Scaling Localization with AI+HT

1 Problem

Translators were working from stand-alone strings with no way to know where a piece of text actually lived in the product. That lack of context drove constant back-and-forth questions, slowed delivery, and made costly language UAT cycles feel unavoidable for every release. Meanwhile the localization footprint kept growing — more languages, more products — without a way to scale quality or spend efficiently.

2 Approach

I built a localization program around context and a blended AI+human translation workflow, rather than treating translation as a black box handed off to an agency.

• Deployed Context Capture so every translation segment linked back to a real-time preview of the UI copy on the UAT environment — translators could see exactly where a string appeared, instead of guessing at meaning from an isolated sentence.
• Layered in an AI-assisted Human Translation (AIHT) workflow: generative AI produced a natural first-pass translation, then professional linguists reviewed it for accuracy — pairing AI speed with human quality control rather than replacing one with the other.
• Published localization checklists for designers, UX writers, and developers covering everything from text-expansion-safe layouts to avoiding string concatenation and embedding editable text (not text baked into graphics).
• Tied translation requirements into the GitHub content pipeline — Jira tickets and Figma links attached to every pull request gave Smartling's translators full visual context automatically.

3 Result

The program scaled from 6 to 18 supported languages while cutting delivery time by 50% and cost by 40%. In 2023 alone, providing UI context in Smartling drove an over 80% decrease in source-language and translation issues raised by translators and reviewers.

On quality: benchmarking AIHT against traditional human translation showed no meaningful quality degradation — a 98.9% MQM score for pure human translation versus 98.8% for AIHT — while AIHT delivered a 26% cost saving per word ($0.13 vs. $0.18). That let the team hold the line on quality while reinvesting the savings into expanding language coverage.