Machine translation of course content is good enough today to save most of the work, and still not good enough to publish unread. The model does not know your internal vocabulary, the name your company uses for a process, or that a product name must stay in English. Someone who speaks the language and knows the subject should click through every version once.
Keep a short list of terms that must not be translated, such as your company, product and system names, and the terms your legal or safety team requires. Check them first; they are where translations go wrong most often. For regulated content like safety or compliance training, the reviewer should be someone qualified to sign it off, not only a fluent speaker.
Pay attention to fill-in-the-blanks questions. The accepted answers are words, so they are translated too. Make sure each one is the word a learner would actually type, and add alternative spellings, such as colour and color. Translated text also tends to be longer than English, so look at tight spots: buttons, tab labels and flip cards.
If you later add or remove slides in the original, the translated versions no longer match its structure. The export tells you which language needs translating again before it can join the package, so you never ship a version that is out of step.
What is not translated
Images stay the same in every language, and so do recorded narration files. Replace screenshots with text in them, and re-record narration or let the browser read the translated transcript aloud.