. "Price" arrives as price, prize, kimat, kimmat, keemat, daam, rate, kitna, kitne. Build each intent's keyword list from your own inbox history, not from a dictionary: export the last few hundred inbound messages, sort them by intent, and lift the real spellings. Repeat every quarter, because the "unmatched message" log is the cheapest product research you will ever get.

Testing before launch

A short checklist that catches most defects:

1. Send ten real customer messages from your inbox history through the bot in each language and score the replies.
2. Test at least one mixed-language sentence per flow.
3. Open every template on a low-cost Android handset and check rendering and truncation.
4. Confirm that a contact with no language set receives a sensible default rather than nothing.
5. Have a native speaker read every outbound message aloud, and change anything that sounds like a form letter.
6. Check the handover message: when the bot gives up, the human it hands to should be able to read that language. Routing a Tamil conversation to an agent who cannot reply is worse than not offering Tamil.

How many languages to offer

Start with English plus the one language most of your customers use, and add a third only when the first two are genuinely well maintained. Every extra language multiplies templates, flows, keyword lists, testing and staffing. A business doing two languages well converts better than one doing six badly — and unlike a website, a chatbot in a language nobody on your team can speak creates a promise you cannot keep.

Frequently Asked Questions

Can a WhatsApp chatbot understand Hinglish typed in Roman letters?

A scripted bot only if your keyword lists include the real spellings customers use, which you should lift from your own inbox history. An AI-based bot handles transliterated Hinglish natively and is the better choice when most of your traffic is free text.

Do I need separate WhatsApp templates for each language?

Yes. Templates are approved per language version, and each is reviewed separately. The contact's stored language decides which version is sent, so contacts without a language set will fall back to your default.

Why was my Hindi template rejected when the English one was approved?

Usually translation quality rather than policy — machine-translated text that reads awkwardly, wrong register, or placeholders left in an unnatural position. Have a native speaker rewrite the sentence rather than translating word by word.

How many languages should I support?

Start with English plus the single language most of your customers use. Add a third only once the first two are well maintained, since every language multiplies templates, flows, testing and the need for agents who can actually reply in it.

Should numbers and product names be translated?

No. Keep digits in Latin numerals and leave brand and product names in English. Transliterating them slows customers down, breaks search and looks unprofessional.