Translating a professional text with AI: the trap to avoid

You paste an email in your language into an AI assistant, you ask “translate into English,” and you get a correct text in a few seconds. Except that a “correct” text is not always a text that works for a foreign reader. That is where most AI translations stumble: they translate the words, not the intent.

Translating is not localising

Two very different reflexes hide behind “translate a text”:

• Translation converts the text from one language to another, keeping the meaning.
• Localisation (or adaptation) goes further: it adjusts the tone, cultural references, formats (dates, currencies, units), courtesy phrases and sometimes even humour, so the text sounds natural in the target culture — not only in the target language.

A literally translated contract can be legally correct and yet sound strange, even clumsy, to a native reader. A courtesy phrase translated word for word can seem too formal, or on the contrary too casual, depending on the context. AI excels at the first reflex. It needs to be guided for the second.

Why AI translates well, but adapts poorly by default

A general-purpose AI assistant, prompted with a simple “translate this text,” produces a translation faithful to the meaning — already real progress compared with the tools of a few years ago. But without further instruction, it does not know:

• who the text is addressed to (a client, a partner, a superior);
• in which country and culture it will be read (British English is not American English, and professional conventions vary even more);
• whether certain references (a public holiday, an idiom, a numeric example) should be adapted rather than translated as is.

The model does not guess this context: you have to give it. It is the same logic we detail in 7 principles to write better prompts — an LLM produces what you ask for, not what you forgot to specify.

The 4-step method

1. Give the context before the text

Do not paste the text to translate directly. First specify: the target language and country (not just “English,” but “American English, for a client based in New York”), the document type (email, contract, presentation), the recipient and the expected register (formal, cordial, technical).

Translate the text below into American English, for a professional
email addressed to a client. Register: cordial but formal.
Adapt the courtesy phrases to American conventions rather than
translating them literally. Text: [paste]

2. Ask for adaptation, not just translation

Explicitly add the instruction to adapt cultural elements: dates, number formats, local references, units of measurement. A “15/03/2026” becomes a problem in American English if no one specifies the expected format (month/day/year).

Also adapt dates to the American format (month/day/year),
amounts to dollars if an amount in euros appears (give the
approximate conversion in brackets), and flag any cultural
reference that might not be understood as is.

3. Ask for an explanation of the tricky choices

On an ambiguous passage, an idiom or a phrase with no direct equivalent, ask the model to flag its choice rather than deciding silently. That way you keep control of the decisions that matter.

If an expression has no direct equivalent, propose your translation
but flag it in brackets with a brief explanation of your
choice, so I can validate or adjust.

4. Reread with a native eye, not just a grammatical one

An AI translation can be grammatically impeccable and yet sound “translated” — a little too literal, a sentence rhythm that betrays the source language. For a high-stakes document (contract, public communication, important client file), have it reread by someone whose native language it is, or ask the AI itself for a second pass:

Reread this translation as if you were a native [target-language]
writer. Flag any sentence that sounds "translated" rather than
natural, and propose a rephrasing.

What AI does not replace

On a contract, a regulatory document or any legally binding text, an AI translation remains a first draft, not a final deliverable. The stakes of legal liability exceed what a prompt-based quality check can guarantee: a professional translator or a lawyer must validate the final version. Specialised tools like DeepL Pro, built specifically for translation and for preserving document formatting, are also an alternative to consider for intensive use — they do not replace a general-purpose assistant, but the two complement each other depending on the document.

And as with any professional use of AI, never paste a confidential contract or clients’ personal data into a consumer tool without first checking the data-handling terms. Our confidentiality habits for AI at work detail this point.

Key takeaway

A good AI translation begins before the text to translate: context, recipient, register and an explicit cultural-adaptation instruction. Without these details, you get a text accurate in its words and sometimes wrong in its intent. With them, the gap with a professional translation narrows sharply — for anything that does not legally bind your company.