This is a good observation. I discussed this above briefly, but did not give any technical reply because the model would have to get a lot more complicated. One way you could model this is at every point in time, the server can choose to work slowly or quickly, but working quickly has a higher cost. The server then chooses to work slowly or quickly at each time point based on minimizing the long run average cost (to the server and to the customers for waiting). The structure of the solution would be such that when the line exceeds N customers, the server would work quickly, but when the line is at most N, it would work slowly. Figuring out N would really require knowing a lot more about the real parameters (arrival rate, various possible service rates, costs, etc.) so it doesn't lend itself to an easy example.