📋 Playbooks

Rolling out multilingual support: start with two languages, not twelve

Updated 2026-09-19 · 4 min read · By the Convo360 team

In short

Roll out multilingual support by picking the two languages your existing customer data shows the most demand for, not the two you assume. Add a short language menu at the start of the conversation, translate templates and the AI agent's knowledge base for those languages only, then expand once quality is confirmed by native speakers.

Key takeaways

Use actual customer data — call recordings, support tickets, pincode spread — to pick your first two languages, not assumptions.
A one-question language menu at the start of the chat avoids guessing wrong and frustrating customers.
Every WhatsApp template needs separate Meta approval per language — plan for this in your rollout timeline.
Voice language availability varies by provider; check before promising a language for calls, not just chat.
Have a native speaker review AI answers before launch — translation tools alone miss tone and local phrasing.

Why twelve languages on day one fails

It's tempting to switch on every supported language at once. Convo360 covers Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Odia and Punjabi alongside English. But launching all of them together means you can't quality-check any of them properly, and you're translating templates, knowledge base articles and voice scripts ten times over before you've validated the first one.

Two languages, done well, serve most of a regional business's customers and leave room to learn what actually needs fixing before scaling further. A staged rollout also spreads out the Meta template approval workload and the native-speaker review effort, both of which get harder to manage in parallel across many languages at once.

Picking your first two languages with data

Don't guess based on where your office is. Look at what you already have:

Call recordings or support tickets: which language do customers switch to when a query gets detailed?
Customer pincodes or delivery addresses: cross-reference with state language data.
Existing WhatsApp conversations: note which customers reply in Hindi or a regional language even when your templates are in English.

Pick the top two by volume. A Mumbai-based D2C brand often finds Hindi and Marathi outrank the rest; a Chennai-based service business usually finds Tamil dominant, with Hindi second for pan-India orders. A useful way to lay this out is a simple rollout calendar:

Phase	Timeframe	Focus
Phase 1	Weeks 1–2	Data pull, pick top 2 languages, draft templates
Phase 2	Weeks 3–4	Submit templates for Meta approval, translate knowledge base
Phase 3	Weeks 5–6	Native-speaker QA, Preview testing, soft launch
Phase 4	Weeks 7+	Monitor Flow Logs, then evaluate a third language

At the start of a new conversation, send a short Buttons or List Menu block: 'Choose your language / अपनी भाषा चुनें / உங்கள் மொழியைத் தேர்ந்தெடுக்கவும்'.
Use Set Variable to store the choice as a system variable for the rest of the conversation.
Branch the flow with a Condition block so templates, AI agent responses and voice scripts all reference that variable.
For returning customers, remember the last language chosen using a custom field so they aren't asked again.
For voice, ask the same question at the start of the call using a Gather or DTMF menu block, since a caller's chat language preference doesn't always carry over to how they want to speak.

See /docs/flow-builder/ for building branching menus.

Templates need separate approval per language

Every WhatsApp template is submitted to Meta per language: a Hindi version of your order-confirmation template is a separate submission, not an automatic translation of the English one. Budget extra time in your rollout for this, since approval can take from a few hours to a couple of days per template. Group your template list and submit all languages together to avoid a staggered, half-translated launch, where some customers get Hindi order updates and others are stuck waiting on an English-only template that hasn't been recreated yet.

Keep a simple tracking sheet of template name, language and approval status. This becomes essential once you're managing two languages across five or six templates each, since it's easy to lose track of which version is still pending.

Voice language availability

Chat and voice don't always support the same languages equally. Convo360's voice stack uses providers like Sarvam (strong across Indian languages), Smallest.ai and ElevenLabs. Availability varies by provider and region, so confirm which languages your chosen telephony and voice combination actually supports for calls before promising it to customers. Chat-only languages can still launch on WhatsApp even if voice isn't ready yet.

If a language you need isn't available on your first-choice voice provider, check whether a different provider covers it before dropping the language from your voice rollout entirely. Convo360 lets you choose the voice provider per agent, so chat and voice don't have to run on the same underlying provider.

QA with native speakers before launch

Machine translation gets grammar right and tone wrong. Before going live, have a native speaker read through the AI agent's knowledge base answers, the WhatsApp templates and a handful of test conversations. Look specifically for phrases that are technically correct but sound stiff or overly formal. Hinglish and natural code-mixing often reads better to customers than textbook-correct regional language.

Run test conversations using Preview / Try-it in each language before publishing, and only expand to a third language once the first two are stable. A simple QA checklist helps: does the greeting sound natural when read aloud, do numbers and currency read correctly, and does the agent switch language mid-conversation only when asked, not accidentally.

Measuring the rollout

Once live, track conversation volume by language in Analytics, alongside the AI agent's fallback rate, how often it says it doesn't understand, separately for each language. A high fallback rate in a new language usually means the knowledge base translation missed some content, not that customers are asking unusual questions.

Also compare resolution rate and reply-to-human handoff rate between your original language and the new ones. If a new language shows a noticeably higher handoff rate after a few weeks, that's a clearer signal to invest more QA time before adding a third language than any fixed timeline would be.

Step by step

Pull language signal from existing dataReview call recordings, tickets and customer locations to identify your top two languages by real demand.
Build the language selection menuAdd a Buttons or List Menu block at conversation start and store the choice in a variable.
Translate and submit templatesCreate language-specific versions of each WhatsApp template and submit all for Meta approval together.
Confirm voice availabilityCheck which languages your chosen voice provider supports for calls before enabling voice in that language.
QA with a native speakerHave a fluent speaker review AI answers and templates for natural tone before launch.
Launch and monitor before adding a third languageWatch Flow Logs and conversation quality for two to four weeks before expanding further.

Example conversation

मुझे हिंदी में जवाब चाहिए

जी बिलकुल! मैं अब आपसे हिंदी में बात करूंगी। आपका ऑर्डर किस बारे में है?

Plan your multilingual rollout

Start free in minutes — no code, no credit card. Bring your own AI keys and pay no per-message markup.

Start 14-day free trial →See pricing

Frequently asked questions

How do I know which two languages to start with?

Look at your existing support tickets, call recordings and customer locations rather than guessing. The languages your customers already switch to when explaining details are your best starting point.

Does adding a language mean redoing all my templates?

Yes, each template needs a language-specific version submitted to Meta for approval, since Meta treats each language as a distinct template, not an automatic translation. Plan this as part of your rollout timeline, not an afterthought, and submit all languages together to avoid a staggered launch.

Can chat support more languages than voice?

Yes. Voice language availability depends on your chosen voice provider, such as Sarvam, Smallest.ai or ElevenLabs, and can lag behind chat, which uses the underlying LLM directly. Always confirm availability with your provider before promising a language for phone calls.

Is Hinglish or code-mixed language supported?

Yes. Natural code-mixing works for both chat and voice because the underlying LLM handles it directly, and it often sounds more authentic to customers than strictly formal regional-language text, so don't force your AI agent's replies into textbook phrasing.

How long before adding a third language?

Give the first two languages two to four weeks of live conversations and review Flow Logs for confusion points before expanding further. If fallback and handoff rates in both languages look stable by then, you have a reasonable baseline to plan the next rollout against.

What's the fastest way to tell if a language rollout needs more QA?

Watch the AI agent's fallback rate and human-handoff rate for that language in Analytics. A rate noticeably higher than your original language usually means the knowledge base translation needs another pass.