The documentation right now [0] is pretty light on how this works. I watched the IO presentation[1] on App Actions and it was also unclear.
While semantic intents exist locally, it is unclear where Google maps from a user action back to that semantic intent, and many of their examples involve online google services (like Google Search and Google Assistant). They also say the intent must be registered in a Google database.
Look at the slide within the video at 23:30 specifically.
My best guess right now is the contextual suggested actions (ex. call a specific person) that show up in the launcher are done with on-device machine learning as she says, but more generic actions (ex. open the dialer to call somebody, not a specific person) that show up in search results would be done with something like Firebase App Indexing and would require those actions to be submitted to Google.
[0] https://www.twit.tv/shows/all-about-android/episodes/368
My 6 year old lumias still have better power usage than the latest Samsungs. No cloud AI necessary.