AFAIK, those devices have two systems.
The subsystem only listens for the keyword. It doesn't have any capability to process the whole question you ask. Its only job is to listen for the keyword and then send the relevant audio to the main system.
The main system is what's connected to the Internet. Only it can send the audio file to the mother ship to be processed.
With this setup, they improve general security by ensuring only the keyword audio is accessible to the processing center of the device.