Variation Isn't Error: What Deming Teaches About Metrics

A building-materials distributor, 180 people, six branches. The Monday meeting's metric is order-picking turnaround. In March it hit 94% on time, in April 89%, in May 92%. At the June meeting, with 88%, the branch manager whose number dropped was called in to explain. He explained: someone was out, it rained, a big order came in. They agreed on an action plan. The next month it hit 93% and nobody said a word.

This happened twelve times that year. Twelve explanations, twelve action plans, and the metric kept wandering between 88% and 94%. The operation didn't improve by a single point. What improved was the managers' skill at producing explanations.

01The idea: two kinds of variation #

W. Edwards Deming published Out of the Crisis in 1982, at age 81, after three decades helping Japanese industry rebuild its processes. The book lays out the 14 points for management, but the idea underpinning all of it goes back further, to Walter Shewhart in the 1930s: every process varies, and there are two kinds of variation.

Common cause is the variation built into the system itself. It's there every day, produced by the combination of method, equipment, information and rule. No single point within that range has its own explanation — asking 'why did it drop to 88%' is like asking why a die rolled a 3.

Special cause is what comes from outside the system: an identifiable event with a traceable origin. That one does deserve targeted investigation.

The expensive mistake is treating common cause as if it were special. Deming called this tampering: reacting to noise as if it were signal increases variation instead of reducing it. And in the chapter on leadership, he's explicit that blaming a worker for variation that belongs to the system is both unfair and ineffective — most causes lie in the system, and the system is management's responsibility.

02The red bead experiment #

Deming used to demonstrate this in the classroom with a simple exercise. A box with white and red beads; volunteers use a paddle with fifty holes to draw fifty beads at a time. The red ones are defects. The procedure is rigid: no shaking, no picking, no changing anything.

Some draw 9 reds, others 16. The 'supervisor' praises the one who drew 9, warns the one who drew 16, fires the worst performer, promotes the best. In the next round the roles flip — because the proportion of red beads in the box never changed. Individual performance there is noise. The only way to improve the outcome is to change the box.

It's uncomfortable to admit how many results meetings in our own lives were the red bead experiment with coffee.

03What changed and what hasn't since 1982 #

The industry changed: Deming was writing for manufacturing, and today most of the work in a mid-sized company is administrative — proposals, approvals, records, service, closing. Data availability changed: in 1982 you had to build a control chart by hand; today the record already exists in some system.

The math hasn't changed. A purchase-approval process varies exactly the way an assembly line varies. And the managerial reflex hasn't changed either: faced with a worse number, hunting for someone to blame is still faster than investigating a system.

One thing actually got worse. Because we can now measure everything weekly, we've created twelve times more opportunities to react to noise than when the report was monthly and arrived on paper.

04The cost of reacting to noise #

Want to see how this looks inside a real operation? Explore the platform.

Let's scale this to a 180-person mid-size Brazilian company with 12 managers (figures in Brazilian reais). Explicit assumptions, all conservative — and estimates, not measurements:

The math: 72 annual hours at a fully loaded rate of R$ 90 an hour comes to roughly R$ 6,5 mil per manager recurrently involved; counting four people per cycle, that lands around R$ 26 mil. The bigger cost, though, isn't on the spreadsheet — it's the manager learning that the right response to a bad number is a good story.

05Where this breaks down in practice #

1. The metric has no range. There's a target and there's actuals, but nobody knows what the process's normal variation looks like — so every deviation looks like a deviation.
2. The data gets re-typed. If the number comes from a spreadsheet filled in by hand on Friday, part of the variation is the process and part is the data entry, and you can't tell them apart.
3. The real special cause gets lost in the noise. When everything turns into an action plan, the event that genuinely deserved investigation gets the same attention as the month that swung by chance.

06What data-driven management answers #

Separating common cause from special cause demands something mundane and hard: the metric needs to come from the work itself, timestamped at every step. When an order records when it came in, when it was picked and when it went out, the normal range shows up on its own — and so does the exception.

It's the same argument behind bringing work, process and analysis into one place — what the market now calls Collaborative Work Management. In practice, that's what the processes and operations module does: every recorded step becomes data, and the metric stops depending on whoever filled in the spreadsheet. It's also worth seeing how this changes by area.

One risk worth naming: fine-grained process data can turn into a surveillance tool. If variation readings are used to pressure people, you've just rebuilt the red bead box with a nicer dashboard — and at that point the real discussion becomes a question of what theory you hold about people.

The number that swung wasn't asking for someone to blame. It was just describing the system that management built.