The system
The pipeline that keeps enterprise translation consistent at volume
The model is one stage of five. The stages around it are what distinguish an enterprise translation system from pasting text into a public tool.
01
Classify by risk
Each content class — support article, product page, contract, safety notice — is assigned a treatment lane before any translation happens.
- Content classes
- Routing policy
02
Apply memory and glossary
Previously approved translations are reused from translation memory, and your glossary locks how names, titles and technical terms render — before the model generates anything.
- Translation memory
- Glossary
03
Machine translate the remainder
The engine translates what memory did not cover, constrained by the glossary, with engines chosen and tested on your own content rather than assumed.
- MT engine
04
Post-edit to the lane's bar
Human editors review machine output to the depth the lane requires — light review for low-risk content, full revision for customer-facing pages — and their fixes flow back into memory.
- Post-editing
- Reviewer roles
05
Evaluate beyond the score
Automated metrics screen for regressions, but quality is judged by structured human review on sampled output — adequacy, terminology, register — because a fluent translation can still say the wrong thing.
- Sampled review
- Quality criteria
The human-only lane
Contracts, regulatory submissions, safety-critical instructions and anything with legal effect bypass the machine lane entirely and go to qualified human translators — with the memory and glossary still applied, so even the human lane stays consistent with the rest.
The routing policy, not the engine, is what makes the system safe to scale.
The boundary
Where machine translation earns trust and where it must not be used alone
Between Arabic and English, machine translation has become genuinely usable for content whose cost of an occasional awkward sentence is low: knowledge-base articles, internal announcements, product descriptions, first drafts that a bilingual editor will polish. In those lanes it changes the economics of operating bilingually — content that would have waited weeks for a translation budget ships in both languages as a matter of course.
The same technology is not acceptable alone where words carry legal or physical consequence. A contract's governing-language clause, a regulator's required wording, a safety instruction on industrial equipment — an error there is not an awkward sentence but a liability, and machine fluency makes such errors harder to spot, not easier. Our designs are explicit about this boundary: those classes route to qualified human translators, and no throughput target is allowed to move them. Being clear about where the machine stops is what lets you use it confidently everywhere else.
A typical engagement
What a bilingual content operation looks like once the lanes exist
Illustrative engagement, not a named client
An organisation publishes policies, product content and customer help in English first, with Arabic following weeks later through ad-hoc agency requests. Terminology drifts between documents, the same sentences are re-translated and re-billed, and legal texts sit in the same queue as blog posts.
- 01Content classes are inventoried and each is assigned a lane — machine with light review, machine with full post-editing, or human-only — with the legal and safety classes fixed in the human lane by policy.
- 02A bilingual glossary of names, titles and technical terms is built with your reviewers and enforced across every lane, and existing approved translations are loaded into translation memory so past spend keeps paying.
- 03The pipeline is wired into the content systems — CMS, knowledge base, document management — so translation is triggered by publishing events rather than by someone remembering to email a file.
Arabic and English publish together for routine content, terminology stops drifting because one glossary governs every lane, and human translation budget concentrates on the documents where only human judgement is acceptable.
Our role
What we build around the translation models and what stays with your editors
- 01
The architecture and the routing
We design the lanes, test candidate engines on your own content, and build the routing so no document reaches a lighter treatment than its risk class allows.
- 02
The terminology assets
We build the glossary and translation-memory infrastructure and connect them to every lane, machine and human alike, so consistency comes from the system rather than from individual memory.
- 03
The evaluation discipline
We set up the sampled human review — adequacy, terminology, register, direction-specific issues between Arabic and English — and the reporting that shows quality per lane over time, so the routing policy is revisited on evidence.
What your organisation is willing to publish at each quality bar is an editorial decision that stays with you; we build the lanes, the terminology control and the evidence that let your editors govern volume instead of chasing it.
Common questions
What buyers ask
Can machine translation be trusted with our contracts and legal documents?+
Not on its own, and our designs do not pretend otherwise. Legal text carries consequence in its exact wording, Arabic-English legal pairs are full of terms whose everyday translation differs from their legal one, and a machine's fluency hides its errors precisely where they cost most. In our architectures, contracts, regulatory submissions and anything with legal effect route to qualified human translators as a matter of policy — supported by the same glossary and translation memory, so the human lane stays consistent with everything else. Machine translation earns its keep in the high-volume, lower-risk lanes, which is where most of the volume actually is.
How do you evaluate translation quality beyond BLEU and automated scores?+
Automated metrics compare output to reference translations and are useful as regression alarms — if a score drops after an engine change, something moved. But they reward similarity to one reference, not correctness, and a translation can score well while mistranslating a name or missing the register your brand requires. So quality is decided by structured human review on sampled output: bilingual reviewers rate adequacy (is the meaning preserved), terminology (does it follow the glossary), register and direction-specific issues, against written criteria. The samples are drawn per lane and per content class, so you learn where the system is weak, not just whether it is weak somewhere.
What does a glossary change in an Arabic-English translation pipeline?+
It is the difference between a translation function and a pile of translations. Without one, your product names, job titles, regulatory terms and brand phrases render differently across documents, translators and years — three Arabic versions of one department name is a small confusion internally and a real one in a customer letter or a tender. The glossary fixes each term's approved rendering in both directions and is enforced in every lane: it constrains the machine engine, guides post-editors, and binds human translators. Combined with translation memory, which reuses whole approved sentences, it is why the tenth thousand pages come out consistent with the first.
In this cluster
Next step
Scope it with us
A translation architecture that names which of your content classes go through machine, post-edited and human lanes, and the terminology assets that keep all three consistent.

