Each stage is an M/M/c queue. Jobs arrive as a Poisson stream with rate λ; stage i has ci parallel workers, each with exponential service time of mean 1/μi. The utilisation is ρi = λ / (ci μi). If ρ ≥ 1 the queue grows without bound: that stage is the bottleneck and no improvement elsewhere helps.
Erlang C. With offered load a = λ/μ the probability that an arrival must wait is Pw = [ac/(c!(1−ρ))] / [Σk<c ak/k! + ac/(c!(1−ρ))], the mean wait in the queue is Wq = Pw / (cμ − λ) and the time in the stage is W = Wq + 1/μ. By Burke's theorem the departures of a stable M/M/c queue are again Poisson, so stages in series behave as independent queues (a Jackson network) and the total cycle time is the sum of the three Wi.
Little's law. In any stable system the long-run average number of jobs inside equals throughput times average time inside: L = λ W. It needs no assumption about the distributions, so it is used as an independent check of the simulator's bookkeeping.
Honest simplifications. Real services are rarely exponential (lower variance shortens queues; the Kingman formula scales waiting by (ca²+cs²)/2), arrivals vary over the day, workers take breaks, and jobs are not identical. Costs are illustrative. Early in a run the averages are still converging, so measured numbers wobble; the checks below simulate 200,000 minutes with a fixed seed.