Localization is an operations problem, not a translation problem
Producing one localized video is easy. Keeping eleven of them current through a year of product changes is where every team gives up.
Ask any team why their video content is English-only and you will hear a cost answer. It is usually not the real answer.
The real answer becomes visible about four months in. The team localized eight videos, felt good about it, and then shipped a product change that invalidated six of them. Updating six videos in eleven languages is sixty-six jobs. Nobody does sixty-six jobs. So the localized library quietly becomes a liability — worse than not having it, because customers in those markets are now being shown something wrong in their own language.
The three states of a localized asset
Any localized asset is in one of three states, and most tooling can only distinguish two of them.
- Current — matches the source, which matches the product.
- Stale — the source changed and this language has not been regenerated.
- Diverged — someone made a manual correction in this language that would be lost by regenerating.
The third state is where localization operations actually break. A native reviewer fixes an awkward phrase in German. Three weeks later the source changes, everything regenerates, and their fix is gone. It happens twice and the reviewer stops reviewing.
Pinning
The fix is to make divergence a first-class state rather than an accident. A language can be pinned at the segment level: this paragraph has been humanly corrected, regenerate everything around it but leave this alone.
That turns a regeneration from a destructive operation into a merge, which is what makes it safe to run automatically.
Staleness as a dashboard, not a discovery
The second requirement is that staleness is visible without anyone going looking for it. When the source changes, every derived language should immediately be marked, with a diff of what changed and an estimate of which segments are affected.
Teams that have this treat localization as a weekly review — here are the eleven things that went stale, approve the regeneration. Teams that do not have it treat localization as a quarterly project, which means for most of the quarter their localized content is wrong.
The billing dimension
This is also why we do not charge video minutes for localized versions. If regenerating eleven languages costs eleven times a video, teams will not regenerate, and the library goes stale. Pricing has to agree with the workflow you want people to have.