AI localization at scale. Your standards, on the record.
Large batches and ongoing programs, from product UI to legal and medical text: the full seven-step cycle runs on every file, with your glossary applied and a quality report on each.
Tell us how much content you have, and whether it’s a one-off batch or ongoing updates. We’ll scope it in the quote.
Your program grows; the method doesn’t thin out
One large batch, or release after release. Scale changes the size of the job, not the checks.
- Counted at intake. An agent counts new, matched, and repeated words, and finds what changed since your last batch. Your quote shows the breakdown.
- Unchanged text isn’t translated again. Text that hasn’t changed since your last batch comes from your translation memory (TM).
- A level per content type. You set the level for each content type, and your quote records it.
- Your TM comes back. Your translation memory comes back updated, with a TMX file on request.
- A report on every file. Every file, in every batch, comes back with its quality report.
What you send, what comes back
| You send | You get back |
|---|---|
| Files in your format (for example, DOCX, XLSX, CSV, XML, JSON, XLIFF, HTML, or PO), or a job in your translation management system (TMS) | The same files in the same format, with tags and placeholders intact |
| Your glossary, translation memory (TM), and style guide, if you have them | The quality-gate result and an issue log with every delivery |
| The level for each content type: AI, or AI with + Human check, + Human edit, or + Full review | A quality report on every file that states the level that ran |
| For Arabic, the variety each content type is written in | For Arabic, the variety named in the report |
The devil’s advocate (the agent that challenges every finding) and the judge run on every file, at every level. If a file fails the quality gate, it doesn’t ship. At every level, you get proposed glossary changes and open questions; at + Human edit and above, the editing linguist’s queries are added.
Domains we run through the cycle
Every domain runs the full cycle at the AI level, with a quality report on every file.
Example
| Domain | Typical content | Source segment | Where errors tend to hide |
|---|---|---|---|
| Product and UI | UI strings, release notes, in-app messages | Tap Save to keep your changes. | String length, placeholders, and your approved UI terms |
| Help and support | Help-center articles, support macros | If the code expires, request a new one. | Product terms that drift from your UI, and steps out of order |
| Marketing | Campaign pages, emails, product descriptions | Your weekend, sorted. | The tone your style guide sets, and wordplay that doesn’t carry over |
| Legal | Terms of use, policies, contracts | The Customer may end this Agreement with 30 days’ notice. | Defined terms, modal words such as may, shall, and must (obligation, permission, or possibility, depending on context), and every number and date |
| Medical | Patient leaflets, device instructions, clinical summaries | Take one tablet twice a day with food. | Doses, units, numbers, negations, and your glossary terms |
| Technical | Manuals, specifications, knowledge bases | Tighten the bolt to 25 N·m. | Units, values, warnings, and tags |
The source segments are illustrative, written for this page.
Legal and medical text runs through the same full AI cycle as everything else. The report records every finding, every ruling, and the level that ran.
Your glossary and style rules, applied every time
The AI translates against your glossary, translation memory, and style rules. A terminology check flags every segment that departs from them, a devil’s advocate challenges each finding, and a judge rules on it and records the reason. The rules that can be tested reliably run as automatic checks in the quality gate.
No glossary or style guide yet? We build them with you from your own content before production starts, and we can align your past translations into a TM.
Inside your platform, or from your files
We work inside your TMS, including Phrase, XTM, Smartling, Crowdin, and Lokalise, or from the exports you send. Files come back in your format, with tags and placeholders intact.
If your product ships often, continuous localization keeps each language in step with your releases and doesn’t translate unchanged text again.
Arabic, in MSA or the dialect your customers speak
We translate into Modern Standard Arabic (MSA) by default, and into Gulf, Egyptian, Levantine, Maghrebi, Iraqi, or Sudanese Arabic when your audience speaks it.
- Set at intake. We fix the variety for each content type before work starts.
- A glossary per variety. The same concept can carry a different approved term in each variety.
- AI prompted for that variety. The brief names it, and the AI is prompted to write in it.
- A dialect-consistency check. It flags drift back into MSA, words from another dialect, and mixed registers.
- The report names the variety it was written in.
When a linguist edits: AI post-editing (AIPE)
Human review is an add-on, per content type, when you ask for it. At + Human edit and above, a native linguist edits a file the cycle has already checked, challenged, and judged: that is AI post-editing (AIPE).
What a managed AI localization program should include
Whoever you work with, ask for these. Each one is part of ours.
- Your glossary and style rules applied to every file, with a check that flags departures from them.
- Separate checks, a challenge to every finding, and a recorded ruling, so no finding rests on a single agent’s word.
- A quality gate on the delivered file. A file that fails it doesn’t ship.
- A report on every file that states the level that ran.
- For Arabic, the variety fixed at intake and named in the report.
- Human review as an add-on, per content type, when you ask for it.
- A route for errors you find after delivery. We correct language errors in the delivered translation free of charge, at every level, and each confirmed lesson goes into your glossary or rules. Your quote or agreement sets the period for reporting errors and how we settle a disagreement about whether something is an error.
- A clear answer on AI training (see the FAQ below).
Questions about AI localization
Do you translate legal and medical content with AI?
Yes. Legal and medical files run the full AI cycle, like every other domain. The report records each finding and ruling, and states which level ran.
Can different content run at different levels?
Yes. You set the level for each content type, and your quote records it. Choose your level in four questions.
Which languages do you work in?
Arabic (MSA and six dialect groups), English, French, Spanish (Spain), Spanish (Latin America), German, Chinese (Simplified), Turkish, Portuguese (Brazil), Italian, and Korean. Each language pairs with English in both directions, and Arabic pairs with each of the others. See all language pairs.
How do you handle large programs?
From one large batch to an ongoing program, release after release, the same seven steps run on every file, and each file comes back with its report. Tell us how much content you have, and whether it’s a one-off batch or ongoing updates. Your quote sets the scope.
What if we find an error after delivery?
We correct language errors you find in a delivered file free, at every level, and each confirmed lesson goes into your glossary or rules: see how free correction works.
Is our content used to train AI models?
No. We don’t use your content to train or tune our own AI tools, and when we use AI on it, we use only business AI services whose terms bar training on it. See how we handle your content.