Here's a couple of different ways I work and have been seeing other folks work with AT in graphical mode:
1) Select a phrase of about 15 seconds, record it in and then do "make curve" where AT just traces your singers exact pitch. Then look and see (with your eyes and ears) which notes are being scooped too much, where they're going flat, etc. Usually it's just a note here and there. I'll get the "freehand drawing tool" out and just make a slight dot at the beginning and end of a spot I want to fix. (Doing that frees that section to be manipulated without dragging the pitch of the rest of the phrase.) Sometimes the line has a natural break to it (where there's a breath or other pause) and I don't need to separate it with the freehand line tool. I'll go with the grabber tool and grab the end of the phrase that needs the work, and yank and stretch it around until it fits what sounds good. Sometimes I'll leave a bad note in there if it's just simply too unnatural sounding to get it to a good note. I'll also mine around other places to see if I got a better performance of that one stellar note that I can paste in there. Generally those notes are passing tones in a run and get forgotten quickly in the grand scheme of the song.
2) When I'm in a hurry to burn through a bunch of vocals, I'll set the key as best as I can in the auto window, then go over to graphical, select my phrases, hit "make auto" and just manually clean up the little squirrely parts where the singer lands halfway between 2 notes (and Autotune wants you to get your money's worth, so it pitches you up and down to get both notes) and do basic damage control fixing.
3) When I'm in a bigger hurry than that, I'll bounce the track to another using autotune in auto mode, then I'll go back and fix any sections with graphical mode if there were any warbles or burbles, or repitching to not quite the right note.
4) I've seen other folks work in such a way that they grab a large section of audio (up to most of a verse or chorus) Then go in with the line tool and draw stair-step the exact melody the singer "should" have been singing. (if the singer were a midi device) Then go and adjust the retune speed knob. Start with the numbers down way low (where you'll have the robotic sound) and adjust it higher (slower retune speed) until the vocal sounds natural. You want to find a happy medium between robotic and AT getting around to actually tuning the note.
I'm a Protools user, so here's some PT specific notes:
I never use Audiosuite simply because my processor seems to be too slow for working with it in graphical mode. I have a friend with a computer that runs 2x as fast as mine and has no problems working with it in Audiosuite.
Instead, I route my track with AT inserted on it through a bus and bounce to another (adjacent) track.
After fixing the entire track, I'll go and listen for pops at region boundaries and fix them with fades and nudging regions around.
When I'm done tuning a vocal, I make at least 2 playlists on the vocal track. 1 of the original vocal comp and 1 with autotune printed on it with the regions consolidated. That way I can quickly find the original vocal. If the singer still thinks the note sounds funny (usually there's more than just pitch wrong with sour notes) I can also quickly show them the before/after. Or if they think they can hear the autotune artifacts (or if I hear a blurp that I didn't catch earlier) I can grab the original, fix it and slug it back into the final comp quickly.
Hope this is of some use.
Greg Thompson