Norman would refer to these constraints as "signifiers", indicators of possible affordances. It's interesting how weak voice assistants are at signifying what you can actually do with them.

The dev team can add helpful responses that signify to users the available set of voice commands for tasks it can complete based on keywords it can recognize from a user utterance or simply letting them know they didn't understand their response and they can get a list of actions spoken to them by asking for help. (I've worked on published Alexa skills for several large tech companies.)

I think a cool immersive middle ground will be smart surfaces embedded in wall materials that can display things and will simply list out all actions available or anthropomorphize the smart assistant as like a virtual servant that follows you around serving up facts and doing monotonous IoT actions for you.

Now the privacy and surveillance implications of something like this is another story...