Use case · Multilingual support
Support in your customers’ languages, checked by people who speak them
Customers write in Hindi, Tamil, Arabic, English or a mix of them. The assistant detects the language and script, answers from knowledge written or reviewed for that language, keeps your product terms consistent with a locked glossary, and passes a reply to a human speaker when its confidence is low. Each language goes live only after it passes its own tests.
At a glance
- What it takes on
- Detect language and script
- Answer from per-language knowledge
- Lock the glossary
- Keep tone and register
- Connects to
- Your helpdesk or shared inbox
- Website chat and WhatsApp Business Platform
- Per-language knowledge base
- Glossary and style notes
- OpenAI, Anthropic Claude or Google Gemini
- Typical rollout
- Week 1 → Weeks 5–8See the plan
The problem
Where the work gets stuck today
Many support teams cover one or two languages well and the rest with a translation tool and hope. Replies come back grammatical but use the wrong word for your product, switch between formal and informal address, or change what a policy means. Customers who write in Hinglish or dialect get answers in a register nobody speaks.
Language models handle many languages, but unevenly: strong in some, weaker in others, and weaker again on your own terms. A system that treats every language the same either over-trusts the weak ones or forces everything through English.
| Problem | Why it happens | What the system does |
|---|---|---|
| Reply in the wrong language | Detection runs on a greeting or one word | Detects language and script on the whole message and keeps the conversation’s language |
| Mixed-language messages | Hinglish, or Arabic with English terms, confuses detection | Treats mixed input as its own case and replies in the customer’s main language |
| Product terms translated | Brand and plan names pass through translation | A locked glossary of do-not-translate terms and approved translations |
| Policy meaning drifts | Policies machine-translated on the fly | Policy text in each language reviewed by a speaker before use |
| Weak language, confident tone | One confidence threshold for every language | Thresholds per language; low confidence goes to a human speaker |
Task catalog
What the system takes on
Task 01
Detect language and script
Every message is classified by language and script, including Hindi and Arabic typed in Latin letters, with a confidence score. The conversation keeps the customer’s language unless they switch.
Task 02
Answer from per-language knowledge
Where you have content in a language, answers come from it. Where you do not, they come from the English source, translated with the glossary and held for review until a speaker approves the wording.
Task 03
Lock the glossary
Brand names, plan names and regulated terms are either never translated or translated one approved way. Every reply is checked against the glossary before it goes out.
Task 04
Keep tone and register
Formal or informal address, honorifics and greetings follow style notes you approve for each language, so the assistant never switches register halfway through a conversation.
Task 05
Route by confidence
Each language has its own threshold. Below it, the reply becomes a draft for a speaker, or the conversation moves to a person who speaks the language, with the transcript and a translation.
Task 06
Test each language before launch
Every language gets its own evaluation set of real questions, mixed-language and dialect messages included. A language is switched on only once it passes, and results are kept per language.
Step trace
One run, step by step
A customer asks why a refund has not arrived. Detected as English with high confidence, answered from the English refunds article with the glossary applied, and sent automatically.
- Detected: English, Latin script
- Knowledge: English refunds article
- Glossary: “store credit” as approved
- Route: automatic reply
The same question in Hindi, written in Devanagari. Answered from the Hindi refunds article a Hindi-speaking reviewer approved, with the brand name left untranslated.
- Detected: Hindi, Devanagari
- Knowledge: Hindi article, reviewed
- Glossary: brand name not translated
- Route: automatic reply
The question in Tamil. There is no Tamil refunds article yet, so the answer is translated from English with the glossary and held as a draft for the Tamil-speaking agent.
- Detected: Tamil, Tamil script
- Knowledge: English source, translated
- Glossary: approved Tamil term for store credit
- Route: draft for a Tamil speaker
The question in Arabic. Answered from the reviewed Arabic article, laid out right to left, with formal address as the Arabic style notes require.
- Detected: Arabic, Arabic script
- Knowledge: Arabic article, reviewed
- Style: formal address
- Route: automatic reply
Romanised Hindi mixed with English and an unclear order reference. Confidence falls below the Hindi threshold, so a Hindi-speaking agent takes over with a translation attached.
- Detected: Hindi and English, Latin script
- Confidence: below the Hindi threshold
- Glossary: not reached
- Route: a human speaker, transcript attached
Read the trace as text (Sample data)
- 00:00CustomerStepMessage in Tamil asking why a refund has not arrived.
- 00:00Language detectorDoneTamil, Tamil script, high confidence. Conversation language set to Tamil.
- 00:01RetrieverCheckNo Tamil refunds article yet; the English source is retrieved instead.
- 00:01TranslatorDoneAnswer translated with the glossary: brand name kept, store credit in its approved Tamil term.
- 00:02Policy checkApprovalTranslated policy wording is not speaker-reviewed yet, so the reply is held as a draft.
- 00:04Tamil-speaking agentDoneOne phrase corrected, then sent. The correction is saved for the Tamil article.
Four scripts
The same answer in four languages
EnglishRoute: Sent automatically
Your refund for order A-1042 was approved yesterday. It usually reaches your card within a few working days. If you chose store credit, it is already in your account.
HindiRoute: Sent automatically
ऑर्डर A-1042 का आपका रिफ़ंड कल मंज़ूर हो गया। यह आम तौर पर कुछ कामकाजी दिनों में आपके कार्ड तक पहुँच जाता है। अगर आपने स्टोर क्रेडिट चुना था, तो वह पहले से आपके खाते में है।
TamilRoute: Held for a Tamil speaker
உங்கள் ஆர்டர் A-1042-க்கான ரீஃபண்ட் நேற்று அங்கீகரிக்கப்பட்டது. அது பொதுவாக சில வேலை நாட்களுக்குள் உங்கள் கார்டுக்கு வந்து சேரும். நீங்கள் ஸ்டோர் கிரெடிட்டைத் தேர்ந்தெடுத்திருந்தால், அது ஏற்கனவே உங்கள் கணக்கில் சேர்க்கப்பட்டுவிட்டது.
ArabicRoute: Sent automatically
تمت الموافقة أمس على استرداد قيمة طلبك رقم A-1042. يصل المبلغ عادةً إلى بطاقتك خلال بضعة أيام عمل. وإذا اخترت رصيد المتجر، فهو متاح الآن في حسابك.
Glossary rules
Sample data| Term | Rule | Why |
|---|---|---|
| Brand and product names | Never translated, and written the same way in every language | Customers search for the name they know |
| Plan and feature names | Kept in English, with an approved explanation the first time they appear | They must match the app and the invoice |
| Refund, store credit, exchange | One approved translation per language | The policy’s meaning depends on the exact term |
| Legal and regulated wording | Quoted from the policy text reviewed in that language | A paraphrase can change an obligation |
Confidence routing
| Signal | What happens |
|---|---|
| High confidence, reviewed content in the language | The reply is sent automatically |
| High confidence, translated content not yet reviewed | A draft is held for a speaker of that language |
| Low confidence in the language or the answer | A person who speaks the language takes over, with transcript and translation |
| Complaint, legal, medical or financial topic | Always a person, whatever the confidence |
Systems involved
What it connects to
- Your helpdesk or shared inbox
- Conversations, language tags and handoffs to the agents who speak each language.
- Website chat and WhatsApp Business Platform
- The channels customers write in, on accounts in your name.
- Per-language knowledge base
- Articles written or reviewed in each language, versioned against the English source.
- Glossary and style notes
- Do-not-translate terms, approved translations and register rules, stored as data the checks read.
- OpenAI, Anthropic Claude or Google Gemini
- Answering and translation, with the model chosen per language from evaluation results.
Guardrails
- A language goes live only after it passes its own evaluation set, mixed-language and dialect messages included.
- Translated content that no speaker has reviewed is never sent automatically. It waits for a speaker of that language.
- Glossary terms are checked on every reply. A reply that breaks the glossary is held, not sent.
- Confidence thresholds are set per language, so a weaker language is never trusted like a strong one.
- Complaints and legal, medical or financial topics go to a person in every language.
- We do not promise native-level fluency. Quality is measured per language and reported before and after launch.
Human checkpoint
For each live language, a named reviewer who speaks it approves the knowledge content, glossary entries and style notes, and works the queue of held drafts. Agents who speak the language receive low-confidence conversations with the transcript, a machine translation for context and the assistant’s draft. Their corrections are saved and reviewed weekly, and recurring fixes go into the glossary or the language’s articles.
What we plan for
Typical targets
Typical target
95–100% in live languages
Messages answered in the customer’s language
Basis: Detection is checked on each language’s evaluation set before launch; messages in languages not yet live get an English reply and the offer of a person
Typical target
None, by design
Replies sent that break the glossary
Basis: Every reply is checked against the glossary before sending, and a reply that fails is held for review
Typical target
10–30% in its first month
Conversations handed to a speaker, new language
Basis: Set by that language’s threshold and tuned from reviewed drafts; it falls as reviewed articles replace translated answers
Rollout
Week by week
Week 1
Language mix measured from your ticket history; glossary, style notes and the first languages chosen, with a reviewer named for each.
Weeks 2–3
Detection, per-language retrieval and glossary checks built, plus an evaluation set per language from real, anonymised messages.
Week 4
English and the first additional language live, with cautious thresholds so more drafts go to reviewers; thresholds tuned from their edits.
Weeks 5–8
Further languages added one at a time, each after passing its tests, while reviewed articles replace translated answers.
FAQ
Frequently asked questions
Do we need to translate our whole help centre before launch?
No. The assistant can answer from your English content with the glossary and hold those replies for a speaker to check. Reviewed replies become the start of each language’s own articles, so translation effort follows the questions customers really ask.
How do you decide which languages to switch on first?
From your own data: the languages in your ticket history, the volume in each, and whether someone who speaks it can review. A language without a reviewer can still be detected and routed to a person, but it does not get automatic replies.
Can customers type Hindi or Arabic in English letters?
Yes. Romanised Hindi and Arabic in Latin letters are common, so detection checks script and vocabulary separately. Replies usually follow the customer’s choice of script, and your style notes decide when to answer in the native script instead.
Will the assistant translate our legal terms and policies?
Not freely. Legal and policy wording comes from a version reviewed in that language and is quoted, not paraphrased. Where no reviewed version exists, the assistant gives the English text with a short explanation and offers a person.
What does each extra language add to running costs?
Mostly model usage for translation, since some scripts take more tokens per sentence, plus reviewer time in the first weeks. The proposal estimates it per language from your volumes, and it falls as reviewed articles replace live translation.
How is answer quality measured for each language?
Each language has its own evaluation set: real questions, mixed-language messages, dialect variants and glossary traps. A speaker grades accuracy, tone and terms, and results are kept per language, so a strong English score never hides a weak Tamil one.
What happens when a conversation moves between languages?
The assistant follows the customer and the handoff notes record both languages. An agent who takes over sees the original messages with a translation, and can reply directly or have the assistant translate their draft for review.
Next step
Have a workflow in mind? Let’s scope it together.
Book a free 30-minute strategy call. Bring one process you want to automate and we will talk through scope, integrations and a sensible first step.
Free 30-minute call · no commitment