AI Translation and Localisation
Translation at volume with your terminology enforced, quality tiered by what the content is for, and a clear list of what should not be machine translated at all.
Machine translation is now good enough that the interesting decisions are no longer about quality in general but about which content can go straight through, which needs a human editor, and which should never be machine translated regardless of how good the output looks.
AI translation and localisation is the systematic translation of content using machine translation with enforced terminology, quality tiering by content risk, human review where required, and adaptation of formats, conventions and cultural references so the result reads as written for the market rather than translated into it.
Quality tiers, decided by consequence
One quality standard for all content is either wasteful or dangerous. We tier by what the content does:
| Tier | Content | Process |
|---|---|---|
| Raw machine | Internal reference, archive, gisting | No review; clearly labelled as machine translated |
| Light post-edit | Support articles, product descriptions, UGC | Human check for meaning and terminology |
| Full post-edit | Marketing, UI text, customer communications | Human editor for fluency, tone and brand |
| Transcreation | Campaigns, taglines, brand messaging | Human writer working from intent, not text |
| Not machine translated | Legal contracts, regulated claims, safety instructions, medical information | Professional human translation, reviewed |
That last row is a policy decision we recommend making explicitly and writing down, because the cost of an error in those categories is not proportional to the saving from automating them.
What makes translation work at scale
Terminology enforced, not suggested
Product names, feature names, legal terms, scheme names and anything with a fixed meaning belong in a glossary applied at generation and verified after it. Inconsistent product naming across languages is the most common and most visible failure, and it is entirely preventable.
Translation memory still matters
Previously approved translations should be reused rather than regenerated, for consistency and for cost. A memory of your approved segments is an asset that improves every future project, and it is frequently discarded when teams move to machine translation.
Context, not sentences
Translating string by string produces text that is individually correct and collectively incoherent, particularly in interfaces where a button label depends on the screen it sits in. We supply surrounding context and screenshots where the platform allows.
Localisation beyond language
Dates, numbers, currency, addresses, name order, units, right-to-left layout, text expansion breaking interfaces, images containing text, and cultural references that do not travel. These are what make content feel local, and they are what a purely linguistic process misses.
Measure per language with native reviewers
Automated metrics are useful for regression detection and poor at judging whether text is good. Native reviewers who understand the domain assess a sample per language, and no language goes live unattended until it clears the bar they set.
Build the review workflow
Editors need the source, the machine output, the glossary, the memory match and a fast way to correct and approve, with corrections feeding back into memory. A review process routed through spreadsheets and email absorbs most of the saving the automation created.
Machine translation of user-generated content needs a different check
Translating reviews, comments or support messages can turn a mild complaint into an offensive statement, or launder abuse into fluent text. Where user content is translated for customers to read, we add safety checking on the output and label it as machine translated.
How it fits with the rest of your stack
- Real-time conversation between customers and agents, or multilingual assistants, is covered under multilingual and translation bots.
- Search across languages, where a query in one language must find content in another, is part of semantic search.
- Website and product localisation at scale integrates with your content management and string management systems rather than living beside them.
- Regulated markets frequently require specific language versions to be certified or professionally translated, which is a legal constraint rather than a quality preference.
How the engagement runs
Tiers and terminology are agreed first; nothing goes live in a language until a native reviewer signs it off.
Content tiering and glossary
Content classified by consequence, the do-not-machine-translate list agreed, glossary and existing memory assembled.
Pipeline and terminology enforcement
Translation pipeline with glossary enforcement, memory reuse, context passing and localisation formatting.
Quality measurement
Native reviewers assess samples per language and tier, the launch bar per language agreed.
Review workflow
Editor interface with source, output, glossary and memory, corrections feeding back into memory.
Rollout and handover
Languages released as they clear the bar, with per-language reporting and the process for adding the next one.
What you receive
Content in every market you serve, at a quality level chosen deliberately per content type.
Content tiering policy
Which content gets which process, including what is never machine translated.
Translation pipeline
Integrated with your content systems, with glossary enforcement and memory reuse.
Terminology glossary
Per language, enforced at generation and verified after, maintainable by your team.
Localisation handling
Formats, layout, expansion, name order and image text handled per locale.
Native reviewer assessment
Per language and tier, with the launch bar and ongoing sampling process.
Editor workflow
Review interface with corrections feeding back into translation memory.
Is this the right engagement?
Worth being direct. AI Translation and Localisation is the wrong spend in some situations, and those are listed rather than buried.
Good fit if
- You publish content in several languages and volume exceeds human translation capacity.
- Terminology consistency matters commercially or legally.
- Native reviewers can be made available for the languages you serve.
- Content can be tiered by consequence rather than treated uniformly.
- There is a content system the pipeline can integrate with.
Choose something else if
- All content is legal, regulated or safety critical, where professional translation is required.
- No native reviewer exists for any target language.
- Volume is low enough that professional translation is simpler.
- The requirement is real-time conversation, which is multilingual bots.
Frequently asked questions
Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.
Is machine translation good enough for customer-facing content?
For many content types with a human post-edit, yes. For marketing and brand messaging, transcreation by a human writer produces materially better results. For legal, regulated and safety-critical content, we recommend professional human translation and say so rather than automating it.
How do you keep product names consistent across languages?
With a glossary enforced at generation and verified afterwards, not with an instruction in a prompt. Translated product and scheme names are the most visible failure in multilingual content and are entirely preventable.
Do we still need translation memory?
Yes. Reusing approved segments improves consistency and reduces cost, and your memory is an asset that improves every future project. Teams frequently discard it when moving to machine translation, which is a mistake.
How do you assess quality in a language nobody internally speaks?
With native reviewers who understand the domain, assessing a sample per language and tier. Automated metrics are useful for catching regressions and poor at judging whether text reads well, so no language goes live unattended before a reviewer sets and clears the bar.
What about translating customer reviews or support messages?
It works, with an extra safety check on the output, because translation can turn a mild complaint into something offensive or render abuse fluently. Machine-translated user content should also be labelled as such for the people reading it.
Often paired with this
Most clients combine two or three engagements from the Natural Language Processing pillar. These are the ones that most often run immediately before or after.
Document Summarisation
Summaries written for a named reader and decision, checked for faithfulness against the source.
Read more →Resume Parsing and Candidate Matching
Parsing and matching that assists shortlisting, with bias testing, explanations and the decision left with a person.
Read more →Semantic and Vector Search
Search that finds the right thing, measured on your real queries with hybrid retrieval and reranking.
Read more →Is this the right engagement?
Tell us what you are trying to build. If a different service fits better, or if you do not need us at all, we will say so.