Sampling is a solved mathematical problem: given a population, a confidence level and an acceptable margin of error, there is a number of items you need to review. Review fewer and the result is not wrong exactly. It is just too imprecise to act on.
The industry convention of "review ten calls per agent per month" is not sampling. It is a workload, chosen because it is what a human analyst can get through, and the margin of error it produces is usually wide enough to swallow the differences you are trying to detect.
Nothing stops you auditing everything, which sidesteps the question. If you would rather sample, the sampling calculator works out the number and shows you what the conventional ten-per-agent actually buys.
Sampling calculator