Answer
Continuous Localization: How Translators and Agencies Run the Delivery Model
Continuous localization pulls new strings from the repo on every merge, routes them through AI first-draft, review, and LQA, and ships locales hours behind
Continuous localization is the delivery model where translators and agencies receive commit-level string deltas from a client's repository through a TMS connector, process them as micro-jobs through AI first-draft, human review, and LQA, and return approved translations so the localized build trails the English source release by hours per merge. This page covers the production-floor mechanics an agency runs — the connector, the micro-job queue, the LQA gate against ISO 17100, the retained-capacity billing model, and the four triggers that force a project off waterfall — and sits inside our broader coverage of How App Localization Works and What Clients Are Buying.
Key points
- Continuous localization pulls new and changed strings from the client's repository on every merge, routes them through AI first-draft plus human review plus LQA, and ships localized builds hours behind source — making it the natural delivery model for any team running CI/CD pipelines.
- Four triggers force a move off waterfall: release cadence under two weeks, more than 500 new or changed strings per sprint, more than six active target locales, and in-app experimentation generating strings between releases.
- Agency pricing stacks translation at $0.10–$0.25 per source word, review at $0.04–$0.08 per word, and LQA at $40–$90 per 1,000 words translated against an MQM-style scorecard aligned to ISO 17100.
- Translation memory inside a continuous cycle trims per-word cost by 20–40%, and best practice pushes new strings to translators within 24–48 hours of string freeze — teams that apply this discipline have reported localization cycle time falling by more than 50%.
- A mid-size program of 5 languages across 50,000 words lands at
- Continuous localization pulls new and changed strings from the client's repository on every merge, routes them through AI first-draft plus human review plus LQA, and ships localized builds hours behind source.
- Four triggers force a move off waterfall: release cadence under two weeks, more than 500 new or changed strings per sprint, more than six active target locales, and in-app experimentation generating strings between releases.
- Agency pricing stacks translation at $0.10–$0.25 per source word, review at $0.04–$0.08 per word, and LQA at $40–$90 per 1,000 words translated against an MQM-style scorecard aligned to ISO 17100.
- Translation memory inside a continuous cycle trims per-word cost by 20–40%, and best practice pushes new strings to translators within 24–48 hours of string freeze.
- A mid-size program of 5 languages across 50,000 words lands at $25,000–$75,000 depending on TM leverage, turnaround speed, and certification requirements.
5,000–$75,000 depending on TM leverage, turnaround speed, and certification requirements.
- The PM role shifts entirely under continuous localization: instead of dispatching and tracking individual purchase orders, the PM focuses on exception handling and capacity planning across a standing translator pool.
- Programs operating at full TM leverage have demonstrated human localization cost reductions exceeding 40% while simultaneously expanding language coverage — proving that cost discipline and scale are not mutually exclusive in the continuous model.
What is continuous localization and how does it work?
Continuous localization is the delivery model where a TMS connector pulls new and changed strings from the client's repository on every merge, routes them through AI first-draft, human review, and LQA inside the sprint, and returns approved translations so the localized build trails the English release by hours. The connector is persistent, the job queue is diff-based, and the reviewer pool stays on standby for sprint-aligned sampling. This model has become the standard delivery approach for software teams that have adopted CI/CD pipelines, because a waterfall handoff cannot keep pace with a weekly or faster deploy cycle. Teams that operate this way have reported localization cycle time falling by more than 50% and publishing delays collapsing from one to two weeks down to near-real-time — without adding internal headcount.
What is the difference between continuous localization and adjacent delivery models?
Continuous localization is one of four related but distinct concepts that come up in software localization conversations. Understanding the differences matters because each model implies a different contract structure, a different tool stack, and a different day-to-day workflow for the PM and the translator pool.
| Term | What it is | When you offer it |
|---|---|---|
| Continuous localization | Connector-driven delivery mechanic that moves string deltas from repo to localized build within hours | Weekly or faster commit cadence with 500+ strings per sprint |
| Agile localization | Sprint-cycle discipline that integrates translation work into development planning | Any sprint-based engineering team, with or without a connector |
| Waterfall handoff | Single export per release, offline translation, re-import | Quarterly or longer release trains with batch sign-off |
| In-sprint translation | Translation work scheduled inside the same sprint as the strings' development | Teams running agile with a dedicated LSP contact per sprint |
What does the string flow look like from commit to localized build?
On every merge the connector detects changed strings, pushes them into the TMS as a micro-job, assigns the micro-job to the translator pool, runs AI first-draft and human review, applies the LQA sampling gate, and writes the approved target strings back to the build within hours. The webhook fires on commit or merge from GitHub or GitLab, and the diff — not the full file — enters the queue. This diff-based architecture is what keeps micro-job sizes manageable: translators receive segments sized in minutes, not multi-day batches. React, Vue, and native mobile codebases are typical commit sources; the same pipeline logic applies to SaaS products that continuously ship in-app copy and experiment strings between formal releases.
The string flow runs in six steps: the developer merges to the repo, the webhook fires to the TMS, the diff is detected, a micro-job is created, AI first-draft runs, and the human translator reviews the draft before LQA sampling. D[Micro-job created in TMS queue]
D --> E[AI first-draft runs]
E --> F[Human translator reviews draft]
F --> G{LQA sampling gate}
G -- Pass --> H[Approved strings pushed back to build]
G -- Fail --> F
H --> I[Localized build ships hours behind source]
Who touches a string inside a continuous lane?
Four roles touch each string: the connector (automated), the AI first-draft engine, the human translator reviewing the draft, and the LQA reviewer sampling the output against an MQM-style scorecard aligned to ISO 17100. The AI first-draft stage runs before human review on every micro-job, and the translator post-edits the draft rather than translating from scratch. This post-edit model is what makes the economics work: translation memory supplies matches, the AI fills gaps, and the human reviewer adds judgment — each role doing only what the previous layer does not. The LQA reviewer operates on a sampling basis calibrated to sprint volume, not on a full-file pass, which is what keeps the gate from becoming a bottleneck inside a high-frequency commit environment.
- Connector — detects the diff, creates the micro-job, carries metadata and context.
- AI first-draft engine — produces a seed translation leveraging TM and glossary.
- Human translator — reviews and post-edits every segment.
- LQA reviewer — samples output against the scorecard before build-back.
Why us
How much does translation memory reduce per-word cost in a continuous cycle?
Translation memory inside a continuous cycle reduces per-word costs by 20–40%, because repeated UI strings across sprints are not re-translated — only the diff is routed to the translator, and TM leverage is highest where micro-batches share vocabulary with prior merges. The glossary and style guide travel with the TM, and the AI first-draft engine pre-populates matches before the human translator touches the segment.
How does the retained-capacity billing model work?
Under continuous localization the agency bills retained monthly capacity or per-word with an SLA on turnaround, not one-off POs — the PM role shifts from job dispatch to exception handling and capacity planning, and the client secures guaranteed hours each sprint instead of waiting on a quote per release. Billing-model and PM-mode differences for multi-locale programs are documented on Localization Management: How Translators and Agencies Run Multi-Locale Programs.
When should a project move from waterfall to continuous localization?
A project moves to continuous localization when four triggers appear together: release cadence under two weeks, more than 500 new or changed strings per sprint, more than six active target locales, and in-app experimentation generating strings outside the main release train. One trigger alone does not justify the switch; the combination is what breaks file-based editing.
- Release cadence under two weeks. Waterfall handoff cannot keep pace with a bi-weekly or weekly release train.
- More than 500 new or changed strings per sprint. Manual export/import loses TM leverage at this volume.
- More than six active target locales. Coordination overhead across locales exceeds PM dispatch capacity.
- In-app experimentation generating strings between releases. A/B copy, feature flags, and in-product messaging spawn strings outside the main train.
What is the scaling threshold where an i18n-editor stops working?
An i18n-editor stops scaling at roughly 5,000 translation keys, four or more target locales, or any continuous-localization cadence that commits weekly — file-based editing loses TM leverage, has no reviewer gate, and cannot run the LQA sampling that certified lanes require under an MQM-style scorecard aligned to ISO 17100. See i18n-editor: What Translators Use to Edit JSON Translation Files for the scaling detail and when a TMS becomes mandatory.
How does continuous localization compare to traditional waterfall handoff?
Continuous localization moves string-level micro-batches through a repo connector on every commit and ships locales hours behind source; waterfall batches 5,000+ words per release cycle, exports once, translates offline, re-imports, and ships locales weeks or months behind source. The operational gap between the two models is not merely one of speed — it is a difference in how risk accumulates. Under waterfall, untranslated strings pile up invisibly between exports: a developer pushes a change on Monday, and by Thursday the French and German versions of that string may still be sitting in a manual export queue while the English build is already live. That latency compounds across every sprint, and by the time a release ships, the localized versions can be a full cycle behind.
| Dimension | Continuous localization | Waterfall handoff |
|---|---|---|
| Trigger | GitHub/GitLab webhook on commit or merge | Manual export at end of release cycle |
| Batch size | String-level diff; minutes per micro-job | 5,000+ words per release |
| Release trail | Hours behind source per merge | Weeks or months behind source |
| PM mode | Exception handling and capacity planning | Dispatch-and-track per PO |
| Billing | Retained monthly capacity or per-word with SLA | Per-PO per release |
| Reviewer pool | Sprint-aligned, on standby | Spun up per batch |
| TM leverage | High — 20–40% per-word reduction | Lower — fewer cross-sprint repeats captured |
| String freeze to translator SLA | 24–48 hours | Defined per PO, often days to weeks |
What changes operationally for the PM and the translator pool?
Operationally, the PM moves from dispatch-and-track to exception handling and capacity planning, translators work on micro-jobs sized in minutes rather than multi-day POs, and the reviewer pool stays on standby for sprint-aligned sampling instead of one-shot batch review. This restructuring changes how capacity is contracted. Under waterfall, the agency scales up for each PO and then scales back down. Under continuous localization, a standing pool is retained across sprints, shifting billing to retained monthly capacity or per-word with an SLA. Teams that make this transition reduce publishing delays from one to two weeks to near-real-time., demonstrating that the operational change is not cosmetic — it fundamentally alters the relationship between source release and localized release. Translators in a continuous lane must also adapt their working habits: context arrives as micro-jobs with metadata rather than as packaged XLIFF files, and the translator's throughput expectation is measured against sprint velocity rather than weekly word counts.
What quality gates run inside a continuous localization lane?
Two quality gates run inside a continuous lane: review at $0.04–$0.08 per source word that checks every segment, and LQA at $40–$90 per 1,000 words that samples output against an MQM-style scorecard aligned to ISO 17100, scoring accuracy, fluency, terminology, locale convention, and style. The LQA gate is blocking — a failed sample sends the batch back to review before the build is pushed.
How is the LQA scorecard structured?
The LQA scorecard aligned to ISO 17100 scores five MQM defect categories — accuracy, fluency, terminology, locale convention, and style — on sampled segments at $40–$90 per 1,000 words translated, and runs as a blocking gate before the localized build is pushed to release.
| Category | What it scores |
|---|---|
| Accuracy | Mistranslation, omission, addition against source meaning |
| Fluency | Grammar, spelling, punctuation in the target locale |
| Terminology | Adherence to the client glossary and TM entries |
| Locale convention | Dates, numbers, currency, address format for the target |
| Style | Register, tone, and style-guide compliance |
How tight is the string-freeze-to-translator SLA?
Best practice is to push new or changed strings to translators within 24–48 hours of a string freeze — any longer and the sprint cadence breaks, because the reviewer and LQA gates still need to run before the localized build can trail the source release by hours.
Pricing
How does an agency price a continuous localization program?
A continuous localization program stacks four US-dollar line items: 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 translated, and retained monthly capacity for PM and exception handling — with TM leverage trimming 20–40% off per-word spend across sprints. The retained capacity line item is the structural difference from waterfall pricing: because the translator pool and the PM are on standby between micro-jobs rather than spun up per project, that availability has a cost that sits outside the per-word rate. Programs that invest in TM leverage early — building glossary and TM while string counts are still manageable — have demonstrated human localization cost reductions exceeding 40% at scale, which means the up-front investment in TM infrastructure pays back quickly across a multi-sprint program.
| Line item | Unit | US price range |
|---|---|---|
| Translation (AI first-draft + human review) | Per source word | $0.10–$0.25 |
| Second-pass review | Per source word | $0.04–$0.08 |
| LQA sampling gate | Per 1,000 words translated | $40–$90 |
| Retained monthly capacity | PM + exception handling | Scoped per program |
| TM leverage adjustment | Deduction on per-word | −20% to −40% |
What does a mid-size five-language program actually cost?
A mid-size multi-locale program covering 5 languages and 50,000 words costs
A continuous localization program stacks four US-dollar line items: 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 translated, and retained monthly capacity for PM and exception handling — with TM leverage trimming 20–40% off per-word spend across sprints.
| Line item | Unit | US price range |
|---|---|---|
| Translation (AI first-draft + human review) | Per source word | $0.10–$0.25 |
| Second-pass review | Per source word | $0.04–$0.08 |
| LQA sampling gate | Per 1,000 words translated | $40–$90 |
| Retained monthly capacity | PM + exception handling | Scoped per program |
| TM leverage adjustment | Deduction on per-word | −20% to −40% |
What does a mid-size five-language program actually cost?
A mid-size multi-locale program covering 5 languages and 50,000 words costs $25,000–$75,000 depending on TM leverage, turnaround speed, and certification requirements — the floor applies to high-TM software strings, the ceiling to low-TM content with certified-lane LQA on every batch.
- Floor ($25,000) — repeat UI strings, 40% TM leverage, standard SLA, no certification.
- Mid-range ($40,000–$55,000) — mixed UI and help-content, 20–30% TM leverage, 24-hour SLA.
- Ceiling ($75,000) — low-TM new product copy, certified lane with LQA on every batch, same-day SLA.
5,000–$75,000 depending on TM leverage, turnaround speed, and certification requirements — the floor applies to high-TM software strings, the ceiling to low-TM content with certified-lane LQA on every batch. The spread is wide because TM leverage is the single largest variable: a program running at 40% TM match rates pays materially less per sprint than one introducing net-new marketing copy in every cycle. Agencies model three scenarios at proposal stage — high TM, mixed, and low TM — so the client understands the cost envelope before committing to a program structure.
Floor (
A continuous localization program stacks four US-dollar line items: 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 translated, and retained monthly capacity for PM and exception handling — with TM leverage trimming 20–40% off per-word spend across sprints.
Line item Unit US price range Translation (AI first-draft + human review) Per source word $0.10–$0.25 Second-pass review Per source word $0.04–$0.08 LQA sampling gate Per 1,000 words translated $40–$90 Retained monthly capacity PM + exception handling Scoped per program TM leverage adjustment Deduction on per-word −20% to −40% What does a mid-size five-language program actually cost?
A mid-size multi-locale program covering 5 languages and 50,000 words costs $25,000–$75,000 depending on TM leverage, turnaround speed, and certification requirements — the floor applies to high-TM software strings, the ceiling to low-TM content with certified-lane LQA on every batch.
- Floor ($25,000) — repeat UI strings, 40% TM leverage, standard SLA, no certification.
- Mid-range ($40,000–$55,000) — mixed UI and help-content, 20–30% TM leverage, 24-hour SLA.
- Ceiling ($75,000) — low-TM new product copy, certified lane with LQA on every batch, same-day SLA.
5,000) — repeat UI strings, 40% TM leverage, standard SLA, no certification.
- Mid-range ($40,000–$55,000) — mixed UI and help-content, 20–30% TM leverage, 24-hour SLA.
- Ceiling ($75,000) — low-TM new product copy, certified lane with LQA on every batch, same-day SLA.
The billing model itself is a differentiator when pitching continuous against waterfall: retained monthly capacity with a per-word SLA gives the client budget predictability and gives the agency stable revenue, whereas per-PO waterfall billing creates revenue peaks and troughs that make capacity planning difficult for both sides. Agencies that articulate this trade-off concretely are better positioned to close continuous engagements with software clients who are already managing recurring SaaS operating costs and prefer predictable line items.
Computer-Assisted Translation Tools: The Working Translator's Field GuideWhich tools and platforms power continuous localization in practice?
Continuous localization runs on three tool layers: a TMS of record wired to the repo through GitHub or GitLab webhooks, a translation memory and glossary stored centrally, and an LQA scorecard tool that runs the ISO 17100-aligned sampling gate before build-back. The TMS is the system of record; the connector is the trigger mechanism; the TM is the economic engine that drives the 20–40% per-word cost reduction that makes the model financially sustainable at high commit frequency. This stack is the direct localization counterpart to the CI/CD pipelines that software teams have adopted to accelerate delivery — just as CI/CD automates build and test, the TMS connector automates string extraction, job creation, and build-back.
When does a lightweight editor stop being sufficient?
A simple i18n editor or file-based workflow is sufficient up to approximately 5,000 keys, four target locales, or a weekly commit cadence — beyond any of those thresholds, a full TMS connector becomes operationally necessary. Below those limits, manual string exports and spreadsheet-based review keep pace; above them, the volume and frequency of micro-jobs exceeds what any manual handoff handles without introducing the same latency problem that continuous localization eliminates. Agencies advising early-stage software clients track these three scaling signals and recommend a TMS migration proactively rather than reactively.
- TMS of record — stores all translation memory, glossary, and job history; serves as the authoritative ledger for every string across every locale.
- Repo connector — GitHub or GitLab webhook on commit or merge, diff-based micro-job queue; fires automatically without PM intervention.
- Translation memory and glossary — stored centrally, consumed by AI first-draft and human review; primary driver of per-word cost reduction across sprints.
- LQA scorecard tool — MQM-style, five error categories, aligned to ISO 17100; runs on a sampling basis calibrated to sprint volume.
- CI/CD build-back integration — approved target strings written back into the pipeline for the next build, completing the automation loop without developer involvement.
How we run continuous localization lanes
We run continuous localization lanes where the client's repo is wired to a TMS via a webhook, and we deliver approved target strings back to the build within hours of the merge. Our AI first-draft runs on every micro-job and is post-edited by a human translator before the LQA gate — not instead of one.
We decline to ship raw AI output to a localized build, even on low-visibility strings. The LQA sampling gate aligned to ISO 17100 runs on every batch in a certified lane, and we refuse to skip it to hit a shorter SLA — the certification requirement is not compressible, only the cycle around it is.
We also refuse to quote a continuous program on a per-PO basis. The economics only work for the client and for us under retained monthly capacity with an SLA, because the PM mode is exception handling, not dispatch-and-track. A client asking for per-PO pricing on a weekly commit cadence is a client whose project belongs on waterfall or has not yet crossed the four-trigger threshold.
We will not claim universal acceptance of certified translations produced in a continuous lane. We prepare the certificate of accuracy and the sign-off; the receiving authority — a court, a registry, a regulator — decides acceptance, and we tell the client to check with theirs.
The questions below answer the operational edges that translators and agencies hit when they first move a client off waterfall and onto a continuous lane.
Frequently asked questions
What is continuous localization?
Continuous localization is a workflow model in which translation and localization work is released and updated in small, frequent increments rather than in large, periodic batches. Borrowed from continuous delivery in software development, it replaces the traditional "big-bang" release cycle — where all content is translated at once before launch — with an ongoing pipeline that processes new or changed strings as soon as they appear. For agencies and translators handling certified documents, this means integrating translation production, QA, and certification steps into a repeatable, near-real-time process that scales with client content velocity.
Related
- Crowdin Translation: How Translators and Agencies Run Projects Inside Crowdin
- Define Localization: What the Term Means in Professional Translation
- Define Transcreation: What It Means Inside a Translator's Workflow
- E-commerce Localization: What Translators and Agencies Actually Deliver
- Expo Localization: What Translators and Agencies Need to Know About expo-localization
- Go i18n: How Translators Handle Golang Localization Handoffs