Your step does not take eight minutes. It takes a shape.

Ask anyone how long a step takes and you get one number. Eight minutes. It is a reasonable answer to a reasonable question, and it is the single most expensive simplification in process work, because the one number is an average and the average is exactly the part that does not cause the problem.

The spread causes the problem. And unlike the average, the spread is not something anyone quotes from memory.

Why one number is not enough

Two teams both average eight minutes a request.

Worked example: two teams, identical averages, at the same arrival rate

	Team A, every request	8 min
	Team B, most requests	4 min
	Team B, one in six	28 min
	Both teams, mean handling time	8 min
	Average wait before being started, A against B	3 min vs 21 min

Illustrative figures at the same utilization. The averages are identical and the queues are not close, because waiting time responds to the spread of handling times as well as to the mean. The long tail on Team B is doing all of it.

Both teams would report eight minutes. One of them has a queue and the other does not, and no model built on the average alone can tell you which is which. This is the mechanism behind the last ten percent of utilization: the spread is one of the two inputs, and quoting a mean discards it.

What to collect

You do not need a study. You need thirty to fifty observations of the same step, timed the same way, and the discipline to record every one rather than the ones that felt typical.

Time the whole step, not the work. From the moment the item became available to the moment it left. If somebody picked it up, put it down and came back, that is part of it. The gap is usually where most of the time is.

Record every observation, including the ridiculous one. The instinct is to drop the outlier because something unusual happened. Something unusual happens at a rate, and that rate is the finding. An outlier dropped is a tail deleted, and the tail is the thing that was worth measuring.

Note what was different about the long ones. Often the tail is not one distribution at all. It is two kinds of work with different shapes that nobody had separated, and discovering that is worth more than any curve you could fit to the mixture.

Reading what you get back

Three shapes cover most of what a process throws up, and you can tell them apart by eye from a plot of the raw numbers.

What the plot looks like	What it usually means
A tight cluster, nearly symmetric	A machine-paced or well standardized step. The average is close to safe here
A hump with a long right tail	The normal case. Most items are quick, some are slow, none are faster than a floor
Two humps	Two kinds of work wearing one name. Split them and model each

The third is the most common surprise and the most valuable. Two humps means the step you are modeling is really two steps, and the fix is a classification rather than a curve.

Where this goes next

Once you have the shape rather than the number, the natural next question is what tool will actually take it. Some ask for a mean and quietly assume the rest. Others fit a curve to observed data and let you see the fit before you accept it, which is the feature that matters here, because a distribution chosen by eye from a menu is a guess wearing a technical name.

That is a tooling question rather than a method one, and Process Improvement Reviews covers it across the products in this category, from published sources and against one rubric.

What to do with it

Stop quoting single numbers for anything that varies. Quote the median and the ninetieth percentile. Two numbers, both from data you already have, and the pair says more than any average.

Plot the raw observations before fitting anything. Thirty dots on a line takes two minutes and tells you which of the three shapes you have. Fitting first and looking second is how a two humped step gets modeled as one.

And keep the outliers. They are the tail, the tail is the queue, and the queue is the reason anybody asked.