I find that using a certain amount of each technique (compression on tracking, compression at mix, and level automation) gives the most natural-yet-engaging sounding results.
To try and replicate the effect of compression by drawing fader moves is usually pointless - way too time-consuming, and still won't give me the "sound" of a particular piece of gear, or even plug-in. But to maximise intelligibility of a lyric and bring out all sorts of nice vocal details just by compressing (or limiting) the shit out of the channel is a wild goose chase, since in my experience the vocal will have had the life squeezed from it long before that point.
Two principal differences between compression and level automation: 1) they have two different guiding "brains" - the very consistent, precise but unintelligent GR circuitry of the former, and the broader-brush but intelligently interpreting brain of the mixer with a lyric sheet. 2) with the possible exception of the Waves vocal rider, which I haven't tried, compression does not take into account the volume or density of the rest of the music.
So compression will automatically give me some overall stability of volume and apparent tone, but only automation can emphasise a particularly important lyric, or bring up a "t" that's buried under a hi-hat hit, or duck intrusive breaths that the compressor is unwittingly amplifying, or introduce a crescendo into the chorus, or ... you get the idea. What is essential is for the automation to be POST-COMPRESSION, otherwise the compressor will just even out the differences you've spent two hours writing in.
To sum up, I can get a decent balance with compression, but it never "sounds like a record" until I've done some level automation, even sometimes down to sub-syllable level.