Chatbot Localization
Six Steps From One Language to a Market, With the Engine Choice Made Per Surface
Quick answer: Turning a working chatbot into a multilingual chatbot is not one translation job; it is about nine small ones, and they want different tools. A greeting and a menu are short, stable and served thousands of times: machine-translate them once, have a native speaker edit them once. A knowledge-base answer changes weekly and should never be translated at build time; either the model answers in the customer's language directly (Tidio's Lyro, 48 languages) or the platform translates the reply at answer time from your one-language content (Intercom's real-time translation for Fin). A WhatsApp template is approved per language, so its translation is a compliance step, and the fallback message is the first line to have a human write. On prices read 5 September 2026, the engine is the cheap part: $40 to $50 a month for 1,000 translated conversations on Google Cloud's per-character meters, and nothing extra on Intercom or Tidio in the help articles we read. What decides whether the bot feels local is a native speaker's afternoon on the top flows and highest-traffic messages and a fifteen-question test in every language before launch, and this guide schedules both.
Multilingual chatbots: 7 truths makes the argument that translation is not localization and that understanding is per-language. This guide assumes you accept the argument and need the procedure: which text to translate, with which engine, in what order, and how to know when it is done. The engine's mechanics and its meters are in the machine translation entry; which language a message is in is the language detection entry. Neither is repeated here beyond what a step needs.
Step 1. Choose languages from where demand is, not from the vendor's count
The vendor's language count is the least useful number in the decision. Tidio's help center lists 48 languages Lyro will answer in; Intercom's Fin list has 65 entries once regional variants are counted (twelve of them English, the base plus eleven regional). Neither tells you which two languages your customers write in. Three sources do. Your own transcripts and tickets, tagged by detected language for a month, are the first. Your web analytics by country and browser language are the second. Search demand by country is the third, and it is the one that surprises: the head term for this guide, "multilingual chatbot," has 78 percent of its global search volume outside the United States by Ahrefs' count on 5 September 2026, and our reviews have repeatedly found brand demand outside the US several times the US figure for platforms that market in English: Manychat's review records 215,000 monthly brand searches across Latin America against 70,000 in the US, and Typebot's records a Brazilian brand volume 32 times its US one.
Two rules come out of that. First, a language is not a market. Intercom's list separates Spanish from Spanish (USA), Portuguese from Brazilian Portuguese, French from French (Canada), and German from German (Formal), and register and vocabulary are the reason such splits exist; if your customers are in Mexico and Madrid, you have two localizations that share an engine. Second, the list should be short enough to test. Step 6 asks for fifteen questions and a native-speaking reviewer per language; if you cannot staff that for a language, do not switch it on. Tidio's own help center makes the same point from the other side: one of the two examples Tidio gives for restricting Lyro to "Specific languages" is that "your live agents can only speak some of Lyro's supported languages."
Step 2. Inventory the nine surfaces a customer can meet
Walk the bot as a customer and write down every piece of text that reaches the screen. Sorted by how often it changes and how often it is served, it comes to nine kinds, and the reason to sort them is that the next step assigns each a different engine.
| Surface | Example | Changes how often | Served how often | Who should own the final wording |
|---|---|---|---|---|
| 1. Greeting and opening question | "Hi, what can I help with today?" | Rarely | Every conversation | Native speaker |
| 2. Menus, buttons, quick replies | "Track my order" / "Talk to a person" | Rarely | Most conversations | Native speaker, with length checked on a phone |
| 3. Scripted flow messages | Order-status steps, booking steps | Monthly | Often | Machine first, native edit |
| 4. Knowledge-base answers | FAQ, help articles, policy text | Weekly | Per question | Nobody translates these; see step 3 |
| 5. Fallback and error messages | "I didn't catch that", "we're closed" | Rarely | On every failure | Native speaker, first |
| 6. Business-initiated templates | WhatsApp order updates, reminders | Per campaign | In bulk | Native speaker, then Meta's reviewer |
| 7. Forms and attribute collectors | "What's your order number?" | Rarely | Often | Machine first, native edit; check formats (step 4) |
| 8. Widget and channel UI strings | "Send", "Attach", "We reply in minutes" | Never | Always | The platform's own localization |
| 9. Agent-side text | Canned replies, notes, inbox view | Weekly | Per handoff | The agent, with inbox translation as a reading aid |
Surface 8 is worth checking before you buy rather than after: it is the platform's own interface language, and our reviews score it under localization. SendPulse ships its admin interface in five languages, one of the three localized layers that put it first on our multilingual list; Tidio's review records seven UI languages; Intercom's records twelve. Surface 6 is the one with a regulator: a WhatsApp template has a language code, the same name in a second language "counts as a second template," and each is approved separately, so a translation error there is a rejected submission, not a typo.
Step 3. Assign an engine pattern to each surface
Intercom's help center is unusually clear that its translation features are separate machines: its AI Inbox Translation "(what teammates see)," Fin's multilingual support "(what customers receive from Fin)" and Multilingual workflows "(what customers see in automated workflow messages)" are, in its words, "separate features. They must be configured independently, and disabling one has no effect on the others." Two of those, workflow Auto-translate and Fin's real-time translation, put text in front of customers; inbox translation is an agent's reading aid and returns in step 5. Add Tidio's model, which "works natively in the customer's language instead of translating messages to English (internally) first," and the hand-authored alternative that needs no engine, and you have the four customer-facing patterns below.
| Pattern | What it does | Fits surfaces | Worked example (read 5 September 2026) | The trap |
|---|---|---|---|---|
| A. Translate once at build time, then edit | Machine-translates static copy into every language when you save; a native speaker corrects it | 1, 2, 3, 5, 7 (on Intercom, attribute collectors "may not be translated" by Auto-translate, so plan surface 7's prompts as pattern D) | Intercom Multilingual workflows: Auto-translate covers "chat messages and reply buttons," "available on all channels"; attribute collectors "may not be translated"; Custom Answers not supported | The engine translates from the workspace default language at the time you save; a workflow written in Spanish under an English default "treats the Spanish text as the English 'original'" and is mistranslated everywhere |
| B. Answer natively, no translation | The model reads the question and the source content in whatever languages they are in and writes the reply in the customer's | 4 | Tidio Lyro: 48 languages; data sources "processed directly in their original language" | Register drifts unless instructed (Lyro Guidance can be told "informal German"); the ceiling is the model's competence in that language, which is per-language and worth testing (step 6) |
| C. Translate the reply at answer time | Content stays in one language; when there is none in the customer's language, the platform translates the answer it generated | 4 | Intercom Fin Real-time translation: Fin "will first search for the answer using content that is available in the language a question was written in," and otherwise translates from "your chosen fallback language"; the source "will remain in its original language" | Only the configured default language is used as a fallback; content in a third language is ignored |
| D. Separate flow per language | You author each language by hand and route by detected language | 1, 2, 3, 5, 6, 7 when register, formats or law differ by market | Intercom's detection-article FAQ: to control the first message's language, "set up separate workflows for each language and use the 'Detected language' attribute"; turn Auto-translate off on those or they are re-translated | Five times the maintenance; the flows drift apart within a quarter unless one person owns all of them |
The decision rule is the change rate from the step 2 table. Text that changes rarely and is served constantly (surfaces 1, 2, 5, and the flow steps and form prompts of 3 and 7) gets pattern A plus a human edit, or pattern D if the markets differ enough that a shared source sentence is wrong for one of them; surface 8 is the platform's own localization and is checked before purchase. Text that changes weekly (surface 4) gets B or C and is never translated at build time; translating a knowledge base into six languages gives you six copies to keep in sync and six chances for retrieval to pick the wrong one. Surface 6 is always D in effect, because Meta approves each language on its own. Surface 9 is a reading aid for the agent and should not be mistaken for a customer-facing translation; Intercom's article states it flatly: "What you see in the Inbox is not always what the customer received."
Step 4. Adapt the things no engine adapts
A correct translation can still be the wrong text for a market. This is the short list we check per language, in the order that customers notice.
Register. Spanish tú against usted, German du against Sie, Brazilian você against Portuguese tu: the engine picks one, and picks it per sentence. Decide per market, write it into the instruction (Tidio documents telling Lyro to "use informal German, as opposed to formal German"; Intercom separates German (Formal) as its own language) and have the reviewer in step 6 flag every slip.
Formats. Dates (5/9 is May in the US and September almost everywhere else), decimal separators, currency symbols and their position, phone-number patterns, and address order. Surface 7, the forms, is where these bite: a collector that validates a US ZIP code rejects every Brazilian CEP. Write the validation per market, not per language.
Names and nouns. Product names, plan names, and button labels that are also brand terms should not be translated, and engines translate them. Google's API lists glossary creation as a free operation on its pricing page, and most engines have some do-not-translate mechanism; an LLM answering natively needs the rule in its instructions. Add the list to whatever the platform calls its guidance and test it.
Money and law. Local payment methods and their names (Pix, Boleto, OXXO, iDEAL) belong in the flow, not in a translation; a consent line for marketing messages has a required wording in some jurisdictions and a required opt-in mechanism on WhatsApp regardless. Our GDPR chatbot compliance guide covers the European case; the point here is that the localized consent text is a legal review, not a translation.
Length. Translated German runs longer than its English source (localization vendors quote a wide range; we have not measured it), and quick-reply buttons have character limits on every channel. Read every button on a phone in every language, once.
Step 5. Wire detection, fallback and the human
Localization fails at the seams: the first message, the unsupported language, and the handoff.
Decide the language once, from the strongest signal. Intercom's detection article (13 May 2026) is the most explicit order in the five articles read for this guide and a reasonable default for any platform: a per-customer language override first, then AI detection "if high enough confidence" (Fin needs "at least 10 characters from customer messages" and ignores button clicks), then the first browser language that is enabled in your workspace, then the workspace default. Once set, "Fin and other workflows will continue in that language for the remainder of the conversation." A bot that re-detects on every message flips on a borrowed English word. When the first message is too short to detect, a one-tap language choice beats a guess; the language detection entry covers that case in detail.
Write the unsupported-language message in the languages you do support. Tidio's behavior is the model: if a customer writes in a language not on your list, "Lyro will clearly say it doesn't support that language, and will use the default." That sentence is surface 5, and it should offer the handoff rather than end the conversation, because the customer who wrote in Turkish is still a customer.
Route the handoff to a speaker, not to a translation. Inbox translation lets an English-speaking agent read Portuguese; Intercom gives each teammate "10 free translated conversations per month" before a Copilot seat is needed. It is a good reading aid and a poor substitute for a Portuguese-speaking agent when the conversation is about a refund. If you have agents in the language, route by detected language; if you do not, say so in the fallback text and set the expectation on reply time. Our handoff design guide covers the queue mechanics.
Keep the fallback intent honest in every language. It is the message customers see most when things go wrong, and it is the one to write by hand first, because a machine-translated "I'm sorry, I didn't understand that" in stiff textbook register is the moment the bot stops feeling local.
Step 6. Test fifteen questions per language, then price the month
Before a language goes live, run the same short test in it that you ran in the first language. Our Tier 1 reviews use a 15-question fact-extraction set against a single support page; in the Tidio review that produced 87 percent (13 of 15) grounded answers in English and Polish and 80 percent (12 of 15) in Mexican Spanish and Brazilian Portuguese, with the failures "clustering on idiomatic phrasings rather than semantic content." The acceptance rule our multilingual list uses is that a non-English language should land within ten points of English on the same script; below that, the language is not ready, whatever the vendor's count says. Log the failures by surface from step 2 so that a wording problem on a button is not mistaken for a retrieval problem in the knowledge base.
Then the native-speaker pass the 7 truths guide calls the cheapest localization test available (one reviewer, one afternoon, the top three flows), applied here to the twenty highest-traffic messages the same guide singles out (surfaces 1, 2, 3, 5, 6 and 7 cover them), marking everything that reads translated and every register slip from step 4. Fix what they flag, greeting and fallback first.
The bill, for a 1,000-conversation month in one added language, on public prices read 5 September 2026 and our own assumptions (eight customer messages of about 60 characters and eight bot replies of about 250 characters per conversation, 2.48 million characters in the month):
| Line | Pattern | Cost | Basis |
|---|---|---|---|
| Translation engine, if you bring your own | A or C on a per-character API | $39.60 to $49.60 | Google Cloud Translation NMT: (2,480,000 − 500,000 free) × $20 per million = $39.60; its Translation LLM at $10 per million in and out is $49.60, or the same $39.60 if the monthly credit applies to it, which the page does not say |
| Translation engine, on the platform | C (Intercom) or B (Tidio) | $0 as a separate line | None of the four Intercom help articles read for this guide lists a charge for real-time translation or workflow Auto-translate (Intercom's pricing page was not read for it); a Fin answer costs $0.99 per outcome in our Intercom review whether or not it was translated; Lyro is metered by conversation allowance on Tidio's plans, in any of 48 languages |
| Native-speaker review | All | One afternoon per language, at whatever a reviewer in that language costs you | Illustrative only: four hours at $40 is $160; the figure is yours to replace, and it recurs whenever surfaces 1, 2, 3, 5 or 7 change |
| Per-language test | All | About two hours of your own time per language | Fifteen questions, logged by surface |
| Template resubmission | Surface 6 | $0 in fees; days in calendar time | Each WhatsApp template language is reviewed separately by Meta |
The engine is the smallest line. The recurring one is the review, and the way to keep it small is to keep the machine-translated surfaces (1, 2, 3, 5 and 7) short and stable, so that a change in the knowledge base, which is answered natively or translated at answer time, never triggers a re-review.
What our fifteen reviews record about localization
All fifteen reviews score a localization dimension and none has measured a translation feature's output as distinct from a model's native answers; the search strings and counts are printed in the machine translation entry's sources rather than repeated here. What the corpus adds to this guide is the pass mark. Three platforms on our multilingual list (Botpress, Tidio, Wati) passed both of that list's tests, measured accuracy within ten points of English in three or more languages and a verified localized interface in five or more, while AiSensy was left off because, as verified at review time, its agent understood 19-plus languages but replied in English only. This guide names no best platform for localization; it names the settings to find in whichever one you have.
FAQ
What is the difference between translating a chatbot and localizing it?
Translation converts the words; localization makes the bot right for a market, which also means register (tú or usted), date and currency formats, form validation, local payment methods, legally required consent wording, product names left untranslated and a fallback message a native speaker wrote. Steps 4 and 5 are the localization work that no translation engine does.
Should I machine-translate my chatbot?
For the short, stable surfaces (greeting, menus, flow steps, form prompts, fallback), yes, once, followed by a native speaker's edit; for knowledge-base answers, no. Let the model answer in the customer's language (Tidio's approach) or let the platform translate the reply at answer time from your one-language content (Intercom's), so that the source stays in one place. The machine translation entry has the meters.
How do I localize a chatbot for WhatsApp?
Template messages are the extra step: each language is a separate template with its own language code, submitted and approved separately by Meta, so translate them once, have a native speaker check them, and budget calendar days for review. Session replies inside the 24-hour window follow the same rules as any other channel. Our WhatsApp template entry and WhatsApp automation guide cover the mechanics.
How many languages should my chatbot support?
As many as you can test and staff. Each language needs a fifteen-question test, a native reviewer's afternoon and, ideally, an agent who speaks it for handoffs. Pick from your transcripts, analytics and search demand by country rather than from the vendor's list, and restrict the platform to that set so it says "not supported" in a language you do serve rather than guessing.
How do I test a chatbot in a language I don't speak?
Run the same fifteen-question script you used in your first language, translated by a native speaker, and score grounded answers the same way; then have that person read the twenty most-served messages and mark anything that sounds translated. The acceptance rule we use on our multilingual list is within ten points of the English result on the same script.
Related guides
- Multilingual chatbots: 7 truths — the argument this guide assumes: translation is not localization, understanding is per-language, every channel adds a layer.
- Best multilingual chatbot platforms — the ranking by measured language coverage, and the ten-point acceptance rule used in step 6.
- Machine translation (glossary) — the engine, the four insertion points and the meters behind step 3 and the bill.
- Chatbot language detection (glossary) — why the first message is the hard case, and the signal order step 5 adopts.
- Chatbot handoff design guide — the queue mechanics behind routing a handoff to a speaker.
- GDPR chatbot compliance guide — the consent wording that is a legal review, not a translation.
- WhatsApp message template (glossary) — the per-language approval that makes surface 6 a compliance step.
- Chatbot QA testing protocol — the test discipline step 6 applies per language.
Sources
- Intercom Help, Understanding translation behavior in Intercom (article 16296591, written by Beth-Ann Sher, "Updated over 3 weeks ago" on the read date), read 5 September 2026: the three-feature table and the sentence that they "are separate features. They must be configured independently, and disabling one has no effect on the others"; "What you see in the Inbox is not always what the customer received"; "Each teammate gets 10 free translated conversations per month"; "Intercom uses AI and large language model (LLM) providers for translations."
- Intercom Help, Set up Fin AI Agent's multilingual support (article 8322387, written by Beth-Ann Sher), read 5 September 2026: the supported-languages list (65 entries by our count: twelve English entries, the base plus eleven regional; four French; three German; plus Dutch (Belgium), Spanish (USA), Portuguese and Brazilian Portuguese); "Fin will first search for the answer using content that is available in the language a question was written in"; "when Fin can't find relevant support content in the customer's language, it will translate existing support content in your chosen fallback language"; "The original source content that Fin references or pulls from will remain in its original language"; "Content in other languages will not be used as a fallback."
- Intercom Help, Multilingual workflows (article 10522253, written by Beth-Ann Sher, dated 30 July 2026), read 5 September 2026: "Currently supported: Chat messages and reply buttons. Translations are available on all channels"; attribute collector and Custom Answers not supported; the two troubleshooting cases (Spanish content under an English default "treats the Spanish text as the English 'original'"; a French workflow with Auto-translate on is re-translated, and the fix is to "Turn off Auto-translate on any workflow that is already authored in its intended delivery language," which "removes all previously generated language variations").
- Intercom Help, Automatic language detection in conversations (article 9423767, written by Paula Lopez, dated 13 May 2026), read 5 September 2026: the 10-character minimum, button clicks ignored, once-per-conversation detection ("Fin and other workflows will continue in that language for the remainder of the conversation"), the order (language override, AI detection, first enabled browser language, workspace default), and the FAQ answer recommending "separate workflows for each language" with the "Detected language" attribute.
- Tidio Help Center, Lyro Multilanguage support (article 12643043611036, dated 31 October 2025), read 5 September 2026: "forty-eight (48) supported languages"; "works natively in the customer's language instead of translating messages to English (internally) first"; data sources "processed directly in their original language"; "Lyro will clearly say it doesn't support that language, and will use the default"; the Specific languages option and its rationale ("your live agents can only speak some of Lyro's supported languages"); the Lyro Guidance "informal German, as opposed to formal German" example.
- Google Cloud, Cloud Translation pricing (cloud.google.com/products/translate/pricing), read 5 September 2026: NMT text translation "First 500,000 characters per month — Free (applied as $10 credit every month)" and "Over 500,000 characters per month — $20.00"; Translation LLM $10.00 per million input and $10.00 per million output. The $39.60 figure is (2,480,000 − 500,000) × $20 per million.
- Chatbotscape, WhatsApp message template (/glossary/whatsapp-message-template): a template "has a name … a language code, a category," and "the same name in a second language counts as a second template." Meta's own template documentation was not re-read for this guide.
- Chatbotscape review corpus (the fifteen platform reviews listed at /reviews), read 5 September 2026; the four corpus search strings and their counts (machine translation 0, auto-translate 0, localization 15, multilingual 7) are printed in the sources of /glossary/machine-translation. Passages cited here:
tidio-review.mdline 692 (four-language results and the "idiomatic phrasings" sentence) and line 810 (seven UI languages);intercom-review.mdlines 69–70 (12-language interface);manychat-review.mdline 479 (215,000 LATAM against 70,000 US brand searches);typebot-review.mdline 731 (Brazilian brand volume 32× US);best/best-multilingual-chatbot.mdlines 68–72 (the two tests, the three platforms that passed both, the AiSensy exclusion) and the SendPulse rank card (five admin languages, three localized layers). - Ahrefs Keywords Explorer, US overview, queried 5 September 2026: 'multilingual chatbot' 100/mo US, 450 global; the other figures in this guide's keyword note.
- Chatbotscape evaluation methodology. /methodology (continuously updated).
About this guide
Chatbotscape launched in 2026 as an independent review site for chatbot platforms. This guide is part of our SMB chatbot Academy and is written for the owner or support lead of a small business adding languages to a working chatbot. It reads Intercom's and Tidio's documentation as worked examples of the four engine patterns because they document them in the most detail.
Methodology
The four Intercom articles, the Tidio article and the Google Cloud pricing page were read in full on the day of drafting, and every quotation carries its article. The corpus figures are our own published measurements, with the evidence class of each stated in Sources, and every corpus search is printed with its string so that it can be reproduced. The nine-surface inventory, the four engine patterns, the decision rule by change rate and the per-language test protocol are editorial constructions, drawn from the vendor documentation cited and from our Tier 1 testing procedure, and are presented as ours.
The editorial judgments on this page: treating Intercom's three translation features plus Tidio's native answering as a complete menu of engine patterns; assigning knowledge-base answers to native answering or answer-time translation only, which is a maintenance judgment rather than a measured comparison of output quality; adopting the ten-point-from-English pass mark from our multilingual list and the fifteen-question set from our Tier 1 protocol; and giving the native-reviewer cost as an illustrative figure the reader must replace.
See our methodology for how platform facts are verified.
Last updated
6 September 2026.