Use case assessment
Is AI translation worth building?
Genuinely AI, thoroughly solved, and never worth building yourself.
Real AI, mature products, and no differentiation available from building your own. The decision is which content gets human review, not which model.
What it usually means
Translating product interfaces, documentation, support content or marketing material between languages.
Is it genuinely AI?
Yes, and it is among the most mature applications there is. Quality on common language pairs is high enough that the interesting question moved years ago from can it to where is it not enough.
Where it is still not enough
Anything with legal effect. Terms, privacy notices, contracts. A translation error in a binding document is a liability, and the cost of review is trivial against that.
Marketing that carries tone. Machine translation produces correct sentences that land flat. If the copy was written to persuade, translated copy needs someone who can write.
Domain terminology. Regulated, medical and technical vocabulary where a near-synonym is wrong. Glossary enforcement helps and does not close it.
Languages with less training data. Quality varies enormously by pair, and assuming uniform quality is how a bad launch happens in one market.
What the simpler version looks like
Tiered handling. Machine translation with no review for internal and low-stakes content. Machine translation with human post-editing for customer-facing material. Human translation for anything legal.
Deciding those tiers is the whole project, and it takes an afternoon.
What it costs to run
Per-character or per-word on a mature product, which is predictable. The real budget is post-editing on the tiers that need it, and glossary maintenance.
Where it sits under the EU AI Act
Translation is not an Annex III high-risk use.
Article 50 is worth checking on published output. AI-generated text published on matters of public interest carries a disclosure obligation, enforceable since 2 August 2026 — subject to the exception where the content has undergone human review and a person holds editorial responsibility for it.
Which is another argument for the post-editing tier: it resolves quality and the disclosure question in the same step.
When it is worth it
Buy, always. Adopt when you have multiple markets and content volume that makes human translation of everything impractical, which is most companies operating across Europe.
Do not build. There is no differentiation available and the mature options are better than what you would produce.
Common questions
Should we build our own translation system?
No. This is among the most mature AI applications available, there is no differentiation in building it, and the existing products are better than what you would produce. The decision is which content gets human review.
Where is machine translation not good enough?
Anything with legal effect, marketing that carries tone, domain terminology where a near-synonym is wrong, and language pairs with less training data. Quality varies enormously by pair, and assuming it is uniform is how a market launch goes wrong.
Does translated AI content need disclosure under the EU AI Act?
AI-generated text published on matters of public interest carries an Article 50 disclosure obligation, enforceable since 2 August 2026 — unless it has had human review with a person holding editorial responsibility. A post-editing tier resolves quality and disclosure together.
Turning a per-client survey into one framework — multilingual handled from the start, because retrofitting language means reissuing.