You can make "lossless" current sensing which is as good as your
transistor matching, though it takes some elaborateness to drive
out all the systematic error sources (for example your sense path
FET needs to be maintained at the same headroom as the switch
FET, in-the-moment, or the difference between a linear-region
switch and a saturation-region sense device makes natural match
moot. Your advantage is zero added switch drop. Your costs are
complexity, the probability that where you sense has a different
current density than elsewhere in the power device, and so on.
But you can do better than 10% tolerance if you are good with
the physical design and analysis of layout effects.