OnlineDocTranslatorGuides

Running your practice

Localization Workflow: How Translators and Agencies Wire the Production Pipeline

A localization workflow routes source strings through intake, AI first-draft, post-edit, review, LQA, and certified sign-off — using connectors, TM, and QA

Inside the production system a translation agency runs, the workflow is the routing diagram that moves source content from a client repository to a signed, delivered target. This page is the production-pipeline deep dive under How App Localization Works and What Clients Are Buying, aimed at translators and ops leads wiring the pipeline themselves.

How it works

What is a localization workflow in a translation agency's production system?

1

Lane-aware routing — regulated, commercial and certified work travel through separate templates inside one TMS, each with its own step sequence and reviewer pool.

2

Priced stages — each step (translation, review, LQA, PM) carries its own rate line and is tracked independently so margin analysis is clean.

3

Connector topology — strings move repo-to-TMS automatically on commit or content publish, not by email attachment or manual XLIFF export.

4

Auditable sign-off — the certified lane ends with a named, credentialed translator's signature on the certificate of accuracy, with notarization available as an optional line item.

How it works

What are the core stages of a localization workflow end to end?

1

How does intake and scoping open the workflow?

Intake opens the workflow by registering source files, locale list, deadlines, and lane assignment — regulated versus commercial versus certified. The TMS applies the matching workflow template before any segment moves. Scoping also sets the TM leverage report, the engine selection, and the QA profile that will fire on every string change.

2

Where does pseudo-localization sit as a build-step checkpoint?

Pseudo-localization sits before translator time is spent. It is a build-step that surfaces i18n defects — hard-coded strings, truncation, placeholder breakage — so engineers fix the source before translation cost accrues. The build runs with inflated pseudo-strings that mimic length expansion in German or French.

3

How do TM leverage and AI pre-translate feed the post-edit stage?

TM leverage pre-fills 100% and high-fuzzy segments. The AI engine drafts the rest. Both route to the post-edit queue where a human translator corrects against the same term base and style guide. Raw machine output is never shipped — see AI Localization: How Translators and Agencies Run the First-Draft + Human-Certification Workflow for the full first-draft + human-certification pattern.

4

What does review add on top of post-edit?

Review is a second-pair-of-eyes bilingual pass at $0.04–$0.08 per source word that checks every segment against source. It is distinct from LQA, which only samples output against a scorecard. Review catches accuracy and completeness defects segment by segment. LQA scores the output population statistically.

5

How does certification close the workflow for regulated and legal content?

Certification closes the workflow on the certified lane: a named credentialed translator signs the certificate of accuracy on every page. Notarization is added as an optional $10–$50 line item when the receiving authority — USCIS, a US state court, a university registrar — requires it. The signed artifact is attached to the delivery back through the connector.

How it works

How is the localization workflow wired between client systems and the TMS?

1

Which repository and CMS connectors sit at the front of the pipeline?

Connectors sit between the client's Git repository, headless CMS, or product strings store and the TMS. The common integrations include:

  • GitHub Actions and GitLab CI for repo-to-TMS sync on commit.
  • Contentful and other headless CMS connectors for marketing content.
  • Framework-specific pipelines for React i18n, Shopify, Webflow, and Pendo — see React i18n: How Translators Fit Into a React Localization Pipeline.
  • XLIFF and JSON import/export for teams without a live connector.

2

How do TM leverage rules decide what skips translation entirely?

TM leverage rules route segments by match percentage:

  1. 100% matches route straight to review.
  2. 85–99% fuzzy matches route to discounted post-edit.
  3. No-matches route to full translation or AI pre-translate followed by post-edit.

Penalty rules downgrade matches that come from the wrong domain, so a marketing match does not fill a regulated segment.

3

Why is the glossary and style guide a workflow asset, not a document?

The glossary and style guide live inside the TMS as queryable assets. The AI engine, the QA checker, and the human post-editor all read from them during the same job. Terminology consistency survives across locales and release cycles because no one is reading a PDF in a separate tab.

How it works

How do traditional, agile and automated localization workflows differ?

1

What does a batch waterfall workflow look like now?

A batch workflow collects a finished source build, exports XLIFF or JSON once, routes it through translation, review and LQA on a two-to-six-week cycle, and reintegrates it at a release milestone. It is still common for documentation, legal content, and certified lanes where auditability outranks cadence. Translation memory leverage rules apply fully in a batch model: 100% matches go straight to review, 85–99% fuzzy matches attract a discounted post-edit rate, and no-match segments go to full translation or AI pre-translate. The batch model is predictable and auditable, but its fixed cycle length makes it a poor fit for product teams shipping weekly.

2

How does agile localization change the handoff cadence?

Agile localization moves strings on commit instead of on milestone. Review and LQA happen inside the same two-week cadence as the engineering team, which means the TMS must handle micro-jobs — diff-sized batches of new or changed segments — rather than monolithic file packages. The SLA clock on each lane template is calibrated to sprint length, not release length. Missing-key detection can be configured to automatically create translation jobs the moment a new string is added to the source repository, keeping the pipeline continuously populated without PM intervention.

3

What does a fully automated AI-first workflow add?

A fully automated workflow adds an AI first-draft stage ahead of post-edit. Quality scores are assigned at segment level, so each string carries a confidence signal that the TMS uses to route it: high-confidence segments go to lightweight review, low-confidence segments go to full post-edit or human translation. This routing logic is what makes AI pre-translate safe to operate at scale — the engine's output is never shipped raw. QA checks fire automatically on every string change. Human time stays on review, LQA, and the certification sign-off, compressing turnaround from days to hours without removing the human accountability layer.

How it works

What tool stack runs underneath the localization workflow?

1

What does the translation management system own?

The TMS owns workflow routing, vendor orchestration, engine selection, TM and term base storage, and reporting. It is the system of record that every other tool reads from and writes back to. A string that enters the TMS never leaves it until delivery — it only gets handed to the CAT editor and the QA tool through API calls.

2

What does the CAT editor own?

The CAT editor owns the segment-level production surface where the post-editor works. The screen shows source, pre-translated target, TM matches, term base hits, and inline QA warnings in the same view. The post-editor never alt-tabs to look up terminology.

3

Where do AI first-draft engines plug into the workflow?

AI first-draft engines plug in as a pre-translate step feeding post-edit, not as the shipped output. They sit as one of four tool categories alongside CAT, TMS and QA tools. Raw machine translation output is never shipped. Pseudo-localization sits earlier in the workflow as a checkpoint that surfaces i18n defects before translator time is spent.

4

What are the priced stages of the workflow on a US agency P&L?

US agencies price the stages as follows: translation at $0.10–$0.25 per source word, review at $0.04–$0.08 per word, LQA at $40–$90 per 1,000 words, and project management at 10–15% of project value. AI pre-translate shifts the mix toward review and LQA rather than translation.

StageRate basisUS range
TranslationPer source word$0.10–$0.25
ReviewPer source word$0.04–$0.08
LQAPer 1,000 words sampled$40–$90
Project management% of project value10–15%
Notarization (certified lane)Per document$10–$50

Pricing

How do workflow templates separate lanes inside one TMS?

Each workflow lane gets its own template — a defined step sequence, a dedicated reviewer pool, a calibrated SLA clock, and a QA profile — while sharing the program-wide term base and translation memory across all lanes. This architecture means regulated, commercial, and certified work run through a single production system without cross-contaminating reviewers, boilerplate text, or QA thresholds. A change to the shared term base propagates immediately to every lane; a change to one lane's QA profile has zero effect on any other.

The TMS enforces lane boundaries at the routing level, not through manual PM decisions. When a job enters intake, the TMS reads content-type metadata — file format, subject domain, client tag, or regulatory flag — and assigns it to the correct template automatically. This removes the most common source of lane contamination: a PM manually selecting the wrong workflow under deadline pressure.

Certified lane — routes to a named credentialed translator who signs a certificate of accuracy on every page; notarization is an optional line item priced at

Which quality gates run inside the localization workflow before release?

1

How is the LQA gate priced and scored?

The LQA gate samples translated output at $40–$90 per 1,000 words and scores defects against an MQM-style typology aligned to ISO 17100. The scorecard covers accuracy, fluency, terminology, locale convention, and style. LQA is a blocking gate before client delivery — the file does not ship until the score clears the threshold.

2

What triggers an automated QA check on every string change?

Automated QA checks fire on every commit or segment save. They cover:

  • Placeholder integrity ({name}, %s, ICU message format).
  • Number and date format per locale.
  • Tag integrity in HTML, XML, and Markdown segments.
  • Length limits for UI strings.
  • Glossary adherence against the term base.

Human reviewers only see substantive linguistic issues because the automated layer has already caught the mechanical ones.

How do agencies keep source and translated strings in sync when the source changes often?

Agencies keep source and target in sync by moving strings out of the repository and into a TMS of record via a persistent connector, syncing on every commit, and routing changed segments to a diff-based micro-job queue. Translators see only the delta — the net-new or modified segments — not the full file. They are never re-translating content that has already been approved and stored in TM.

The connector is the critical component: it watches the source repository or CMS for changes, extracts only the affected strings, packages them as a micro-job, and places that job into the correct lane template inside the TMS. New source content is automatically detected and extracted from connected CMS platforms — including common systems used for web, documentation, and marketing content — with no manual file exports required. This detection-to-job creation step is what eliminates the version-control failures that plague email-based handoff workflows.

Continuous localization uses this exact pattern: repo-to-TMS on every commit, micro-jobs for diffs, review and LQA completed inside the sprint cadence. Missing-key handling can be configured to fire automatically the moment a new string appears in the source, so the translation pipeline is always populated in near-real time. Version-locked builds — where a specific source tag is pinned to a specific translated output — are maintained in the TMS so that rollbacks on the engineering side do not create orphaned translated segments.

CI/CD automation via version-control pipeline integrations further reduces manual work while preserving the review gate before any translated string is deployed to production. The result is a workflow where source velocity and translation velocity stay matched without increasing PM headcount.

How it works

What are the biggest bottlenecks in a localization workflow and how are they fixed?

1

How do context packs cut reviewer back-and-forth?

Context packs — a screenshot, developer comment, character limit, or UI location attached to every segment at intake — cut reviewer back-and-forth by giving post-editors and in-market reviewers the information they need to resolve ambiguity inside the CAT editor. Without context, a single ambiguous string can generate a clarification chain that adds three to five business days to a sprint. With a context pack attached at intake, that same string is resolved on first pass. The CAT editor surfaces the context alongside the source, the pre-translated target, TM matches, term base hits, and QA warnings — all in one view — so the reviewer never needs to leave the tool to find an answer.

2

How does a connector eliminate manual file handoffs?

A connector eliminates manual file handoffs by piping strings repo-to-TMS automatically on commit or content publish. PM time shifts from zipping XLIFFs and tracking email attachments to exception handling on the jobs that fall outside normal routing. Translators never touch the source repository. Engineers never attach files to a ticket. The pipeline runs on commits, not on scheduled exports, which means delays introduced by timezone gaps and forgotten handoffs disappear from the critical path.

3

How does TM fragmentation slow a program and how is it fixed?

TM fragmentation occurs when different lanes, clients, or file formats accumulate separate translation memories that are never reconciled. Leverage rates drop because a perfect match stored in one TM is invisible to jobs running against a different TM. The fix is a shared, program-wide TM maintained in the TMS as the single source of truth: 100% matches route to review at no translation cost, 85–99% fuzzy matches attract a discounted post-edit rate, and no-match segments are the only strings that consume full translation or AI pre-translate budget. Consolidating TM into one managed store can reduce net translatable word counts by 20–40% on mature programs.

4

Why does pseudo-localization belong at the front of the workflow, not the back?

Pseudo-localization is a build-step checkpoint positioned at sprint zero — before any translator time is spent — that surfaces internationalization defects such as hardcoded strings, truncated UI containers, and encoding failures. Finding these defects in pseudo-loc costs nothing beyond a CI pipeline run; finding them after translated content has been delivered requires re-translation, re-review, and re-delivery. Moving pseudo-loc to the front of the workflow is one of the highest-leverage fixes available to a team experiencing repeated post-delivery defect cycles.

What economics does a mid-size localization workflow carry?

A mid-size multi-locale program covering five languages and 50,000 source words budgets at

A mid-size multi-locale program covering 5 languages and 50,000 words budgets at $25,000–$75,000 end to end. Translation runs $0.10–$0.25 per word, review $0.04–$0.08 per word, LQA $40–$90 per 1,000 words, and PM at 10–15% of project value. AI pre-translate shifts the mix toward review and LQA rather than translation. See Localisation Strategy: How Agencies Plan, Price, and Deliver Multi-Locale Programs for the planning view on this pricing basis.

Line itemRateVolume (5 × 50k)Range
Translation$0.10–$0.25 / word250,000 words$25,000–$62,500
Review$0.04–$0.08 / word250,000 words$10,000–$20,000
LQA$40–$90 / 1,000 wordsSampled$2,000–$8,000
PM overhead10–15%Of project value$3,700–$13,500

5,000–$75,000 end to end before AI leverage adjustments. Translation is the largest line item at $0.10–$0.25 per source word on a US agency P&L. Review adds $0.04–$0.08 per source word and covers every segment checked against source. LQA is priced at $40–$90 per 1,000 words translated and is scored against an MQM-style scorecard aligned to ISO 17100. PM overhead runs 10–15% of total project value and covers vendor orchestration, SLA tracking, and delivery.

Line itemRateVolume (5 × 50k)Range Translation$0.10–$0.25 / word250,000 words

A mid-size multi-locale program covering 5 languages and 50,000 words budgets at $25,000–$75,000 end to end. Translation runs $0.10–$0.25 per word, review $0.04–$0.08 per word, LQA $40–$90 per 1,000 words, and PM at 10–15% of project value. AI pre-translate shifts the mix toward review and LQA rather than translation. See Localisation Strategy: How Agencies Plan, Price, and Deliver Multi-Locale Programs for the planning view on this pricing basis.

Line itemRateVolume (5 × 50k)Range
Translation$0.10–$0.25 / word250,000 words$25,000–$62,500
Review$0.04–$0.08 / word250,000 words$10,000–$20,000
LQA$40–$90 / 1,000 wordsSampled$2,000–$8,000
PM overhead10–15%Of project value$3,700–$13,500

5,000–$62,500

Review$0.04–$0.08 / word250,000 words

Frequently asked questions

What best practices make AI-powered localization workflows reliable at scale?

AI-first pipelines require deliberate guardrails configured before the first string is processed: curated TM priming, language-pair-specific engine selection, confidence-score thresholds that route low-certainty segments to human post-edit, and locked terminology enforcement via integrated glossaries. Without these guardrails, engine output quality varies unpredictably across language pairs and content types, and QA failures surface late in the pipeline where they are expensive to remediate.

Quality scores assigned at segment level are the routing mechanism that makes AI pre-translate safe at production scale. Each segment receives a confidence signal from the engine; the TMS uses that signal to decide whether the segment goes to lightweight review, full post-edit, or human translation. This per-segment routing logic means the workflow adapts automatically to new content types and language pairs without manual PM intervention — low-confidence outputs are caught before they reach a reviewer's queue as if they were approved.

  • Curated TM priming — load the engine with domain-matched translation memory before the first job runs so that style and terminology consistency start high, not after a correction cycle.
  • Pair-specific engine selection — assign the best-performing engine per language pair rather than applying one global default; quality gaps between pairs are significant and consistent.
  • Confidence-score routing — segments below the defined threshold bypass lightweight review and go to full post-edit or human translation; the threshold is tuned per content type.
  • Locked terminology — the glossary enforces approved terms at engine generation time, not at review time, so term errors do not compound through the post-edit stage.
  • MQM LQA gate — AI-produced output is held to the same MQM-style scorecard as human output; the acceptance threshold is fewer than two errors per 1,000 words before the job is cleared for delivery.
  • Human certification intact — on certified lanes, a named credentialed translator still signs the certificate of accuracy on every page regardless of how much AI pre-translate contributed to the draft.
  • Style guide with disclosure clause — the program style guide documents that AI pre-translate is used in the workflow; this clause protects the agency and aligns client expectations on content that may carry regulatory or legal weight.

Evaluating AI performance systematically — measuring time saved, cost per word, and quality score across content types and language pairs over time — is what moves a pipeline from experimental to production-grade. Organizations that track these metrics together can make confident, business-aligned decisions about which content types benefit from full automation and which require heavier human involvement.

What tools do teams actually use to manage localization workflows end-to-end?

Most professional localization teams combine a Translation Management System (TMS) with CAT tools, version control, and API connectors to orchestrate the full pipeline. Common TMS platforms include Phrase, Lokalise, Crowdin, and memoQ Server; CAT tools like SDL Trados, memoQ, and Déjà Vu handle segment-level translation and leverage translation memory. Agencies handling certified translation production typically layer in QA tools (Xbench, Verifika) and project management software to track deadlines, reviewer assignments, and certification sign-off. Integration via API or file connectors (XLIFF, JSON, PO) keeps content moving without manual re-keying.

How do you handle localization for agile/continuous deployment pipelines?

Agile localization requires treating translation as a continuous process rather than a waterfall phase bolted on at the end of a sprint. Teams typically integrate a TMS directly with their CI/CD pipeline — pushing new or changed source strings automatically on each merge to main, triggering translation jobs, and pulling completed strings back before the next release. String freezes are replaced by rolling translation cycles, often 24–72 hours per sprint increment. Certified translation workflows add a compliance gate: machine-generated first drafts are routed to a credentialed translator for review and sign-off before the string is marked release-ready.

What's the best way to manage translation memory and glossaries across projects?

Centralizing translation memory (TM) and glossaries at the agency or organization level — rather than keeping them per project — is the single highest-leverage practice for consistency and cost reduction. A shared TM means every 100% or fuzzy match from a previous project cuts both turnaround time and per-word cost; industry benchmarks for repetition discounts range from 30% to 100% depending on match quality. Glossaries should be maintained in a structured termbase (e.g., MultiTerm or a TMS-native term database), enforced by QA checks, and versioned alongside source content so certified translators can audit which term set was active at sign-off.

How do you keep source strings and translated strings in sync when the source changes frequently?

String-level change tracking and diff-based updates are the standard solution for keeping source and target content in sync in fast-moving projects. A TMS compares incoming source files against the existing TM and flags only new, modified, or deleted segments — so translators touch only what changed rather than re-translating the full file. Setting clear segment status rules (e.g., a 95–99% fuzzy match triggers human review, not full retranslation) limits rework. For certified translation production specifically, any change to a previously certified string must invalidate its certification status and route back through the review and sign-off chain, even if the change is minor.

What are the biggest bottlenecks in a localization workflow for a small team?

For small localization teams, the most common bottlenecks are reviewer availability, manual file handling, and lack of automation between tools. A single senior translator or project manager who handles both translation and certification review creates a single point of failure that stalls entire batches. Manual export-translate-import cycles add hours of non-billable overhead per project. Investing early in TMS automation, standardized file formats (XLIFF is the safest interchange format), and — for certified work — first-draft generation tools dramatically reduces per-project turnaround without requiring headcount growth. Even small agencies report cutting administrative overhead by 30–50% after automating file routing.

what are localization workflows

A localization workflow is the ordered sequence of steps — from source content extraction through translation, review, QA, and delivery — that transforms source material into a linguistically and culturally adapted target version. For certified translation production, the workflow also includes credentialed translator assignment, certification sign-off, and an audit trail proving chain of custody. Typical stages are: content export → TM leverage → translation (human or MT+post-edit) → linguistic QA → subject-matter or legal review → certification → delivery. Each stage has defined inputs, outputs, and responsible roles, allowing project managers to track progress and SLA compliance at a granular level.

what is localization workflow

A localization workflow is the structured, repeatable process that moves content from a source language through every production step until a finished, quality-assured translation is delivered. It defines who does what, in what order, with which tools, and against which quality standards — covering file prep, translation, editing, QA, and (for certified work) formal sign-off by a credentialed professional. Well-designed workflows eliminate ad hoc handoffs, enforce consistency through shared TM and glossaries, and create the audit records required for certified or legally binding translations. The goal is predictable quality at scale, regardless of team size or language pair.

Related