I generally agree, but there are some cases where I've found natural language actually is the right interface. And complex interfaces with lots of touch points but a simple underlying principle seems to be that sweet spot.
For example, I exclusively use Siri to set reminders for times and places, because to do so using a typical touch interface is much more painful.
Whether or not this CAP idea will meet the bar is an open question for me.
I find it telling that your comparison is to touch interfaces. The (unintentional?) inference being that only touch interfaces are painful enough that current generation NLP can compete with them.
I use keyboard and mouse for setting reminders. The app I really miss now I don't have a smartphone is google maps.
If you don't have a smartphone it will be difficult for you to do this, but try creating a time- or location-based reminder using a keyboard interface. And I'm generously assuming that you're going to run this test at a computer, and not directly after bumping into someone in the hallway who reminds you about something you need to do.
This is a bit off-topic from the CAP thing, where I think it's reasonable assume that you are at a computer.
For example, I exclusively use Siri to set reminders for times and places, because to do so using a typical touch interface is much more painful.
Whether or not this CAP idea will meet the bar is an open question for me.