AI has made translation remarkably easy. A company can copy a text into ChatGPT, Claude, Gemini or another AI tool and receive a fluent translation within seconds. For marketing teams, product departments and customer support, that creates an obvious opportunity: more content can be translated faster and at a much lower apparent cost.
The difficulty usually appears later. Once several departments, languages and content types are involved, the challenge is no longer simply producing a translation. It is keeping multilingual content consistent, reusable and under control.
In other words, generating translations has become easy, while managing the translation process has not.
Translating documents instead of managing a translation process
A typical company starts small. Someone needs a product brochure in Italian, so they upload it to an AI tool, check the result and publish it. A few weeks later, another department translates the website. Later still, the technical team translates documentation and customer support starts using AI for emails.
Each task may work perfectly well on its own. The problem is that these translations are often disconnected from one another. The AI used for the new brochure may not know how a feature was translated on the website three months earlier. It may not know that a technical term has already been approved by the local distributor or that a particular product category has an established name.
As the volume of translated content grows, the company can therefore end up with several perfectly acceptable translations for the same concept. None of them may be linguistically wrong, yet the overall multilingual communication becomes inconsistent.
Traditional translation processes were designed partly to prevent this. Translation memories store previously approved sentences, terminology databases define preferred terms, and style guides establish language-specific conventions. Companies that move directly from manual translation to general-purpose AI sometimes leave this structure behind without realising what they are losing.
The result is a paradox: translations become much faster to produce, but increasingly difficult to maintain.
For occasional translation this may not matter. For a company managing websites, product catalogues, manuals and customer communication across ten or twenty languages, it quickly does. At that point, the relevant question is no longer simply how to translate the next document. It is how that document fits into everything the company has already translated and how corrections and terminology decisions will be reused in future projects.
Terminology does not manage itself
This becomes particularly visible with terminology.
Consider a manufacturer that uses the English term control unit. Depending on the product, context and target language, several translations may be possible. In German, for example, the correct term might be Steuergerät, Steuereinheit or Bedieneinheit. In Croatian, depending on what the unit actually does, possibilities might include upravljačka jedinica, upravljački uređaj or kontrolna jedinica.
These are not necessarily interchangeable.
A control unit controlling an engine is not necessarily described in the same way as a unit used by an operator to control a machine. If the AI sees only the isolated English term, it may choose a perfectly plausible translation that is nevertheless wrong for the particular product.
Even when several alternatives are technically acceptable, inconsistency can create problems. The AI might use one equivalent in a user manual, another on the website and a third in service documentation. All of them may sound natural, which is exactly why the inconsistency can survive unnoticed.
For technical products, software functions, medical terminology, service names and branded concepts, consistency is not simply a stylistic preference. Customers need to recognise the same feature across different documents. Technicians need to know that two texts refer to the same component. Sales teams and local subsidiaries need a shared vocabulary.
Professional translation therefore involves more than asking whether a sentence is correct. It also asks whether the company has already decided how that concept should be expressed.
AI can follow terminology very well when it is given a reliable glossary and appropriate reference material. Without them, however, it has to choose among several plausible alternatives, and those choices may differ from one project to the next.
This is why a relatively simple terminology database can sometimes improve an AI translation process more than a highly elaborate prompt. Knowing how the company wants its most important 200 or 300 terms translated creates a foundation that the AI can actually follow.
Translation is not only text
Another common mistake is to treat translation as if it involved nothing more than copying sentences from one language into another.
Business documents contain much more than plain text.
A Word document, for example, may contain headings, lists, tables, bold and italic formatting, footnotes, cross-references, hyperlinks, text boxes and carefully structured layouts. When the text is simply copied from Word into an AI chat and the translation is copied back, part of that structure may be lost or altered. The translation itself may be good, but someone still has to reconstruct the document afterwards.
The same problem becomes even more obvious with other types of content. Website exports contain HTML. Software localisation includes variables and placeholders such as {username} or %PRODUCT_NAME%. Technical documentation may contain cross-references, model designations, measurements, warning structures and numbered procedures. Marketing systems use variables that are later replaced with names, prices or locations.
These elements may appear inside the text, but they are not ordinary text. Some must remain exactly where they are, some must never be translated, and others have to move together with the surrounding sentence.
General-purpose AI models are increasingly good at handling structured content, but they are not infallible. A placeholder can disappear, a tag can move, spaces can be introduced, numbering can change, or a product code can be interpreted as something that should be translated.
In a single paragraph, such problems are usually easy to spot. In a document with dozens of tables or an export containing 20,000 strings, they become much harder to detect.
There is also a hidden efficiency problem here. If an employee saves 30 minutes by generating a translation with AI but then spends an hour restoring Word formatting, rebuilding tables and checking whether links and references survived the copy-paste process, the apparent productivity gain becomes much smaller.
A scalable translation process therefore needs to preserve both the language and the structure of the original content. This is one of the main differences between occasionally translating text in a chat interface and integrating AI into a professional business process.
When everyone can translate, who owns the translation?
AI also changes the organisation of translation inside companies.
Previously, translation was often centralised because it required either an external translation provider or an internal language team. Today, virtually any department can translate content independently. Marketing may use one AI tool, customer support another, the product team may develop its own prompts, and local subsidiaries may correct translations according to their own preferences.
There is nothing inherently wrong with decentralising translation. The problem arises when the company decentralises production without defining ownership.
Someone still needs to decide which terminology is official, which corrections should be reused, which content requires professional review and how consistency should be maintained between the website, product documentation and customer communication.
Without that ownership, each department optimises its own immediate task while the multilingual content as a whole becomes less coherent.
This problem tends to grow quietly. One inconsistent term is insignificant. Ten are irritating. Hundreds spread across multiple documents and languages become expensive to correct, particularly when nobody knows which version should be considered authoritative.
AI therefore makes translation governance more important, not less. The faster a company can generate multilingual content, the more important it becomes to control what is being generated.
AI translation needs a process, not just a better prompt
When companies first experiment with AI translation, they naturally focus on which model performs best and how to write the perfect prompt. Those questions matter, but they become less important once translation turns into a recurring business activity.
At that stage, the larger questions are whether the AI has access to approved terminology, whether previous translations can be reused, whether document structure and protected elements are preserved, whether corrections are fed back into future work and whether someone is responsible for multilingual consistency.
The companies that use AI most effectively will therefore not necessarily be those with the most sophisticated prompts. They will be the ones that integrate AI into a controlled translation process, rather than treating every new translation as an isolated task.
In Part 2, we look at a different set of mistakes: why fluent AI translations can hide important errors, why the same process should not be used for every type of content or every language, and why companies often overestimate how much time AI translation actually saves.
Need a translation?
Contact us and we'll make you a personalized offer.
Fill out the form and we will prepare a customized offer for you. For more information you can also call us at +386 1 620 26 16.