| Fluent but wrong | A machine produces confident, well-formed sentences even when it has misread the source — the error is invisible until the reader acts on it. | A native-speaker specialist translates for meaning and verifies against the source, so a smooth sentence is also a correct one. |
| Terminology drift across files | Run through a machine, the same term comes out three different ways across a set of documents; a contract and its annex stop matching. | A glossary and translation memory hold terminology steady across every file in the job, so wording stays consistent. |
| Register and tone | An official letter reads as casual, or a warm personal note turns stiff — a machine has no sense of who the text is for. | The specialist sets the register to the purpose — formal for an authority, natural for correspondence — as a native writer would. |
| False friends and ambiguity | Look-alike words and words with two meanings are resolved by surface form, and the sentence flips to the opposite sense. | Ambiguity is resolved from context by someone who knows the field, not by the most probable token. |
| Names, numbers, dates, formatting | Decimal points, thousands separators, date order and accented names are silently changed, and a figure or a name is then misread. | Number, date and name conventions are set to the target language, and names are matched exactly to the identity document. |
| Confidentiality | Pasting a text into a free tool can send confidential content to a third party and train a model on it — a breach for contracts and records. | Texts are processed under the GDPR, only for the agreed purpose; on request entirely without AI and machine translation, with an NDA where needed. |