How to Offer Multilingual Customer Support With AI Translation

Count which languages your customers actually write in, then switch on the translation built into your help desk or inbox so staff read and reply in their own language while customers see theirs. Add a short glossary of your trade terms, label translated replies, and keep prices, terms and safety instructions in pre-checked translated templates rather than live machine translation.

Written support translates well; the errors cluster in a few predictable places: names, addresses, numbers, dates and trade vocabulary. A reply that renders "night latch" as something unrecognisable, or turns a floor number into the wrong flat, does more harm than a reply in English. The glossary and the check steps below exist for exactly those spots.

Follow me on Instagram@sagnikteaches

Find out which languages you really need

Don't start from a vendor's list of 45 languages. Start from your own messages. Export the last three months of written enquiries (email, forms, chat, WhatsApp) and tally them by language. Most tools detect the language automatically; otherwise, an AI assistant can label a pasted batch in seconds. Count two more things while you're there: messages written in English by people clearly struggling with it, and conversations that dragged on because of misunderstanding. Those customers would also be better served in their own language.

Connect on LinkedInSagnik Bhattacharya

Take an illustrative two-van locksmith covering a mixed residential area. Of about 1,100 written messages in three months, around 9% weren't in English: 52 in Spanish, 31 in Portuguese, 14 in Arabic and 8 spread across other languages. A further 30 or so were in English but hard to follow. That's enough to be worth doing properly for three languages, with a general fallback for the rest.

Subscribe on YouTube@codingliquids

Decide a threshold before looking at the numbers, or you'll talk yourself into supporting everything. Mine is roughly this: a language gets named support (glossary, checked templates, reviewed quality) once it's more than about 2% of messages; anything below that gets translation on request with the standard disclaimer.

Four ways to add translation, depending on what you run

If a chatbot answers customers first, check whether it can reply in the customer's language itself. Intercom's Fin, for instance, replies in the languages you've enabled and can translate your support content from a fallback language when real-time translation is switched on. That covers the first response; the translation routes above cover the humans behind it.

The reply flow, step by step

Whatever the tool, the safe flow has the same seven steps. The two check steps are the ones people drop, and the ones that catch the costly errors.

1. Detect the language automatically, and tag the conversation with it so it can be routed and reviewed.
2. Translate the customer's message in to the reader's language.
3. Check the facts that matter in the original, not just the translation: numbers, names, addresses, dates, anything with a digit in it. Compare them character by character against the customer's own message.
4. Write the reply in your own language, in short plain sentences. Idioms, jokes and long sentences translate badly.
5. Translate the reply out, with your glossary applied.
6. For anything sensitive, back-translate: ask the tool to translate the outgoing text back into your language, and read it. If the meaning has shifted, rewrite the English more simply and try again.
7. Send with the translation label, so the customer knows to query anything unclear.

Step 6 takes 20 seconds and is worth it for prices, appointment times, anything about access to someone's home, and complaints. For routine "thanks, see you Tuesday" messages, skip it.

A glossary for your trade vocabulary

General translation engines know ordinary language well and your trade's language poorly. A short glossary fixes the terms that matter and tells the tool what to leave alone. A filled-in example for the locksmith:

Where your tool supports a glossary or custom terms, enter these there. Where it doesn't, or you're using an AI assistant, put them in the prompt:

Translate the reply below into [language] for a customer of a locksmith.
Use a polite, informal register (the customer wrote informally).

Rules:
- Translate these terms by meaning, not word for word: night latch = a spring
  lock that locks automatically when the door shuts; deadlock = a lock that
  only opens with a key; call-out charge = the fee for visiting the property.
- Copy exactly, never translate: brand and model names, street names,
  flat or floor numbers, access codes, our business name.
- Keep all prices in figures with the $ sign, and all times in 24-hour format.
- Keep sentences short. Do not add anything that isn't in the English.

Reply:
[paste reply]

"Do not add anything" matters more than it looks. Language models sometimes smooth a blunt reply by adding a pleasantry or an assurance ("we will certainly help you"), which in a translation you can't read is an unreviewed promise.

The mistake that shows why step 3 exists

An illustrative message to the locksmith at 11pm, in Spanish:

Hola, me he quedado fuera de casa y la llave está puesta por dentro.
Vivo en el tercero B, el portal tiene código 4417. ¿Cuánto cuesta?

The translation the owner saw:

Hello, I'm locked out of my house and the key is in the lock on the inside.
I live on the third B, the door has code 4417. How much does it cost?

Mostly right, and good enough to understand the job. Two things needed checking against the original. "Tercero B" means the third-floor flat B, and floor numbering conventions differ: in many buildings the ground floor isn't counted as the first. The owner's reply asked, in plain English that translated cleanly, "Is that three floors above the street entrance?" And the key "puesta por dentro" (left in the lock on the inside) changes the method and possibly the price, since some locks can't be opened from outside with a key in them. The translation carried that meaning, but only because the owner read it carefully rather than skimming for the address.

The access code, 4417, was copied from the original, not the translation. Numbers rarely go wrong in translation, but when they do the consequences are out of proportion, so the habit costs nothing.

Rewriting a reply so it survives translation

The quickest quality gain costs nothing: write the English differently. Here's a real-looking reply a locksmith's office might send, before and after.

Before:

No worries at all! We can pop round first thing tomorrow, should be with you by 9ish, and if the lock's had it we'll swap it out there and then. Call-out's 85 plus parts, but we'll give you a heads-up before we crack on.

Back-translated from a machine translation, that came out roughly as "No concerns in total! We can jump around the first thing tomorrow, we should be with you for 9 approximately, and if the lock has had it we will exchange it there and then." The meaning survives in places and collapses in others: "pop round", "had it", "heads-up" and "crack on" are idioms with no reliable equivalent, and "85 plus parts" has lost its dollar sign and its meaning.

After:

That's fine. We can visit tomorrow morning and arrive between 08:30 and 09:30. If the lock is broken, we can replace it during the visit. The visit fee is $85. Parts cost extra, and we will tell you the price before we start any work.

Five short sentences, no idioms, a time window in figures, the price with its currency sign, and an explicit promise about telling the customer before work starts. It back-translates almost word for word. Staff who write for translation for a fortnight usually find their English replies get clearer too, which is a small bonus for every customer.

What should never go out as live machine translation

Some messages are too important to translate on the fly. For these, prepare the text once, in each supported language, and have a fluent speaker check it:

• Prices and what they include. Call-out charges, out-of-hours rates, what happens if a lock has to be drilled. Number formats differ too: some languages use a comma as the decimal separator, which can make "$85.50" look like something else.
• Terms, guarantees and cancellation rules. Anything the customer might later rely on.
• Safety instructions. "Don't try to force the door; if someone vulnerable is inside, call emergency services first."
• Complaint responses that touch on fault or refunds.

For three languages, that might be 600-800 words of templates. A professional translator can check that for a modest fixed fee, and it's the one place in this setup where paying a person clearly beats AI. The trade-offs between the two are costed in AI translation vs human translators. Store the checked templates as saved replies in your help desk, one per language, so nobody translates them afresh.

The locksmith's setup and running cost

The locksmith answers written messages from a shared inbox and WhatsApp, with no help desk. Its setup took about a day in total:

• Two hours: tallying three months of messages and choosing Spanish, Portuguese and Arabic for named support.
• One hour: the glossary and the translation prompt, saved in a business AI assistant plan so client details aren't used for training.
• Two hours: writing the English templates (prices, the out-of-hours explanation, safety lines, the cancellation rule), then sending them to a translator.
• One hour: testing with past messages and checking the back-translations.

Running cost is mostly staff attention. If you use an AI model directly through an automation rather than a chat subscription, the translation itself is close to free. At one of the cheapest current models, gpt-5-nano at $0.05 per million input tokens and $0.40 per million output, 100 non-English messages a month of about 80 words each, translated in and out with the glossary prompt, comes to well under a cent. Even a mid-range model costs pennies. The subscription or help desk you run it in is the real cost.

After two months, the owner's rough tally showed non-English enquiries converting to booked jobs at about the same rate as English ones, where before they'd trailed well behind. The numbers are too small to prove much, but misunderstandings, measured as conversations needing more than four back-and-forth messages to book, fell noticeably.

Tone and politeness differ by language

Many languages distinguish formal and informal "you", and getting it wrong reads as either cold or overfamiliar. Spanish and Portuguese both make the distinction; Arabic has its own conventions of address. Set a default in the prompt or the tool's settings (formal for first contact is the safe choice), and switch to informal only if the customer writes that way. Put the rule in your glossary notes so every reply is consistent.

Sign-offs and greetings are worth fixing too. A reply that starts with a literal translation of "Hi there!" can sound odd. Include a preferred greeting and closing per language in your templates.

Keep your own sentences plain. Staff who write "No worries, we'll sort it ASAP" force the tool to guess at three idioms. "That's fine. We will fix it today." translates cleanly into almost anything, and you lose very little warmth.

Checking the translations customers see

Quality drifts silently, because the people reading the output can't judge it. Three routines keep it honest:

1. A monthly fluent-speaker review. Ten translated conversations per supported language, read by someone fluent: a staff member, a trusted customer who offered, or a freelancer paid for an hour. They mark anything wrong, odd or rude-sounding.
2. Confusion signals. Tag conversations where the customer asked the same thing twice or replied "I don't understand". A rise in one language points at a glossary gap.
3. Back-translation on anything new. Each new template or unusual reply gets a back-translation read before it's saved or sent.

A fuller routine for this is in checking AI translations before customers see them.

Surveyors, HVAC engineers and the limits of translation

Different firms hit different limits. A six-person surveying firm can use translation to answer booking questions and explain what a survey covers in a customer's language, but the survey report itself is a document the client may rely on for a purchase. Machine-translating it creates risk the firm can't check. The honest approach: translate booking and process messages, offer a translated plain-language summary clearly labelled as a courtesy with the original report as the authoritative version, and suggest a professional translation if the client needs one. Whether any client-facing document is safe to machine-translate is weighed in whether machine translation is good enough for client documents.

An HVAC installer meets language on site as well as in the inbox: an engineer explaining to a customer why a system needs a part, or a customer explaining a fault. Written channels are covered by the same setup, and many firms give engineers a phone translation app for simple on-site exchanges, but anything contractual (the quote for the part, the warranty terms) should go in writing afterwards from the checked templates.

Phone support is the hardest channel. Some AI phone agents can hold conversations in several languages, but test any such tool thoroughly in each language before offering it, and don't advertise phone support in a language nobody on the team speaks. A clear line on the website, "Message us in your language by WhatsApp or email and we'll reply in it", is honest and gets most of the benefit. If requests in several languages are arriving in one queue, AI ticket triage can tag each by language so the right templates and reviewers are applied.

AI translation in customer support: common questions

Should I tell customers a reply was machine-translated?

Yes. A short line such as 'This reply was translated automatically; if anything is unclear, just reply in your language' is honest, sets expectations about small errors, and invites the customer to flag confusion rather than act on it. Some help desks mark translated messages in the conversation automatically; check what the customer actually sees before relying on that.

Which languages does AI translate reliably?

Widely spoken languages with a lot of online text usually translate well for everyday support messages. Less common languages, regional dialects, slang and mixed-language messages are less reliable. The practical test is your own: have a fluent speaker review ten real translated conversations in each language you support before advertising it, and repeat the check every few months.

Can the AI chatbot answer in other languages on its own?

Many can. Intercom's Fin, for example, detects the customer's language and replies in the languages you've enabled, and can translate your support content from a fallback language if real-time translation is switched on. The same cautions apply as for staff replies: prices, policies and safety instructions should come from pre-checked content, and there should be a clear route to a person.

Further reads

• How to Spot-Check AI Support Replies: A Weekly Sampling Routine — A weekly sampling routine that covers translated replies too.
• How Translators Use AI to Keep Terminology Consistent — How translators keep terminology consistent with AI.
• Is It Safe to Put Client Documents Into AI Translation Tools? — Whether client documents are safe in translation tools.
• AI Chatbot Disclosure: What to Tell Customers at the Start of a Chat — How to tell customers they're talking to AI.
• Best AI Customer Support Software for Small Teams in 2026 — Support tools compared, several with translation built in.
• AI Email Triage for Shared Inboxes: Sales, Support, and Invoices — Sort a shared inbox before adding translation to it.
• How to Write B&B Room Descriptions and Guest Guides With AI — Room fact cards, prompts for each listing, the words that lead to refund disputes, and a quick way to turn a walk round the house into a guest guide.
• How Small Translation Agencies Build an AI-Assisted Workflow — How a small translation agency sets up AI-assisted work: service tiers, the tool stack, intake routing, linguist briefs, quality sampling and pricing by tier.
• AI Customer Service for Small Businesses: What to Automate First — Rank your customer messages by volume and risk, then automate in order: sorting, drafted replies, proactive messages, FAQ answers, and only then actions.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Intercom help on AI Inbox Translations and Fin multilingual support; Zendesk help on conversation translation; Help Scout pricing page; OpenAI API pricing.