"There is no static analysis — you cannot catch missing translations at build time"

Correct — there is no "missing key" concept, because the source string is the key. The trade-off: a typo creates a new translation (and the AI translates it) rather than a build error. The dashboard shows which strings have been seen but not yet reviewed, so you can audit production usage. For teams that want build-time safety, a key-based library (i18next, Lingui, next-intl) is the right tool.