The do what I mean interface of Siri could be limited to actions with insignificant consequences - the human brain incorporates the best speech recognition available and still makes mistakes. Any speech AI performing significant actions would need to outperform the brain.
Or actions which can be undone.
I'd be very concerned if the US army decided to use Siri to control its nuclear missiles. Less so if someone uses Siri to change their thermostat or query their fridge contents.