Additionally, even if you were interested in the actual, correct value, it's trivial and very common to apply calibration and compensation steps, digitally. These steps remove systematic gain and offset errors, sometimes non-linearity and temperature-related drift as well, and only some long-term drift elements remain. This can easily reduce the amount of error by an order of magnitude, if not more.
So, when doing a mental comparison between accuracy and resolution, you need to think about calibrated accuracy (if you are going to calibrate, that is). If you know your raw sensing accuracy is, say, 2% and then decide that logically 8 bits is enough, you don't have enough headroom to improve the data by calibration and run-time compensation.
Not all error is noise. It's often correlated to the signal value (or other measurable factors, such as chip temperature), and whenever it's correlated, it can be corrected, but you need the "wrong" digits to be able to do that.