Localisation teams should use AI where the task is repetitive, checkable and low risk, and keep people in charge where tone, culture, legal meaning or brand voice are at stake. AI is most useful as one stage inside a workflow with clear inputs, review and quality checks, not as a replacement for the whole workflow.
Start with what localisation actually involves
The W3C describes localisation as adapting a product or content to the language, cultural and other requirements of a specific target market, and says it often goes beyond translating text: date, time, number and currency formats, symbols and colours, legal requirements, and address or name formats can all change. That breadth is why a single AI translation step cannot be the whole process. For the conceptual difference, see Localisation vs Translation: What Businesses Should Know.
The same W3C page makes a useful point about timing: internationalisation, the design practice that separates localisable text from code and supports Unicode and bidirectional text, is easier to do early than to retrofit. AI cannot fix a product that hard-codes text into images or layouts.
Where AI fits in the workflow
| Stage | AI can help with | Keep a person responsible for |
|---|---|---|
| Preparation | Extracting strings, flagging untranslatable items, drafting context notes | Deciding what must be adapted, not just translated |
| First draft | Machine translation, terminology-aware drafts | Tone, persuasion, humour, cultural fit |
| Terminology | Suggesting glossary candidates from existing content | Approving the glossary and brand terms |
| Review | Highlighting inconsistencies, formatting errors, missing placeholders | Judging whether the meaning and voice are right |
| QA | Automated checks for numbers, tags, length limits, punctuation | Final linguistic sign-off and in-context review |
| Publishing | Checking language attributes and encoding | Approving legally sensitive or regulated claims |
Customising machine output with your own material
Generic machine translation will not know your product names or house style. Some services let you train on your own previously translated documents. Microsoft documents this in its Custom Translator overview: it builds customised neural machine translation systems using parallel documents so the output reflects your terminology and style, and it can train from a dictionary alone if you have no training data. Microsoft adds that gains in its BLEU quality score of 5 to 10 points are not uncommon with the right type and amount of data. That is a vendor statement about an automatic metric, so it does not tell you how a human reader will judge your content.
The practical lesson: AI output improves most when you feed it approved translations and a maintained glossary. A short briefing checklist helps with that, whether the next step is a person or a tool.
A briefing checklist for any AI or human step
- Audience, market and purpose of the content.
- Brand voice notes with two or three examples of approved wording.
- Glossary of product names, do-not-translate terms and preferred translations.
- Character and length limits for buttons, headlines and subtitles.
- Reference material: previous approved translations and style guide.
- What counts as an error, and who has the final say.
Where AI fails or needs extra care
- Marketing copy and humour. Literal translation of slogans can be flat or wrong. Campaigns usually need adaptation, which is covered for video in Adapting Video Campaigns for Different Markets and Cultures.
- Legal, medical and regulated text. Errors carry real consequences. Use qualified reviewers.
- Names, brands and cultural references. Meanings and associations differ by market, and checks should be done by people who live in the language.
- Language variants. Chinese, for instance, has regional and script differences that affect wording. Launching a Brand Across English- and Chinese-Speaking Markets covers the planning side.
- Silent errors. Fluent output can still be wrong. A fluent mistranslation is harder to spot than an obvious one.
- Confidentiality. Check where text is sent and stored before putting unreleased or personal content into any cloud tool.
Quality assurance that survives automation
Automated checks catch mechanical problems: missing variables, broken tags, inconsistent numbers. They do not catch tone. Keep three layers: automated checks, a bilingual reviewer comparing source and target, and an in-context review of the finished page, app screen or video. For web pages, make sure each page declares its language correctly. The W3C’s guidance on declaring language in HTML says to set the lang attribute on the html element, mark inline changes of language, and use standard BCP 47 tags such as zh-Hans.
Honest limitations of this approach
- Savings depend on content type and on how good your glossary and reference material are. Measure time and error rates on a pilot before committing.
- Reviewers still need time. Treating AI output as nearly finished can make review sloppier, not faster.
- Tools and their language support change quickly, so re-test periodically.
Where Radiance Impact fits
Radiance Impact works in English and Chinese, and its Global team handles localisation and adaptation of content for new markets. If you want to plan an AI-assisted localisation workflow with proper human review, see Localisation and Adaptation, learn more about us, or get in touch.
Sources
- Internationalization and localization, W3C Internationalization
- What is Custom Translator?, Microsoft Learn
- Declaring language in HTML, W3C Internationalization
