An AI Translation Pipeline That Actually Works: Machine Draft, Human Judgment

An export-business client once did the math with us: translating their website and catalog into English and Japanese, with a translation agency charging per word, ran into six figures per redesign — and every new product launch meant another round. So many companies' multilingual sites freeze at "the version translated three years ago," with new products available only in Chinese while overseas customers stare in confusion.

Since LLMs arrived, the cost structure of this problem has fundamentally changed. But first, the cold water: "fully automatic machine translation straight to production" is still a disaster in commercial settings. What actually works is a hybrid pipeline — machine translation for the heavy lifting, human review as the gatekeeper. This piece shares how we've built that production line for ourselves and our clients.

Why LLM translation beats traditional machine translation

Traditional machine translation converts sentence by sentence. LLM translation is different in kind: it understands context, and it takes instructions. That yields three practical differences:

• Glossaries can be enforced. Put your brand names, product names, and industry terms into the prompt as a mapping table and require adherence throughout. "Stainless steel vacuum flask" stays the same English phrase site-wide, instead of one rendering on this page and another on that one.
• Tone can be specified. B2B catalogs need formality; brand stories need warmth. The same batch of content can carry different tone instructions.
• Formatting survives. HTML tags, variable placeholders, Markdown structure — instruct clearly and they're preserved intact, which saves enormous engineering effort in website localization.

What our hybrid pipeline looks like

1. Build the glossary and style guide. This is the foundation of the entire line — invested once, reused forever. Brand terms, do-not-translate terms, unit formats, date formats, all in black and white.
2. Machine first draft. Content is exported from the CMS and sent to the model along with the glossary and style guide, producing drafts in batches. This step replaces 70–80% of the translation workload.
3. Machine self-review. A second LLM pass checks the draft: did any terms drift, are numbers and units correct, are any sentences missing? Cheap, and it catches most of the low-level errors.
4. Human review. A reviewer who is a native speaker or fluent in the target language does exactly two things: fix unnatural phrasing, and verify high-risk content (prices, specifications, legal statements). Because they review rather than translate, they work several times faster than pure human translation.
5. Write-back and version management. Translations return to the CMS, and whenever the source text changes, the corresponding translation is flagged "needs update" — so multilingual versions never quietly go stale.

Note where the human sits: not eliminated, but moved to the position of highest value. Machines handle volume; humans handle judgment — the design philosophy shared across all our AI pipelines.

Three pits everyone falls into

Pit one: direct translation is not localization. A Taiwanese site writes "free shipping"; Japanese e-commerce has entirely different conventions and threshold culture. English product pages follow a different narrative order than Chinese ones. For high-stakes pages (homepage, flagship product pages), we recommend localized rewriting rather than translation; save the translation pipeline for the long tail.

Pit two: regulations and claims language. Efficacy claims for food, supplements, and cosmetics have different red lines in every market. A sentence that's compliant in Chinese isn't necessarily legal once translated into English or Japanese — this needs human gatekeeping and can't be delegated to the model.

Pit three: you translated it, but search engines don't know. Publishing multilingual content is only half the job — get the language annotations and URL structure wrong and Google will confuse your versions. We cover the details in SEO for multilingual websites; read it before building a multilingual site.

What AI translation saves isn't the "translation fee" — it's the opportunity cost of "translation is so expensive we just stopped updating." Content that moves in real time with the business is this pipeline's real value.

Where to start

Don't translate the whole site in one go. Pick the pages most tied to inquiry conversion — product pages, FAQ, contact form — build the glossary and run the pipeline end to end, confirm the quality and cost meet expectations, then expand to the blog and catalog. Building the full pipeline, including CMS integration and version tracking, is a small project with a well-defined scope. If you'd like to evaluate it, talk to us about your language pairs and content volume.