Customer Service
AI Translation for Small Business: Serve Customers in Their Language
In any multilingual city, some of your customers would rather write to you in French, Spanish, Mandarin, Punjabi or Portuguese — and many simply buy from whoever answers them clearly. AI translation is now good enough to handle most everyday business messages in seconds, for cents. The skill is not picking the "best" engine. It is knowing which messages are safe to translate and send untouched, which ones need a human read first, and how to keep your product names and prices saying the same thing in every language.
What it is
Modern translation tools are neural engines trained on enormous amounts of bilingual text. You paste or upload text — an email, a product page, a whole Word or PDF document — and get a fluent translation back, usually in seconds. Independent guides that test these engines report that DeepL tends to read most naturally for European language pairs, while Google Translate covers more languages overall. For a small business, both are far past the threshold for everyday messages; the differences that matter are workflow, terminology control and price structure.
The feature that matters most is the glossary. DeepL's document translation preserves the original file's formatting, and its glossary feature lets you pin specific terms — your product names, service labels, and words that must never be translated — so they come out the same way every time. Style and formality controls matter too: the right register for a customer message in German or French is the formal one, and a tool that lets you set that once beats re-explaining it per message.
What translation AI does not do is understand your obligations. It will fluently translate a wrong price, an expired policy and a promise you never meant to make. That is why the workflow below sorts content by risk before anything is translated.
Who it's for
Best for businesses that regularly serve customers in a second language: shops and clinics in multilingual neighbourhoods, trades firms quoting newcomers, online sellers shipping across language lines, tourism and food businesses with international visitors. If you handle even a few non-English conversations a week, translation stops you losing those customers to silence or to a competitor who replied.
It helps less if your customer base is effectively monolingual, and it is the wrong tool — on its own — for regulated or high-stakes documents. Contracts, consent forms, safety instructions and anything with legal effect need a qualified human translator; the machine can produce a draft for that translator to work from, which still saves money, but it should never be the final step.
What it costs
Prices checked 2026-10-16. There are two ways to pay. Subscriptions are built for people pasting text and files into an app. DeepL's plans, as reported in eesel AI's 2026 review from DeepL's own pricing page: Individual is €7.49 per month billed annually, with 300,000 characters a month, 3 file translations and 1 glossary. Team is €24.99 per user per month billed annually, with 1,000,000 characters per user, 20 file translations, 5 glossaries and shared style rules. Business is €49.99 per user per month with unlimited characters. The free version allows 50,000 characters and one file translation a month — enough to test quality on your real messages before paying anything.
Per-character APIs are built for volume and automation — a website, a help desk or an ordering system translating continuously. Google Cloud Translation charges $20 per million characters for its standard text model, with the first 500,000 characters each month covered by a $10 credit, and it counts every character you send, including spaces and markup. Microsoft's Azure Translator undercuts that: its free F0 tier includes 2 million characters a month, and its pay-as-you-go S1 rate is $10 per million characters for text and $15 per million for document translation.
| Option | Published price (checked 2026-10-16) | Billing unit | Watch for |
|---|---|---|---|
| DeepL Individual | €7.49/month billed annually | Per user, 300,000 characters/mo | 1 glossary only; file translations capped at 3/mo |
| DeepL Team | €24.99/user/month billed annually | Per user, 1M characters/user/mo | Per-seat billing adds up if the whole shop needs access |
| Google Cloud Translation | $20 per million characters; first 500k/mo covered by a $10 credit | Per character sent | Every character counts, including whitespace and HTML markup |
| Azure Translator | Free tier 2M characters/mo; then $10 per million characters | Per character sent | Needs a developer or an app that integrates it; no friendly paste-in window |
The crossover: a typical customer email runs about 1,000 characters. Two hundred translated conversations a month is roughly 200,000 characters — about $2 of Azure usage, $4 of Google usage beyond the free credit, or comfortably inside a single DeepL Individual seat. Subscriptions win on convenience until your volume is automated and large; APIs win the moment software, not a person, is doing the translating.
Target ROI: the honest math EXAMPLE
The return is rarely the translation bill — it is the staff time and the sales that used to leak away. Price both honestly.
Example: Suppose your shop handles 30 customer conversations a month in a second language, and each one takes a bilingual staff member about 20 minutes of translating and drafting — 10 hours a month, about $250 at $25/hr. If AI drafts cut that to 5 minutes of review per conversation, the review load is 2.5 hours, freeing roughly $188 a month against a tool cost between $2 of API usage and a €24.99 Team seat. The second return is harder to count but often bigger: inquiries answered the same day instead of when the bilingual staffer is next on shift. Track reply time and conversion on translated conversations separately, or you will never see it.
What to actually measure:
- Minutes of human time per translated conversation, before and after
- Reply time to non-English inquiries versus English ones
- Characters translated per month, against your plan's allowance
- Corrections caught in review per week — your early quality signal
- Sales or bookings from conversations held in a second language
How to set it up
- Sort your content into three risk tiers. Tier 1, translate and send: hours, directions, order status, appointment confirmations, simple product questions. Tier 2, translate then human-read before sending: quotes, invoices, delivery terms, anything with prices, dates or promises. Tier 3, human translator territory: contracts, policies, safety and medical information, complaint responses with legal exposure. Write the tiers down where staff can see them.
- Test on your real messages, free. Take twenty recent conversations, translate them with the free tiers, and have a fluent speaker grade the output for accuracy and tone. Success looks like this: Tier 1 messages need no corrections beyond the occasional awkward phrase. If numbers, names or negations come back wrong even once, that content moves up a tier.
- Build your glossary before you scale. List your product and service names, brand terms, and anything that must never be translated — plan names, model numbers, your business name. Add the correct target-language term for each specialist word your trade uses. Thirty careful terms prevent most of the embarrassing errors; add to the list every time a review catches one.
- Set formality once. Choose the formal register for customer-facing text in languages that distinguish it, and save that as the default style rather than deciding per message.
- Translate incoming and outgoing as one workflow. Staff read the customer's message in translation, draft the reply in your language, translate, and — for Tier 2 — get the fluent read before it sends. Keep the original and the translation together in the thread; you will need both the first time a customer quotes your own words back at you.
- Sample-audit weekly. Each week, pull five sent translations at random and have them checked. Log every error by type — terminology, number, tone, meaning. A rising error count in any type is your signal to move that content up a tier or fix the glossary.
Watch-outs and honest limitations
Numbers are the danger zone. Dates, prices, quantities and negations — "not included," "no refund after" — are short, easy to skim past, and catastrophic when flipped. The Tier 2 human read exists mainly to check numbers against the source, every time, before sending.
Privacy deserves one decision up front. Text you paste into a free consumer tool may be handled under weaker terms than a paid business plan; one reason to pay is that DeepL's paid plans state customer data is not used to train its models. Customer messages can contain addresses, health details and payment questions — route them through a paid plan with clear data terms, and never through a free tool you have not read the policy of. This guide is general information, not legal advice.
And keep expectations honest for low-resource languages and heavy dialect. Quality is strongest across major European and world languages and thinnest exactly where your customers may need it most. The sampling audit is what tells you where your languages sit — trust it over any engine's marketing.
What to measure in your first 30 days
- Human minutes per translated conversation, trending down
- Errors caught per week in the sample audit, by type
- Reply time to second-language inquiries versus your English baseline
- Characters used versus plan allowance — are you on the right tier?
- Bookings or sales traced to conversations held in a second language