{"course":{"slug":"ai-for-small-business","title":"AI for a Small Business","url":"https://addaly.com/learn/ai-for-small-business","level":"beginner","track":"work"},"lesson":{"slug":"working-across-languages","title":"Serving customers in a language you do not read","url":"https://addaly.com/learn/ai-for-small-business/working-across-languages","position":37,"of":88,"minutes":9,"answer":"Translation quality drops sharply outside the best-resourced languages and fails fluently rather than visibly, so back-translate to catch gross errors, have a real speaker read ten replies before you rely on a language.","keyIdea":"Translation quality drops sharply outside the best-resourced languages and fails fluently rather than visibly, so back-translate to catch gross errors, have a real speaker read ten replies before you rely on a language, and never publish permanent text in a language nobody in the business can read."},"body":{"format":"markdown","text":"## Where this earns the most\n\nA café in Barcelona answering in Catalan, Spanish and English. A guesthouse in Da Nang taking a Korean booking enquiry at midnight. A shop in Leicester whose customers write in Gujarati. For businesses like these, translation is not a convenience; it is the difference between a customer and a missed message.\n\nIt is also the area where quality varies most and where the marketing tells you least.\n\n## Quality is not uniform, and nobody says so\n\nMachine translation is strong between the languages with the most text behind them: English, Spanish, French, German, Portuguese, Mandarin, Japanese, Russian, Arabic, Hindi. Good enough for business correspondence, with occasional stiffness.\n\nIt degrades from there, and the degradation is not gentle. For many languages with tens of millions of speakers (Yoruba, Marathi, Sinhala, Amharic, Quechua, Wolof) output ranges from serviceable to confidently wrong, and the confident wrongness is the problem. It does not come back broken. It comes back fluent and mistaken, which is the same mechanism as everywhere else in this course, applied to a language you cannot check.\n\nFormality is where it fails most expensively. Languages with grammatical politeness levels (Korean, Japanese, Javanese, and the tu/usted distinction across Spanish-speaking countries) can produce text that is grammatically perfect and socially wrong. Addressing an older customer with the familiar form is a small error with a large effect.\n\n## Two checks you can run without speaking the language\n\n**Back-translation.** Translate your message into the target language, then, in a fresh conversation, translate it back. Compare against the original. This catches gross errors (a negation dropped, a number changed, a meaning inverted) and takes twenty seconds.\n\nIt does not catch tone, politeness or idiom, because those survive the round trip. So it is a floor, not a guarantee.\n\n**A person, once.** Have a speaker read ten real replies. Not a professional translator; a customer, a neighbour, a member of staff. It costs a coffee and it tells you whether the output is embarrassing, which is the thing you most need to know and cannot discover any other way.\n\nDo this before you rely on a language, and again if you change tools.\n\n## The rule for publishing\n\nDrafting to a customer in a language you cannot read is a manageable risk, because a single reply reaches one person and can be corrected.\n\nPublishing is different. A website, a sign, a menu, a legal notice or an advertisement in a language nobody in the business reads is a standing statement to everybody, and errors sit there for years. Have those checked by a person. It is a one-off cost on a permanent asset.\n\n## Free paths\n\n- **The general assistants** do translation well for major languages and it is included in what you already have.\n- **Dedicated engines**: DeepL and Google Translate both have free tiers, and DeepL is often better on European languages.\n- **NLLB**, released openly by Meta, covers 200 languages and runs on your own machine. For a lower-resourced language, it is frequently better than the general assistants and it costs nothing.\n- **Local models via Ollama** if the content is sensitive.\n\nWorth testing two or three on your own real messages rather than trusting a comparison chart. The ranking differs by language pair, and your language pair is the only one that matters.\n\n## The parts to keep human\n\n**Anything legal or medical.** Terms, safety instructions, dosage, contraindications. The cost of an error is not symmetrical with the cost of a translator.\n\n**The first message to a new market.** You are establishing what kind of business you are.\n\n**Anything where politeness carries weight.** Which is most of Asia and much of everywhere else.\n\n## A note on your own voice\n\nTranslated text tends to arrive flatter than the original. Idiom, warmth and humour are the first things lost, and a business whose appeal is its personality can end up sounding like a form letter in its second language.\n\nThe fix is the same as everywhere: give it examples. Five real messages written by a speaker of that language, in the tone you want, and ask it to match them. The improvement is immediate and it is the same technique as module two, doing the same work."},"check":{"question":"A guesthouse drafts replies in Korean using an AI tool. Back-translation into English looks correct every time. What important class of error does this check miss?","options":[{"text":"Numbers and dates, which frequently change during translation and would need separate verification against the original message.","correct":false,"tempting":"Picks the error class that back-translation catches best, since a changed figure reappears as a changed figure when the text is brought back into English."},{"text":"Politeness level and tone, which survive the round trip into English because English does not encode them grammatically.","correct":true},{"text":"Vocabulary specific to the hospitality trade, which general translation models render inconsistently and which a back-translation would smooth over into generic wording that reads acceptably in English.","correct":false,"tempting":"Names a real weakness with specialist terms, though a mistranslated trade term usually does return as an odd English word rather than passing unnoticed."},{"text":"Regional variation, since Korean differs between South Korea and diaspora communities and the model may produce a variety unfamiliar to the guest.","correct":false,"tempting":"Raises a genuine but marginal issue for this use, and one that a back-translation would not reveal for the same reason as tone, which is the more damaging case."}]},"prev":{"slug":"phone-and-voice","title":"The phone: what a voice assistant can and cannot do yet","url":"https://addaly.com/learn/ai-for-small-business/phone-and-voice"},"next":{"slug":"measuring-service-not-speed","title":"Did the customer actually get better service?","url":"https://addaly.com/learn/ai-for-small-business/measuring-service-not-speed"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'Serving customers in a language you do not read', in AI for a Small Business. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-for-small-business/working-across-languages (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"Serving customers in a language you do not read","courseTitle":"AI for a Small Business","module":{"index":4,"slug":"customers-and-the-front-desk","title":"Customers, from first message to resolved complaint"},"answer":null,"keyIdea":"Translation quality drops sharply outside the best-resourced languages and fails fluently rather than visibly, so back-translate to catch gross errors, have a real speaker read ten replies before you rely on a language, and never publish permanent text in a language nobody in the business can read.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"where-this-earns-the-most","runs":[{"text":"Where this earns the most"}]},{"t":"p","runs":[{"text":"A café in Barcelona answering in Catalan, Spanish and English. A guesthouse in Da Nang taking a Korean booking enquiry at midnight. A shop in Leicester whose customers write in Gujarati. For businesses like these, translation is not a convenience; it is the difference between a customer and a missed message."}]},{"t":"p","runs":[{"text":"It is also the area where quality varies most and where the marketing tells you least."}]},{"t":"h2","id":"quality-is-not-uniform-and-nobody-says-so","runs":[{"text":"Quality is not uniform, and nobody says so"}]},{"t":"p","runs":[{"text":"Machine translation is strong between the languages with the most text behind them: English, Spanish, French, German, Portuguese, Mandarin, Japanese, Russian, Arabic, Hindi. Good enough for business correspondence, with occasional stiffness."}]},{"t":"p","runs":[{"text":"It degrades from there, and the degradation is not gentle. For many languages with tens of millions of speakers (Yoruba, Marathi, Sinhala, Amharic, Quechua, Wolof) output ranges from serviceable to confidently wrong, and the confident wrongness is the problem. It does not come back broken. It comes back fluent and mistaken, which is the same mechanism as everywhere else in this course, applied to a language you cannot check."}]},{"t":"p","runs":[{"text":"Formality is where it fails most expensively. Languages with grammatical politeness levels (Korean, Japanese, Javanese, and the tu/usted distinction across Spanish-speaking countries) can produce text that is grammatically perfect and socially wrong. Addressing an older customer with the familiar form is a small error with a large effect."}]},{"t":"h2","id":"two-checks-you-can-run-without-speaking-the-language","runs":[{"text":"Two checks you can run without speaking the language"}]},{"t":"p","runs":[{"text":"Back-translation.","b":true},{"text":" Translate your message into the target language, then, in a fresh conversation, translate it back. Compare against the original. This catches gross errors (a negation dropped, a number changed, a meaning inverted) and takes twenty seconds."}]},{"t":"p","runs":[{"text":"It does not catch tone, politeness or idiom, because those survive the round trip. So it is a floor, not a guarantee."}]},{"t":"p","runs":[{"text":"A person, once.","b":true},{"text":" Have a speaker read ten real replies. Not a professional translator; a customer, a neighbour, a member of staff. It costs a coffee and it tells you whether the output is embarrassing, which is the thing you most need to know and cannot discover any other way."}]},{"t":"p","runs":[{"text":"Do this before you rely on a language, and again if you change tools."}]},{"t":"h2","id":"the-rule-for-publishing","runs":[{"text":"The rule for publishing"}]},{"t":"p","runs":[{"text":"Drafting to a customer in a language you cannot read is a manageable risk, because a single reply reaches one person and can be corrected."}]},{"t":"p","runs":[{"text":"Publishing is different. A website, a sign, a menu, a legal notice or an advertisement in a language nobody in the business reads is a standing statement to everybody, and errors sit there for years. Have those checked by a person. It is a one-off cost on a permanent asset."}]},{"t":"h2","id":"free-paths","runs":[{"text":"Free paths"}]},{"t":"ul","items":[[{"text":"The general assistants","b":true},{"text":" do translation well for major languages and it is included in what you already have."}],[{"text":"Dedicated engines","b":true},{"text":": DeepL and Google Translate both have free tiers, and DeepL is often better on European languages."}],[{"text":"NLLB","b":true},{"text":", released openly by Meta, covers 200 languages and runs on your own machine. For a lower-resourced language, it is frequently better than the general assistants and it costs nothing."}],[{"text":"Local models via Ollama","b":true},{"text":" if the content is sensitive."}]]},{"t":"p","runs":[{"text":"Worth testing two or three on your own real messages rather than trusting a comparison chart. The ranking differs by language pair, and your language pair is the only one that matters."}]},{"t":"h2","id":"the-parts-to-keep-human","runs":[{"text":"The parts to keep human"}]},{"t":"p","runs":[{"text":"Anything legal or medical.","b":true},{"text":" Terms, safety instructions, dosage, contraindications. The cost of an error is not symmetrical with the cost of a translator."}]},{"t":"p","runs":[{"text":"The first message to a new market.","b":true},{"text":" You are establishing what kind of business you are."}]},{"t":"p","runs":[{"text":"Anything where politeness carries weight.","b":true},{"text":" Which is most of Asia and much of everywhere else."}]},{"t":"h2","id":"a-note-on-your-own-voice","runs":[{"text":"A note on your own voice"}]},{"t":"p","runs":[{"text":"Translated text tends to arrive flatter than the original. Idiom, warmth and humour are the first things lost, and a business whose appeal is its personality can end up sounding like a form letter in its second language."}]},{"t":"p","runs":[{"text":"The fix is the same as everywhere: give it examples. Five real messages written by a speaker of that language, in the tone you want, and ask it to match them. The improvement is immediate and it is the same technique as module two, doing the same work."}]}],"check":{"question":[{"text":"A guesthouse drafts replies in Korean using an AI tool. Back-translation into English looks correct every time. What important class of error does this check miss?"}],"options":[{"text":[{"text":"Numbers and dates, which frequently change during translation and would need separate verification against the original message."}],"correct":false,"tempting":[{"text":"Picks the error class that back-translation catches best, since a changed figure reappears as a changed figure when the text is brought back into English."}]},{"text":[{"text":"Politeness level and tone, which survive the round trip into English because English does not encode them grammatically."}],"correct":true,"tempting":null},{"text":[{"text":"Vocabulary specific to the hospitality trade, which general translation models render inconsistently and which a back-translation would smooth over into generic wording that reads acceptably in English."}],"correct":false,"tempting":[{"text":"Names a real weakness with specialist terms, though a mistranslated trade term usually does return as an odd English word rather than passing unnoticed."}]},{"text":[{"text":"Regional variation, since Korean differs between South Korea and diaspora communities and the model may produce a variety unfamiliar to the guest."}],"correct":false,"tempting":[{"text":"Raises a genuine but marginal issue for this use, and one that a back-translation would not reveal for the same reason as tone, which is the more damaging case."}]}],"order":[2,1,0,3],"why":null},"prev":{"slug":"phone-and-voice","title":"The phone: what a voice assistant can and cannot do yet","minutes":9},"next":{"slug":"measuring-service-not-speed","title":"Did the customer actually get better service?","minutes":8}}}