Guide · For business
AI translation: when it’s enough, when it isn’t and what damage it can do
Machine translation has become good enough that its errors are hard to find. This page shows which texts it suits and which it doesn’t, which types of error are typical and what happens when nobody notices them. With a decision table and a checklist.
In short
- AI translation is enough when a text is for information, stays internal and an error has no consequences.
- It isn’t enough once a text is published, binding or relevant to safety.
- In between lies post-editing: a machine translation checked by a professional, standardised in ISO 18587.
- The real risk isn’t obvious errors but fluent sentences that say something different from the original. Nobody reports them.
- Settle four things before you first use it: an approved tool, a glossary, the language variant and the level of review.
Four levels, not two
The question is rarely “AI or human”. In between are two levels that suit most corporate texts best: machine translation with light or full post-editing by a professional. Which level fits depends not on the budget but on what an error would cost.
| Level | What happens | Suitable for | Not suitable for |
|---|---|---|---|
| AI translation only | Machine translation without human review | Internal information, gist translation, large volumes, short-lived content | Anything that is published or could have consequences |
| AI plus light post-editing | A professional corrects errors of meaning, figures and terminology | Internal documentation, support content, knowledge bases | Marketing, contracts, safety-relevant content |
| AI plus full post-editing (ISO 18587) | Full check against the source text, including style and terminology | Product information, technical documentation, training material, large volumes with quality requirements | Texts meant to have an effect rather than just inform |
| Specialist translation (ISO 17100) | Translation by one professional, revision by a second | Contracts, operating instructions, approval documents, annual reports, campaigns | Nothing. The question here is rather whether transcreation is also needed |
One point that is often overlooked in practice: a document doesn’t have to be assigned to a single level as a whole. The safety chapter of a manual can be handled as a specialist translation while the appendices and parts lists go through post-editing.
Which text needs which level
Three questions are enough to decide. Will the text be published? Would an error have legal or safety consequences? Should the text persuade or only inform? A yes to either of the first two questions means specialist translation.
| Type of text | Recommended level | Why |
|---|---|---|
| Internal notes, minutes, emails for information | AI only | Nobody makes a binding decision on this basis |
| Support articles, FAQs, knowledge bases | AI plus light post-editing | Errors cost understanding, but rarely money |
| Product data sheets, catalogues, training content | AI plus full post-editing | Large volumes, fixed terminology, visible to customers |
| Operating instructions, safety notices and warnings | Specialist translation | Part of conformity, and errors can put people at risk |
| Contracts, terms and conditions, privacy policies | Specialist translation | One misunderstood sentence shifts obligations and liability |
| Approval and regulatory documents | Specialist translation | Authorities and notified bodies read them too |
| Websites, campaigns, slogans | Specialist translation or transcreation | Impact doesn’t come from literal accuracy |
| Annual and sustainability reports | Specialist translation | Figures, technical terms and a critical readership |
| Job adverts and internal communication | AI plus post-editing | Tone determines the effect, and errors are visible |
What needs adapting beyond the translation itself is explained in our guide Translation or localisation?
Typical types of error and how to spot them
Machine translation makes different errors from people. It rarely fails on grammar and often fails on meaning, consistency and context. The difference from the past is that these errors now read completely naturally.
| Type of error | What it looks like | What helps |
|---|---|---|
| Inconsistent terminology | The same component has one name on page 4 and another on page 40, because the model translates each sentence separately | Provide a glossary and translation memory, then spot-check |
| Numbers, units and formats | Swapped decimal separators, unconverted units, a date that is read differently in the target market | Check figures and units separately, not while reading |
| Negations and restrictions | “Must not be operated without the protective cover” turns into a recommendation | Always have safety-relevant sentences checked by a person |
| Ambiguous words | The German “Anlage” as an investment, a plant or an attachment, depending on what the model thinks is most likely | Provide context and write an unambiguous source text |
| Omissions and additions | Language models smooth texts and occasionally drop clauses or add ones that aren’t in the original | Compare section by section with the source text, rather than just reading the translation |
| Invented passages | Where the source text has a gap, a language model fills it with plausible-sounding text | Have anomalies flagged and check them against the original |
| Form of address and register | Informal instead of formal address, or a form of politeness that is inappropriate in the target market | Decide on the language variant and form of address in advance, including for every new edition |
| Format and markup | Lost tags, broken placeholders, tables that fall apart | Work in the file format rather than copying text, then run a technical check |
| Cultural references | Examples, idioms and images that don’t work or cause offence in the target market | Plan for localisation, not just translation |
The most important consequence: you can’t check a translation by reading it. You can only check it by comparing it with the source text. If you only look through an AI translation for good German or good English, you won’t find the errors that matter.
What can follow
The cost of a poor translation rarely arises where the money was saved. Here are seven ways it comes back.
Liability and product safety
A badly translated warning isn’t a language problem but a product problem. After an accident, the documentation becomes evidence, and the way it was translated becomes part of the question of whether due care was taken.
Conformity and market access
From 20 January 2027, the EU Machinery Regulation applies to machinery. Instructions and declarations in the official language of each country are part of conformity. If a language version is missing or flawed, what is at stake isn’t the translation but your ability to deliver.
Contracts that apply differently from what you intended
In multilingual contracts, one version often prevails in a dispute. If the other version differs, both sides are negotiating different obligations without realising it, until it matters.
Silent errors that nobody reports
The most dangerous case isn’t obvious nonsense but a fluent, plausible sentence that says something different from the original. Nobody complains about it, because it reads correctly. It lives on for years, in every new edition and in every translation memory that takes it over.
Visibility in the target market
Literally translated texts miss the terms people search for in the target market. The page exists but isn’t found. That only shows up in the figures months later.
Confidentiality and data protection
If you upload contracts, personnel data or unpublished product information to any tool that comes to hand, you lose control of it. Without a data processing agreement and without clarity about server location and training, that is a data protection problem, however good the translation is.
Reputation with your audience
A text that sounds machine-made tells readers something about how seriously their market is taken. The damage is hard to measure and hard to repair.
None of this argues against using AI. It argues against choosing the level out of habit rather than by risk.
Checklist before you start
Nine points to settle before the first larger project. After that, every further run is routine.
Using AI translation responsibly
Where AI translation really excels
Used properly, machine translation pushes back the limits of what is feasible: content that used to stay monolingual for cost reasons becomes translatable. That is the real gain, not the price per word.
| Strength | What it means in practice |
|---|---|
| Large volumes | Product catalogues, knowledge bases and archives that would otherwise never have been translated |
| Speed | First drafts in minutes, so decisions don’t have to wait for translations |
| Gist translation | Incoming documents in languages nobody in-house can read |
| Groundwork | A machine draft as a basis that a professional checks in a targeted way, instead of starting from scratch |
| Consistent building blocks | Recurring sentences from the translation memory that don’t need translating again |
The difference between useful and risky lies not in the technology but in the decision beforehand: do you know which level this text needs?
FAQ
When is AI translation enough?
When the text is meant to inform rather than persuade, when an error has no legal or safety consequences and when nobody outside the company will see it. Typical examples are internal notes, minutes and large volumes that only need to be understood.
When isn’t it enough?
As soon as a text is binding, published or relevant to safety: operating instructions, warnings, contracts, approval documents, annual reports, campaigns. What counts here is not the average quality but the worst single sentence.
Isn’t AI translation very good these days?
It is good enough that its errors are harder to find. That is the real difference from the past: machine translation used to be recognisable as such, but today even the wrong version reads fluently. Checking shifts from a feel for the language to comparison with the original.
What is post-editing?
The revision of a machine translation by a professional, standardised in ISO 18587. Light post-editing removes errors of meaning, while full post-editing brings the text up to the level of a specialist translation.
Can I upload contracts and personnel data to an AI tool?
Only to tools your company has approved and for which a data processing agreement is in place. Also check whether your input is used for training. If in doubt, the text doesn’t belong in a public tool.
How do I find out which level my text needs?
Three questions are usually enough. Will the text be published? Would an error have legal or safety consequences? Should the text persuade or only inform? A single yes to either of the first two questions means specialist translation.
More guides
-
AI and translation
Proofreading AI texts
Which errors AI texts contain and how to check them.
-
AI and translation
Prompt engineering
Eight building blocks, 14 prompts to copy, style lists and two PDFs (German).
-
For business
Translation or localisation?
Which content needs which level, with a decision table.