Pareto chart

The principle comes from the economist Vilfredo Pareto; it was Joseph Juran who introduced it into quality, under the formula of the "vital few" as opposed to the "useful many". The chart combines decreasing bars and a cumulative curve.

The 80/20 rule is a frequent observation, not a law. Some processes give 70/30, others 95/5. Trying to find exactly 80/20 in your data, or worse, forcing the categorisation until you get there, amounts to decorating a prejudice. What counts is the shape of the curve: if it is almost straight, there is no vital few, and the tool has just said something useful.

A Pareto by number of occurrences and a Pareto by cost rarely rank in the same order. A rare defect that causes a customer return often costs more than the most frequent defect, repaired in thirty seconds at the station.

There is no right unit in absolute terms: there is the unit that corresponds to what one is trying to reduce. Choosing it explicitly, and saying so, avoids long discussions about a ranking that nobody really disputes.

The second useful reflex is to redo the Pareto after the action. It is the only way to check that the first bar has come down — and not that the ranking has reorganised itself around a cause that was already there.

The Pareto serves first to end a discussion. Faced with a list of ten possible causes, a team debates all ten, and everyone defends the one they know best — that is, the one they encounter, not the one that weighs. The chart replaces that debate with a ranking nobody disputes, because it is built from counts everyone accepted beforehand.

It then serves to protect the project from itself. A team that treats six causes simultaneously will never know which one produced the gain, and will therefore be unable to reproduce anything elsewhere. The Pareto justifies treating only one or two, which is an unpopular choice to defend without a chart.

Its condition of use, on the other hand, is strict: count data over a representative period is needed, and categories that two different people assign the same way. Without that second point, labels are being ranked rather than causes, and the chart lends undeserved authority to an arbitrary filing.

One month of records on a paint line, 1,000 defects logged. Run 384, dust speck 271, insufficient thickness 118, colour mismatch 96, scratch 61, blistering 39, miscellaneous 31.

The cumulative curve reads as follows: 38.4 % for runs alone, 65.5 % with dust, 77.3 % adding insufficient thickness, then 86.9 %, 93.0 %, 96.9 % and 100 %. Two causes therefore carry two thirds of the defects. Note that 80 % is not reached with 20 % of the categories: three out of seven are needed, i.e. 43 %. The chart does not say 80/20, it says what it says — and it is that figure that is used.

The first reflex is to tackle the runs. The second, more profitable, is to stratify before acting: split those 384 runs by the most obvious variable, here the shift. They break down as 261, 71 and 52 across three shifts that each produce about a third of the volume.

That figure changes the action plan entirely. One shift concentrates 68 % of the runs for a third of production, i.e. twice what would be expected of it. The problem is no longer in the process, otherwise the three shifts would be affected alike: it is in a setting, a gesture or an instruction that differs at one station. One no longer seeks to modify an installation, one goes and looks at how a shift works — that is an action with no investment, and it was invisible on the overall chart.

One last figure deserves a glance: the "miscellaneous" category, at 3.1 %. It is negligible here, and that is what allows it to be ignored. Had it weighed 20 %, the first action would not have been to treat the runs but to break down that catch-all — a "miscellaneous" bar that rivals the leading ones signals that the category grid does not describe the problem.