Machine translation can make a government notice readable in a refugee’s language, but a faithful translation of bureaucratic Korean is still bureaucratic, and newcomers often cannot act on it. We report, with the operating organization as a co-author, what happened when a small refugee-serving organization in South Korea put a large language model in front of the translation step at the service it has operated since 2018. Scoped with the frontline staff who field refugees’ questions, the model rewrites the Korean source into plain language, explains Korea-specific terms where they appear, checks whether figures and deadlines have gone out of date, and pins proper names to a fixed glossary before anything is translated into seven languages. It covers most of the catalog and shortens the sentences readers meet. It also fails in ways a translation-only system never does. The stages allowed to search the web made corrections no translation could make, and they also wrote every serious error we found, including policy updates that were simply wrong; the stages without that access wrote a handful, none of them critical and all of them confined to metadata. Value and failure come from one capability, so the lesson is not to withhold it but to put the human check where it acts. For an organization with nobody on staff who reads the seven target languages, that check must be placed on the Korean source before publication, which makes Korean-language staff review the safeguard the whole design depends on.