Somebody still has to decide whether the output is fit to publish
Your engine translates. Your quality estimation flags what it is unsure of. Somebody still has to decide whether the output is fit to publish, sign off that the decision was made by someone qualified, and produce the evidence if anyone asks.
That is the whole of what we do.
What we do
Almost everything is machine-translated now. Almost nothing is independently checked.
The economics of translation changed in about three years. Engines got good, volumes went up, and the cost per word went down. That part worked.
What did not change is where the responsibility sits. When a mistranslated instruction for use reaches a regulator, when a claim in a patent narrows because a term drifted, when a contract clause reads differently in the target language than in the source — the machine is not the party that answers for it.
Someone has to be accountable for the final draft. That is the work we do.
The evidence that human review still matters
We did not translate it. That is what makes our review worth something.
An audit carried out by the party being audited is not evidence. It is the same principle everywhere else in business, and it applies here.
The largest providers in this industry now sell quality checking of AI output — while also selling the AI output. One of them markets its reviewers as agnostic and impartial. We can say the same thing without the conflict, because we do not sell the engine, the platform, or the first draft.
We have no reason to defend the output we are asked to assess, and no commercial interest in which engine you use. When we tell you a file is fit to publish, that judgement is worth something precisely because we had nothing to do with producing it.
That also means we can compare. If you have standardised on one engine across every language, there is a reasonable chance it is the wrong engine for at least one of them — and almost nobody has checked.
How we work
Every job runs the same way, and the process is the deliverable as much as the file is.
- Two stages, two people. Post-editing and review are separate passes. The person who edited the file is not the person who signs it off.
- Named reviewers with recorded qualifications. We record every linguist's qualifications, subject expertise and time in the domain, and we build that record before the work starts rather than after. For regulated content, ask for it and you will get it.
- Structured scoring. Errors are categorised by type and severity using MQM, the industry-standard error framework. You get a score, not an impression.
- A version trail. What changed, when, by whom, against which source version.
- A quality floor. If machine output falls below an agreed threshold, the segment is treated as translation and priced as translation. You should not absorb the cost of a bad engine day, and neither should we.
Where we are strongest
Start with one file
The fastest way to find out whether this is worth your time is to give us something real. Send a file you have already had machine-translated, in a language pair that matters to you, and we will return it post-edited with a scored error report attached.
You will see exactly what we found, how we categorised it, and what the process looks like from your side.
Sources
- WMT25 General MT shared task, Conference on Machine Translation, 2025. 30 language pairs, 60 systems, professional annotators marking errors.
- 2025 entity-translation benchmark covering names, brands and cultural references across major languages.
- Large-scale post-editing productivity study: 90 million words, 879 linguists, 2.5 years of production data. European language pairs only.