It's fantastic for expressing complex requests that would otherwise be a lot of effort/time consuming on a less-information-dense interface.
But voice has a higher "floor" for ease of use. It's easier to tap a button to confirm than it is to speak "Yes, ok" out loud. It's also more socially appropriate in more circumstances.
The other problem is that voice is very low density as an output modality compared to the status quo, which are high-resolution screens. The amount of time it would take to express even the most basic of information (imagine speaking the HN home page out loud vs. reading it!) is pretty extreme.
Where this forms a bad combination are tasks where it's not realistic for the user to utter the full complex request at once, where the user must consult intermediate outputs to determine the next action. In that case you're in a really vicious scenario: the density of voice input is not really necessary, while the low-density of voice output slows the task down dramatically.
For example, think of a use case where you're booking a flight: "I want to see flights from San Francisco to New York".
It's not really possible for the user to fully define a booking in their initial voice input - the user would reasonably want to review choices, pick seats, etc, necessitating that the task be multi-step. Now imagine if voice was the exclusive modality - the UX would be positively painful.
> "that's the future of most/all consumer interfaces with AI, no?"
And this is why I disagree with this statement. I think the idea that voice is the dominant user interface is far from obvious, especially when it comes to AI systems.
LLM hype men tend to hand-wave around the complexities of this: "the AI will automatically pick the best possible flight for you! Why would you even want to review your choices?!" - which conveniently dismisses every area of weakness for this UX with "the AI will do it for you, trust it"... and time will tell but I suspect that will not work out that way.