Skip to content

PATTERN Cited by 1 source

AI-draft, human-approve translation

Intent

Use AI to prepare every translation draft while keeping a qualified translator as the mandatory approval authority before release. The pattern increases throughput without equating model output or edit-distance metrics with customer-facing correctness.

Shape

localized source + context → AI draft → professional review → approve / edit → release
                                      │
                                      └── approved edits → improvement data

When it fits

  • Quality, tone, and cultural interpretation matter enough that a wrong string must not ship automatically.
  • Translation volume makes human first-drafting the bottleneck.
  • The organization can supply message context, terminology, and style policy to the model.

Why it differs from automated evaluator loops

The release gate is a human expert, not an LLM evaluator. This is distinct from patterns/drafter-evaluator-refinement-loop, which relies on an automated critic and retry budget. The two patterns can coexist, but Atlassian's public design uses professional translator approval for every output because semantic quality cannot be reduced to edit distance or an automated verdict. (Source: sources/2026-08-06-atlassian-scaling-localization-at-atlassian-keeping-translation-at-the-pace-of-ai-era-development)

Operating considerations

  • Preserve reviewer ownership of final text and never hide AI provenance from the workflow.
  • Track Translation Edit Rate by language, but do not use it as auto-approval policy.
  • Make approval edits actionable feedback while protecting customer and sensitive text.
  • Monitor reviewer queue latency because it becomes the release bottleneck after drafting is accelerated.

Seen in

Last updated · 620 distilled / 1,953 read