i've been wrestling with this for awhile too.
i try to come from a purely pragmatic approach. personally, i don't make music for purists or other musicians, so i'm not concerned with procedure for its own sake.
thus, the question i ask is this: "what level of hands-on control will give me the ability to improve the result the most?"
one pitfall would be having too much control. a scenario of this would be to attempt to manually sequence every beat or 4-bar phrase individually. the audience wouldn't enjoy it as much because it would fall apart every time you tried to switch motifs, and you'd get a sore arm from constantly hammering at the keys.
another pitfall would be having not-enough control. a scenario of this would be only being able to trigger the beginning of a 3-minute track. the audience would have less fun because it would be hard to change the mood to fit them, and you would be doing less; they can sense that.
in my case, the ideal is a combination of three levels of control:
1) a section of "song level" components, i.e.: MySong_Verse1, MySong_Verse2, MySong_Outro.
2) a section of "phrase level" components, i.e.: Amen_4BarLoop, JB_Bassline8Bars.
3) a section of "trigger level" components, i.e.: Scorpio_DrumRoll_2Beats.
I find that by combining these three levels, there are long enough parts to make it sound like songs, but there are still fine enough grains that i can alter the feel by manually "playing" stuff.