Machine translation in customer service: what to check first

An email in German lands at a rural guesthouse in Asturias. Nobody on the team speaks German, so someone pastes it into a translator, writes the reply in Spanish, runs it back through the translator and sends it. The guest replies delighted, with a "vielen Dank" and a couple more questions. Nobody noticed a thing. A few years ago the same scene would have ended with baffling sentences and a confused guest; today, with the most common languages, you almost never notice.

That shift is worth understanding properly. This piece is for you if your business gets messages in languages nobody on the team really speaks — tourists, overseas customers, residents from other countries — and you're already answering them with machine translation, or thinking about it. You'll see what has genuinely improved, where it still goes wrong, and what to check before you hit send.

What changed: from word-by-word to context

Older translators worked sentence by sentence, almost word by word. Today's AI models read the whole message, understand what it's about and write the reply the way someone who speaks the language would. That's why they now get things right that used to be impossible: tone, polite formulas, idioms.

For a small business this means something very concrete: replying to a customer in English, French, German, Italian or Portuguese no longer requires someone who speaks the language. The result is usually correct and natural.

But "usually" isn't "always". And the errors that remain aren't the old kind. They're no longer clunky sentences the customer spots straight away. They're errors that sound perfectly natural and say something else. Those are the dangerous ones.

Where it still goes wrong — and why you won't notice

If nobody on the team speaks the language, nobody can spot the mistake by reading the output. So it helps to know in advance where mistakes tend to hide:

• Numbers, dates and prices. "03/04" is 3 April in Spain and March 4 in the US. "1.500" and "1,500" mean different things depending on the country. A translation that leaves them as they are, or "localises" them without being asked, can shift an arrival date or a quote.
• Commitments. "We could try to refund part of it" can come out sounding like a firm promise of a refund. A Spanish "maybe" arrives as a "yes".
• Formality. In German or French, switching between formal and informal "you" mid-conversation stands out and can come across as rude. Translators don't always keep the same register from one message to the next.
• Proper names. Your service name, a dish, or a street translated literally ("Calle Mayor" as "Main Street") confuses anyone trying to find you later.
• Less common languages. Quality isn't the same everywhere. In widely spoken languages it's very good; in languages with less text available to train the models, it drops. And those are exactly the ones where fewest people can check anything.

The real risk isn't the translation — it's everyone doing it differently

There's a problem that slips under the radar. If each person on your team translates their own way — one uses a free browser translator, another pastes the email into an AI chat on their personal account, another does it on their phone — you've got three issues at once:

1. Inconsistent results. The same customer gets replies in different tones depending on who answers.
2. Customer data leaving through doors you don't control. That pasted email carries names, dates, sometimes health or payment details. If you don't know which tools are in use, you don't know where that data ends up. We go into it in this piece on GDPR and AI.
3. No copy in your own language. If the conversation only exists in German, when the customer complains a month later nobody will know what was promised.

You don't need to ban anything. You just need to pick one tool, set a few ground rules and have everyone work the same way.

What to check before you send

A short list for when you're replying in a language nobody on the team speaks:

1. Write it in your own language first, and approve that version. The commitment is decided in your language. The translation is just the wrapping.
2. Translate it back. Run the final reply back into your language and read it. If it says something different from what you meant, there's a problem. It takes a minute and catches most meaning errors.
3. Check numbers, dates, prices and addresses by hand. Compare them one by one with the original. Where you can, write dates with the month in words ("3 April") to sidestep format confusion.
4. Keep a fixed glossary. A short list of things that never get translated, or always get translated the same way: your service names, your address, your cancellation terms. Give it to the tool every time.
5. Keep the register steady. Decide whether you're formal or informal in each language and don't switch mid-conversation.
6. Store both versions. The original and the translated message, together in the customer's history.

When not to trust machine translation

There are cases where the risk outweighs the time saved:

• Contracts, legal terms or formal complaints. These need a professional translator or, at the very least, someone who knows the language to check them.
• Health or safety instructions. A dose, an allergy or a warning mistranslated isn't a matter of style.
• Angry customers. A complaint that's already going badly can get worse over a shade of tone. If the customer speaks English and so do you, plain English is sometimes safer than flawless German you can't control.

Where this is heading

This will only grow: more and more customer-service tools translate on their own, with nobody copying and pasting. That's good news, but it changes the question. It's no longer "can we reply in German?" but "who decides how it's done, and what gets checked?".

If your team already uses AI to translate — or for anything else — and everyone does it their own way, sorting that out is exactly what we do in AI for teams: one shared way of working, with approved tools and clear rules on what data can leave the business.