ATLASSIAN 2026-08-06

Scaling localization at Atlassian: keeping translation at the pace of AI-era development¶

Summary¶

Atlassian scaled product localization after AI-assisted development increased words sent to translation by 272% year over year in FY25 H2. Its production response is a coupled system rather than a standalone translation model: an internal or vendor-hub AI prepares a context-rich draft, a professional translator approves every string in Smartling before release, and editor plus code-review validation prevents malformed or under-specified source messages from entering the translation flow. Reviewer edits feed future model improvement, while AI-assisted bulk remediation eliminated more than 20,000 existing internationalization defects. (Source: sources/2026-08-06-atlassian-scaling-localization-at-atlassian-keeping-translation-at-the-pace-of-ai-era-development)

Key takeaways¶

Translation throughput is constrained by source quality as well as model quality. Atlassian treats source strings that are ambiguous, English-only, or assembled from fragments as upstream defects. Validation before translation lets the model and reviewer spend effort on semantic and locale judgment instead of reconstructing intent. (Source: sources/2026-08-06-atlassian-scaling-localization-at-atlassian-keeping-translation-at-the-pace-of-ai-era-development)
The shipping gate is human approval, not an LLM judge. AI drafts every string, but professional translators approve each result in the translation management system before it ships. This differs from the Lyft drafter-evaluator loop, where an LLM evaluator gates a bounded automated refinement cycle. (Source: sources/2026-08-06-atlassian-scaling-localization-at-atlassian-keeping-translation-at-the-pace-of-ai-era-development)
Context is assembled per string. The internal translator receives developer notes plus reference context: similar prior translations, glossary terms, and locale-specific style rules. The post says this context-building approach beat external-vendor drafts in a blind multi-language trial. (Source: sources/2026-08-06-atlassian-scaling-localization-at-atlassian-keeping-translation-at-the-pace-of-ai-era-development)
Translation Edit Rate is a useful operational signal, not a substitute for review. Early production TER ranges from roughly 10% to just under 50% by language. Low edit distance indicates a reviewer is often polishing rather than rewriting, but Atlassian explicitly keeps human approval because edit distance does not measure semantic or cultural quality. (Source: sources/2026-08-06-atlassian-scaling-localization-at-atlassian-keeping-translation-at-the-pace-of-ai-era-development)
Reviewer corrections create a closed quality loop. Translator edits feed the translation model, with Atlassian estimating the resulting improvement can produce up to 50% translation-cost savings over time. The claim is an estimate, not a disclosed realized saving. (Source: sources/2026-08-06-atlassian-scaling-localization-at-atlassian-keeping-translation-at-the-pace-of-ai-era-development)
Plural-aware messages replace English-specific string assembly. The validation system steers code such as {count} {count === 1 ? 'item' : 'items'} selected toward one message with a count argument. This preserves target-language plural rules and lets translators localize the entire semantic unit. (Source: sources/2026-08-06-atlassian-scaling-localization-at-atlassian-keeping-translation-at-the-pace-of-ai-era-development)
Developer descriptions are model inputs. A note that names the UI location, the object counted, and a variable's meaning lets the AI and reviewer choose the intended term and grammar. In the example, a glossary maps “work item” to German Vorgang, and the complete context enables correct ICU plural rendering. (Source: sources/2026-08-06-atlassian-scaling-localization-at-atlassian-keeping-translation-at-the-pace-of-ai-era-development)
Shift-left validation and automated remediation address the installed base. Editor and code-review checks stop new defects. Once Atlassian could inventory frontend problems, AI remediated repeated defects in bulk and opened pull requests for developer approval, clearing 20,000+ issues. (Source: sources/2026-08-06-atlassian-scaling-localization-at-atlassian-keeping-translation-at-the-pace-of-ai-era-development)

Architecture¶

Developer-authored message + description
        │
        ├── editor / code-review i18n validation ──► fix source before intake
        │
        ▼
Translation intake
        │
        ├── string-specific context: developer notes, UI meaning, variables
        └── reference context: similar translations, glossary, locale style rules
        │
        ▼
AI draft (internal Atlassian translator or external vendor hub)
        │
        ▼
Smartling / professional-translator approval ──► released translation
        │
        └── reviewer edits ──► model and context improvement loop

Operational numbers¶

Metric	Value	Interpretation
Product languages	20+	Localization must support many language-specific grammar and style rules.
Translation input growth	272% YoY, FY25 H2	Demand growth caused by AI-assisted feature development.
Early Translation Edit Rate	~10% to <50%	Varies by language; a quality/cost signal rather than a release criterion.
Potential cost saving	Up to 50%	Atlassian estimate from the reviewer-edit learning loop.
Remediated existing i18n issues	20,000+	AI bulk fixes submitted as developer-reviewed pull requests.

Systems extracted¶

systems/atlassian-localization-pipeline — the coupled pretranslation, human-review, source-validation, and feedback system.
systems/smartling — the translation management system that hosts the final professional-review gate in this deployment.

Concepts extracted¶

concepts/translation-context-assembly — combine message-specific meaning with reusable translation memory, terminology, and locale policy.
concepts/translation-edit-rate — reviewer edit distance as an operational signal, not an authorization to skip quality review.
concepts/localization-readiness — a source message's ability to carry locale-independent semantics before translation begins.

Patterns extracted¶

patterns/ai-draft-human-approve-translation — AI drafts at volume while professional translators remain the synchronous release authority.
patterns/shift-left-i18n-validation — editor and review-time checks prevent invalid source messages before translation intake.
patterns/reviewer-edits-as-translation-learning-signal — use approved human edits to improve later draft quality.

Caveats¶

Atlassian does not identify the internal model, external vendor-hub implementation, prompt format, or training/update mechanism.
The blind-trial methodology, language list, quality scores, and statistical confidence are not published.
TER is reported as a range only. Per-language volume, cost, latency, and reviewer turnaround are not disclosed.
The “up to 50%” saving is an estimate. The post does not state a realized cost reduction or its baseline.
Smartling is named as the approval workflow but its tenancy, API integration, and release synchronization architecture are not described.
The validation rule set, false-positive rate, IDE integration, and code-review enforcement mode are not described.
The 20,000+ count does not identify issue categories, remediation precision, or rejected pull-request rate.