By the Onsites AI team · Last updated · 4-minute read
Multilingual support used to be a hiring problem: one agent per language, or a translation agency, or — the quiet default — "just answer in English and hope." The copilot changes the arithmetic. Onsites AI translates a conversation for 3 prepaid credits in either direction, so a two-person team can hold a five-language desk with the same shared queue it already runs. What it cannot do automatically is everything around the translation: which languages to commit to, register — the difference between correct and culturally right answers — and the few sentences that must be perfect on sight. This guide is the operating manual for that layer.
The desk hands you the answer before you commit anything: for one month, tag conversations by the customer's language — the header field does most of it, a tag completes it — then read the sorted histogram. The typical small business finds a spine: 80% of non-native volume in two languages, a long tail of one-offs. Commit formally to the spine (translated macros reviewed by a native speaker if you have one, AI drafts for everything else), and let the long tail ride on translation without ceremony. The common mistake is mirrored ambition: building processes for five languages with a team of three, then under-serving the two that carry 90% of real customers. Your commitment ladder should be: languages the desk answers fluently on the clock (the spine), languages the copilot translates confidently (everything else), and honesty at the boundary — "an answer in your language is being prepared" beats a silently bad translation of a bad English reply. Re-run the histogram every season: multilingual demand moves with your market mix, and the ladder is meant to be climbed deliberately, language by earned language, not declared all at once in an ambitious March and regretted in a swamped May.
Every language splits differently along the formal/informal line, and each split carries commercial weight: German's Sie against du, Spanish usted, Japanese register that signals hierarchy in every sentence ending. Machine translation chooses; the wrong choice is noticed instantly and reads as disrespect or, worse, indifference. Three operational rules keep the register right. First, write source with register in mind: plain, direct, businesslike English drafts translate far better than clever ones — idioms and constructions like "we've got you covered" translate to nothing. Second, set the register once per language in a reviewed macro: your greeting, your closing, your apology line — translated by the best tool you have, checked by a native speaker, then frozen as the house voice. Third, spot-review the spine weekly for a month: read the sent translations of your five most common reply types; when the same failure pattern appears twice, fix the source template, not the translation.
Translation costs 3 credits per action — under three cents — so volume math stays trivial even at scale: a 2-person team handling 1,000 messages a month across five languages, translating roughly half the incoming and a third of the outgoing, burns on the order of 1,500–2,000 credits a month: one $20 pack. Compare that to any historical alternative — per-language agent salaries, agency per-word rates — and the honest statement is that translation is no longer the constraint; register and policy-sense are. The spend-control caveat from the cost guide applies unchanged: prepaid, capped, never a surprise.
Language behavior differs by channel, and pretending otherwise produces odd replies. Email is the most forgiving: formality is expected, translation reads naturally, and a two-line confirmation in the customer's language needs nothing more than the copilot. WhatsApp and LINE are the opposite: short, casual, emoji-tolerant — translated replies should shrink accordingly (the copilot's draft can be told "keep it under two sentences, no salutation"); a translated business letter arriving in a chat feels bureaucratic. WeChat carries its own register system entirely — polite address, tone circles — and deserves the committed-language treatment if Chinese buyers matter at all (the cross-border guide treats this at length). The widget sits in between: the customer saw your site in one language and writes in it, so match the site's language by default. Set the policy per channel in your macros — "email: full sentences and register; chat: two sentences, warm" — and the copilot inherits consistency instead of averaging it away.
Three cases where the honest move is the original language plus a careful note. Money and terms: a refund computation, a discount promise, a contract clause — send it in the language the business record keeps it in, with a plain-language summary in the customer's; a subtle number ambiguity inside a beautifully fluent translation is a dispute nobody can adjudicate. Legal or regulated texts: any language with a version on file — consumer notices, warranty statements — quotes the filing language verbatim. Anger with legal stakes: when a complaint may become a claim, the response should carry a translated courtesy layer over a carefully worded core, reviewed before sending rather than after. The pattern is always the same: translation is for comprehension, not for custody — the record keeps its language, the customer gets theirs, and the two are visibly the same message.
Reserve native-speaker review (yours, a colleague's, or a paid hour) for the sentences that repeat and carry stakes: working-hours and SLA statements in every committed language; price-change and increase notices; refund and money confirmations; the complaint acknowledgement that buys the de-escalation; and your canned answers for the five most common questions in each spine language. Once frozen as templates, these carry the desk's voice in every thread; the copilot's per-message translations then inherit a register someone has actually read. Everything improvised — the odd one-off question in a tail language — can ride on translation alone, which is the entire economic miracle of this feature: the reviewed floor is permanent, and the ceiling keeps rising.
How do we decide which languages to offer support in?
Read the data: tag conversations by customer language for a month, then commit formally to the two languages that carry 80% of non-native volume — reviewed macros in those — and let translation handle the long tail with honest answers at the boundary.
What does AI translation get wrong in customer support?
Not grammar — register. Formality levels (Sie/du, usted/tú), hierarchy in Japanese endings, and idiom are where machine translation reads wrong. Fix the register with reviewed macros in each committed language, spot-review weekly, and write source replies in plain, idiom-free sentences.
How much does multilingual support cost with AI translation?
3 prepaid credits per translation (≈$0.03) in either direction. A two-person team translating roughly 1,500-2,000 times a month spends about one $10-20 pack. The real investment is the reviewed register for committed languages, not the credits.
Which sentences deserve native-speaker review?
The repeated ones: working-hours and SLA statements, price-change notices, refund confirmations, complaint acknowledgements, and the five most-used canned replies in each committed language. Improvised one-off translations can ride on the model.
Can a two-person team really answer in five languages?
Yes — the shared queue is channel-agnostic and language is handled per message. Two agents, translated spine macros, and copilot translation for everything else is the standard shape of a small multilingual desk now.