Erlang B can be used for either case. Correlation is really non existant unless you can prove that the sample can be repeated accurately, many times.
Remember that this is a prediction, based upon numerous assumptions about the sample. It is improbable that all assumptions are in fact true for the (any) sample and therefor there is a ‘margin of error’ in the calculations.
None the less, it is my experience that Erlang B provides a very close estimate of the lines required