Services / AI Translation Review

AI translation review:
error checking and scoring

Your AI translates. We check the output across several models, correct and score it, and the account's named owner signs it off on behalf of Translia — ISO 18587-certified post-editing and evaluation, on demand, with no technology bundle attached.

Start With a Pilot →
The real problem
AI translation is in production.
The quality layer usually isn't.
Most teams now generate translations with MT engines, LLMs, or an internal AI solution — and the output is often acceptable. But acceptable is not defensible: product names drift, tone misses the market, and errors ship silently until a customer notices. What's missing is not better AI; it is a structured human layer that reviews, corrects, and scores the output before it goes live.
Translia provides error checking and scoring of AI-generated translations: every batch is checked across several models, reviewed for the errors both machines and people miss, and signed off by the account's named owner on behalf of Translia. ISO 18587-certified post-editing and LQA against MQM/DQF-based or client-defined error frameworks, across 30+ language pairs in regular production (100+ languages supported), on demand, with no platform fees and no minimums. We review AI output; we don't sell AI translation technology.
What's included
The human layer, three ways

Post-editing (MTPE)

Full post-editing to publishable quality, or light post-editing for understandability — performed by market-qualified linguists under ISO 18587, in your file formats and pipeline.

Quality evaluation (LQA)

Structured scoring of AI output against an error framework (accuracy, fluency, terminology, style, locale conventions) with severity weighting, per-batch reports, and trend tracking over time.

Your framework or ours

Default framework based on MQM/DQF categories; if you run your own error typology, we adopt it as-is and report in your format, so results feed your existing dashboards.

Human layer vs. technology bundle
What you're actually buying —
and what you're not
 Platform-bundled reviewTranslia human layer
Works with your existing AIOften requires their engine or platformYes — any MT, LLM, or internal solution
Platform / subscription feeUsuallyNone — pay for reviewed volume only
Minimum volumesCommonNone — fully on-demand
Error frameworkFixed to their toolingMQM/DQF-based or your own, adopted as-is
CertificationVariesISO 18587 (post-editing) + ISO 17100
Lock-inPipeline migrates to themNone — we plug into your pipeline
Choosing a level
Light or full post-editing —
and how to tell which you need

ISO 18587 defines two post-editing levels, and the difference is not effort but the promise made about the output. Light post-editing delivers text that is accurate and comprehensible but not stylistically polished — errors of meaning are fixed, awkward phrasing is left alone. Full post-editing delivers text that reads as if written in the language, with terminology, tone and local convention brought to publishable standard. Choosing the wrong level is the most common way MTPE budgets get wasted: paying full rates for content nobody reads closely, or shipping light-edited text onto a product page.

 Light post-editingFull post-editing
The promiseAccurate and understandableReads as if written by a human translator
Always fixedMistranslation, omission, added meaning, offensive or unsafe wordingEverything in light, plus terminology, tone, register and local convention
Left aloneClumsy but correct phrasing, inconsistent styleNothing that a reader would notice
Typical useInternal knowledge bases, support tickets, high-volume listings, gistingProduct pages, marketing, UI strings, regulated and legal text
Publishable to customersUsually notYes
Relative costLowerHigher — often 2 to 3 times light

A single content set often needs both. We scope the level per content type rather than per project, so high-volume material is not subsidising the pages that actually convert.

The deliverable
What comes back with every batch

The edited files

Your original format, your file structure — TMX, XLIFF, JSON, CSV or whatever your pipeline expects. No repackaging on your side.

A scored report

Every error logged against the agreed framework with a severity weight, so a batch has a number, not an impression. Comparable across languages and across weeks.

Per-file traceability

Which reviewer, which pass, which decisions. When someone questions a string three months later, the record exists.

A trend line

Recurring error patterns across batches — the signal that tells you whether to fix the engine, the glossary, or the source content.

Why Translia
Built for teams that already chose their AI

Human layer only

No engine to adopt, no platform to migrate to. We review the output of whatever system you run (including your internal AI solution) and hand results back into your pipeline.

Framework discipline

Every review runs against a defined error framework with severity weighting, so quality is measured, comparable across batches and languages, and auditable — not a matter of reviewer taste.

Asian-language depth

Beijing and Hong Kong teams with particular strength in Asian languages and variants — the pairs where generic review networks are thinnest and AI output needs the most correction.

Common questions
AI translation review,
answered

What is error checking and scoring of AI translation?

We check and correct translations produced by MT engines, LLMs, or your internal AI solution: every batch is checked across several models, reviewed for the errors both machines and people miss, and signed off by the account's named owner on behalf of Translia. Scope runs from full post-editing to publishable quality, to light post-editing, to structured evaluation (LQA) against a defined error framework. Translia is ISO 18587 certified for post-editing.

What error framework do you use?

Our default is based on MQM/DQF error categories (accuracy, fluency, terminology, style, locale conventions) with severity weighting and per-batch scoring. Already run your own typology? We adopt it as-is and report in your format, so results feed your existing dashboards.

Do you sell AI translation technology?

No. We provide the human layer only. There is no technology bundle, no platform to adopt, and no lock-in — we plug into your existing pipeline and file formats, whatever AI produces the translations.

Which languages and content types do you cover?

30+ language pairs in regular production and 100+ languages supported, with particular depth in Asian languages and variants alongside major European languages. Typical content: e-commerce and product content, marketing copy, support content, technical documentation, and UI strings.

How do you price it?

On demand, with no platform fees and no minimum volumes. Priced per batch or deliverable, with the scope agreed before work starts — not capacity, seats, or subscriptions.

How fast is turnaround?

Agreed per content type and volume: recurring streams run on fixed cycles with defined SLAs, and capacity is scheduled per language and volume. Tell us your daily or weekly volume by language and we'll propose a turnaround commitment.

Can we test before committing?

Yes. Send a sample of your AI-translated content and we will agree a small pilot on one real file, so you see the review quality, report format and turnaround before committing.

What is the difference between light and full post-editing?

ISO 18587 defines both. Light post-editing produces text that is accurate and comprehensible: errors of meaning, omissions and unsafe wording are corrected, while clumsy-but-correct phrasing is left as the engine produced it. Full post-editing produces text that reads as if written in the language, with terminology, tone, register and local convention brought to publishable standard. Light is appropriate for internal knowledge bases, support tickets and high-volume listings; full is appropriate for anything a customer reads as a representation of the brand.

How do you decide which level a batch needs?

By content type, not by project. We look at who reads the content, whether it carries brand or legal weight, and what happens if a sentence reads awkwardly but correctly. Most clients end up with a mixed scope: full post-editing on product and marketing pages, light on support and internal material. Scoping per content type rather than per project keeps high-volume text from consuming the budget the converting pages need.

What does the quality report actually contain?

Each error logged against the agreed framework with a category and severity weight, a batch-level score derived from those weights, per-file traceability showing reviewer and decisions, and a trend view across batches. The trend is the part clients use most: recurring error patterns tell you whether the fix belongs in the engine, the glossary, or the source content, rather than in more review.

Is our content used to train AI models?

No. We process client content only through enterprise API tiers configured with zero data retention (ZDR) and model-training opt-out — under those terms the content is not retained by the provider and is not used for model training. We never put client content through free or consumer AI tools, every file stays under our standard confidentiality terms, and the last step is always sign-off by the account's named owner on behalf of Translia.

On request we also disclose the specific engines used under NDA and sign your own Data Processing Agreement (DPA).

Further reading
Before you scale AI translation
What AI translation still gets wrong — and where exactly it fails → What second-layer review catches after standard QA → The orchestration layer: what sits between your AI and market-ready output →

Running AI translation without a human layer?

Send one real file and we'll agree a small pilot, so you can judge the difference before committing.

Start With a Pilot →