Understanding Variation: The One Idea That Changes How You See Work

A team closes 214 support tickets in March and 187 in April. Somewhere, a meeting is already on the calendar to discuss the decline. Explanations will be requested, causes proposed, a recovery plan sketched on a whiteboard. And there is a fair chance the entire exercise is about nothing — because the difference between 214 and 187 may be no more than the ordinary wobble of a stable process, the same wobble that will produce 209 in May and earn someone credit for a turnaround that never happened.

Learning to tell wobble from change is, by a comfortable margin, the most consequential idea in process improvement. It was formalized a century ago, it can be taught in an afternoon, and most organizations still run as if it did not exist.

The idea has a birthplace. In the 1920s Walter Shewhart, a physicist at Bell Telephone Laboratories, observed that manufacturing output varied even when nothing about the process had changed — and that engineers kept adjusting machines in response, as if every movement of the needle had its own cause. His invention, the control chart, gave industry a way to separate routine fluctuation from genuine change. W. Edwards Deming spent the next half-century arguing that this distinction was not a statistical nicety but the foundation of management itself.

Every number wobbles

Run any process twice and it will give you two different answers. File the same expense report on two different days and the elapsed time differs. Bake two loaves from one recipe and they brown differently. Nothing meaningful changed; the outputs changed anyway. A process is not a formula — it is a crowd of small influences: the traffic on the network, the mood of the approver, the humidity in the room, the phrasing of one customer’s request. Each influence nudges the result a little, and the sum of the nudges is what statisticians call variation.

This means a single number is never a verdict. It is one draw from a distribution — one photograph of a moving object. The question that matters is never “why is this number what it is?” but “is this number behaving differently from the numbers before it?” Those are different questions, and they lead to very different meetings.

Two kinds of variation

Shewhart’s insight was that variation comes in exactly two kinds, and everything depends on telling them apart. Common cause variation is the routine fluctuation built into the process as it currently exists — the combined effect of hundreds of small influences, none of them dominant, all of them always present. It is predictable in range even though no single result can be predicted, the way a die is predictable without any one roll being known. A process showing only common cause variation is called stable: it will keep delivering results inside the same band until the process itself is changed.

Special cause variation is different. It is a signal that something specific happened — a new hire took over the queue, a supplier changed material, a form was redesigned, a machine began to drift. Special causes are not part of the process’s ordinary character. They are events with names, and they can be found, verified, and addressed one by one.

The two mistakes

Confuse the two kinds and you will make one of two mistakes, each with its own cost. The first is treating common cause as special — reacting to noise as if it were news. Deming demonstrated the damage with his famous funnel experiment: aim a funnel at a target, drop a marble through it, and then adjust the funnel after every drop to compensate for the miss. The adjustments feel diligent. They make the scatter worse. Reacting to individual results from a stable process does not correct the process; it stacks a new source of variation on top of the old ones.

The second mistake is the mirror image: treating special cause as common — shrugging at a genuine signal because “the numbers always bounce around.” A real shift gets absorbed into the folklore of a noisy metric, and a problem that had a findable cause quietly becomes the new normal. Both mistakes are everywhere once you know their shapes:

• Adjusting the process after every result that is not dead centre — tampering that adds variation instead of removing it
• Demanding a root-cause explanation for every monthly dip in a metric that has wobbled the same way for years
• Redesigning the workflow each quarter on the strength of the latest number
• Praising or blaming individuals for results the system produced, then attributing the next wobble to their character
• Dismissing a genuine shift as “just a bad month” because the metric has always been noisy

Plot the dots

The antidote is almost embarrassingly simple: plot the data in time order. Not a table, not a pair of before-and-after averages, not a quarterly summary — a run chart, one dot per result, in the order the results occurred. Averages destroy the very information you need, because two identical averages can hide completely different stories: one process steady, the other swinging wildly or trending toward a cliff. The time-ordered picture keeps the story.

A run chart alone reveals most of what matters: sustained shifts, gradual trends, repeating cycles. A control chart goes one step further, using the process’s own history to calculate the limits of its ordinary variation. Points inside the limits, with no unusual patterns, mean common cause — leave the process alone, or improve it as a system. A point outside the limits, or a long run on one side of the centre line, means special cause — go find what changed while the trail is fresh. The chart is not a report card. It is a decision rule about which kind of problem you have.

What changes when you see it

Once the distinction sinks in, familiar rituals start to look strange. The monthly review that demands a narrative for every uptick. The dashboard that colours a cell red for a result sitting well inside ordinary variation. The exhortation to “do better,” aimed at people working inside a system whose results are the system’s to give. Deming’s harder lesson was that most improvement is management’s work, because only management can change the system that produces the common cause variation — the staffing, the tools, the policies, the design of the work itself.

None of this lowers standards. It aims effort where effort can act: stable-but-inadequate processes get redesigned rather than exhorted, genuine signals get investigated promptly, and people stop being punished for arithmetic.

Before you demand an explanation for a number, ask whether the number needs one.

Learning to see it on purpose

Everything in this article can be practised tomorrow with a pencil and twenty historical data points. Making it a professional habit — the chart rules, the vocabulary, the way the idea threads through DMAIC — is what structured training is for. Our free White Belt program covers variation and the core language of process improvement in about six hours, ending with a 30-question exam and a verifiable certificate. Green Belt, at $299, goes on to control charts and process capability across 35 hours of material built around a simulated project. Start either one, and the next dip in your dashboard will read very differently.