On this page
The roles a hybrid model needsWhat "human" should meanHow work flowsThe loop that makes it improveStaffing by volumeA company announces that translation is now "AI plus human review". Six months later, Korean support articles are fluent and often wrong, the product UI has three names for one feature, and the two Korean reviewers say they spend their days fixing the same errors. The engine was fine. The model had an engine and reviewers, and none of the roles that connect them.
Most enterprise translation is now hybrid in some form: machines draft, humans edit, check and decide. The idea is simple. What makes it work or fail is the operating model around it: who does what, how content is routed, who owns quality, and how many people each role needs. This post lays that out, with Korean as the worked example.
The roles a hybrid model needs
| Role | What they do | Korean-specific notes |
|---|---|---|
| Localization owner | Owns the program: budget, vendors, routing rules, priorities | Often not a linguist; needs a Korean lead to rely on |
| Lead linguist (per language) | Owns quality: style sheet, termbase, scoring, final calls on disputes | Decides speech level, loanword policy, name and number conventions |
| Post-editors | Edit machine output to the agreed level | Must read the source, not just smooth the Korean |
| Translators and revisers | Translate high-risk content from scratch; revise each other's work | Contracts, regulated labels, marketing, patent text |
| Reviewers in context | Check text on screens, devices or final layouts | Catch line breaks, truncation, particle errors with variables |
| Localization engineer | Connectors, file handling, MT configuration, QA automation | Sets up Korean-specific QA checks and encoding |
| Domain or legal reviewer | Confirms meaning in specialist or regulated text | Usually part-time or external |
Small teams combine roles; one person can be lead linguist and reviser. The mistake is not combining them but leaving one out, most often the lead linguist, which leaves nobody to turn individual corrections into rules.
What "human" should mean
The human step is only as good as the person doing it. ISO 17100, the international standard for translation services, lists the competences a translator must have, as summarised by the Institute of Translation and Interpreting: translation competence; linguistic and textual competence in both languages; competence in research and information processing; cultural competence; technical competence; and domain competence. It also builds revision by a second person into the process. ISO 18587 sets requirements for post-editing machine output and for the competences of post-editors, and distinguishes light from full post-editing.
Those lists are a useful hiring check. A bilingual colleague may have strong language skills but no research habits or domain knowledge, and Korean post-editing in particular needs someone who reads every source segment, because fluent Korean output can hide a reversed meaning or a dropped negation.
How work flows
Routing decides which content gets which human step. A simple, explicit table beats a sophisticated one nobody follows:
- Critical content (legal, regulated, safety, prices): human translation and revision by a second linguist; machine output at most as a reference.
- Customer-facing content (UI, website, onboarding, marketing): machine draft with glossary, full post-editing, review in context; marketing rewritten by a Korean writer.
- Support and help content: machine draft with glossary; light or full post-editing depending on traffic; sampled quality checks.
- Internal content: raw or lightly edited machine output, with a way for readers to flag problems.
Quality estimation can refine this within a tier. Phrase's Quality Performance Score, for example, can be used with thresholds to decide which segments skip or receive human editing. The lead linguist should own those thresholds and check them against sampled human scores; see sampling-based and risk-based translation QA.
The loop that makes it improve
A hybrid model gets better only if corrections flow back. Every week, the lead linguist should look at what post-editors and reviewers changed and ask whether the fix belongs somewhere permanent:
Same term fixed again
Post-editors keep replacing the engine's word for a featureAdd the term to the termbase and the MT glossaryOne entry saves the same edit hundreds of times.Register drift
Engine output alternates between formal and polite conversational endingsSet the formality option or add example sentences to the engine or promptThen measure whether the edits drop.Ambiguous English
"Post" translated as mail, as a noun and as a verb in different stringsAsk writers to add context notes or rewrite the sourceSome errors are best fixed before translation.Staffing by volume
We will not give headcount ratios as if they were industry standards, because they depend on content mix and engine quality. A more useful approach is to staff by role coverage:
- Low volume, one or two releases a month: an external Korean lead linguist on retainer who also post-edits and revises, plus a second linguist for revision of critical content. Localization owner part-time in-house.
- Medium volume, continuous releases: a dedicated Korean lead (in-house or vendor-side), a small pool of post-editors, a localization engineer shared across languages, in-context review built into the release process.
- High volume, many content types: separate pools for marketing, product and support, each with a reviewer; the lead linguist spends most of their time on rules, data and disputes rather than on text.
Measure, then adjust: post-editing time per thousand words, edit distance, sampled error scores and Korean support tickets that quote UI text. When edit effort falls on a content type, move it to a lighter tier; when scores fall, move it back.
For the tools behind this model, see the translation tooling stack; for costs, a cost model of a hybrid AI pipeline; and for the trade-offs companies weigh when choosing, AI translation vs human translation.
Need the Korean half of a hybrid model?
We act as your Korean lead linguist and post-editing team: termbase, style sheet, routing advice and review, inside your tools. See our services →