Erlang is a convolution of exponential samples. Thus Erlang 1 (K=1) is the same as exponential. Erlang 2 is a very left skewed bell chopped at 0. As K increases the shape moves towards normal but never samples < 0.
In Simul8 it is commonly used for Mean-Time-To-Repair (MTTR) reflecting its common usage for that in manufacturing. Otherwise, it is used as an extension of the Exponential distribution for any sort of durations, such as between arrivals.