Kaoru Ishikawa's position was that the large majority of quality problems in a factory can be solved by the people who run it, using a small set of simple tools and no statistics beyond arithmetic. He put the share at around 95 per cent. As a claim about the tools that is unfalsifiable; as a claim about who should be doing the analysis it is the whole point. None of the seven needs a statistician.
They are usually taught as a list, which is the least useful way to hold them. The tools are sequential, each consuming the output of the one before it, and using them out of order is how a team spends six months on the wrong defect.
What are the seven basic quality tools?
| Tool | The question it answers |
|---|---|
| Check sheet | What is actually happening? |
| Pareto chart | Which few causes dominate? |
| Cause-and-effect diagram | What could be causing it? |
| Histogram | What shape is the data? |
| Scatter diagram | Do these two things move together? |
| Control chart | Is this variation normal for this process? |
| Stratification | Is this one population, or several? |
The seventh is not agreed, and pretending otherwise is unhelpful
Six of these appear on every list; the seventh does not. Stratification, the flowchart and the run chart are swapped in and out depending on who is listing them. ASQ names stratification and notes that some organisations substitute a flowchart or a run chart; Ishikawa's own presentation grouped graphs and control charts together, which frees a slot.
Nothing hangs on it. Stratification is arguably not a chart at all but a way of looking, which is why it is the one that gets displaced. If your training material says flowchart and the auditor's says stratification, they are not in conflict; the set was never canonised.
Check sheet: what is actually happening?
A form, designed before collection starts, with the categories already printed on it and a space to make a mark. That is all it is, and it is the one everyone skips.
The analytical work is entirely in the design. The categories you print are the categories you will be able to see. A log with three boxes produces a three-bar Pareto whatever the process is doing. The same goes for context — machine, cavity, shift, operator, lot, time of day. If it is not recorded at the moment of the observation it cannot be recovered afterwards, and that is the commonest reason a stratification fails six months later.
A location check sheet — a drawing of the part, marked where each defect was found — frequently solves the problem outright: twelve marks clustered at one corner is a fixture, a gate or a handling point.
When it is the wrong tool: when the data already exists — if the machine or the MES records it, extract it. And a check sheet is blind by construction, so always leave an "Other" line and read it. A large "Other" is the form telling you the categories are wrong.
Pareto chart: which few causes dominate?
Categories as bars, largest to smallest, with the running cumulative percentage as a line across the top. Its job is to turn a list of problems into an order of work.
The decision that matters is the weight. Most are drawn by frequency, because frequency is what the defect log gives you free — and frequency is right only when every occurrence costs about the same. Four hundred scratches that polish out are a smaller problem than twenty seal leaks that scrap the assembly. Draw it both ways; where they disagree, the cost chart sets the agenda.
When it is the wrong tool: it has no time axis, so it cannot tell a category that is steadily worsening from one that spiked once in week two. A flat Pareto is also a real finding rather than a failed chart — it says there is no dominant cause, and a project aimed at any one bar buys a sixth of the problem.
Paste categories and counts into the Pareto chart generator, or see how to build one in Excel.
Cause-and-effect diagram: what could be causing it?
The fishbone, or Ishikawa diagram. The effect goes at the head, major categories branch off a spine — Machine, Method, Material, People, Measurement, Environment — and candidate causes hang off those.
It is the only one of the seven that produces no numbers, and that is what it is for: it makes a group state its hypotheses out loud, including the ones that would otherwise stay in one person's head.
When it is the wrong tool: whenever it is used as the analysis rather than the plan for one. It generates candidates; it ranks nothing and proves nothing. Drawn before anyone has counted anything it is a meeting with a whiteboard, and the cause that gets fixed is the one argued for most forcefully. The fishbone diagram guide covers how to run one so that it ends in a list of checks rather than a decision.
Histogram: what shape is the data?
A bar per interval, height as count. It shows centre, spread and — the part worth having — shape, which every summary statistic destroys. Four shapes carry specific meaning:
- Two humps. Two populations mixed: two spindles, two suppliers, two shifts. No average describes it and no capability index from it means anything.
- A cliff edge. The distribution stops dead. Somebody is sorting, and you are looking at the survivors rather than the output.
- Comb or heaping. Values piling on particular readings — usually gauge resolution or rounding.
- A long tail on one side. Often legitimate: anything bounded at zero is skewed by its own physics.
When it is the wrong tool: it discards time order completely. A process that drifted steadily across a month produces a respectable bell and the histogram will not hint at it. Chart it over time first, look at shape second. Bin count is a real choice too, not a default — the histogram generator reports both standard bin rules and a normality test that declines to declare data normal, because no test can.
Scatter diagram: do these two things move together?
One point per observation, suspected cause horizontal, effect vertical. It is how a hypothesis off the fishbone gets tested: if humidity really drives the reject rate, plot one against the other and look.
When it is the wrong tool: three traps, all common. Correlation is not causation, and both variables may be moving with a third. More damaging: if the input barely varied during the study, a strong relationship appears as a featureless blob — twenty readings taken while the oven sat between 181 and 183 °C say nothing about a temperature effect that is real across the working range. Third, a non-linear relationship can give a correlation near zero over a clean, obvious curve. Look at the plot, never the number instead of it.
Control chart: is this variation normal for this process?
Readings in time order, a centre line, and limits computed from the process's own short-term variation rather than chosen by anyone. Inside them the variation is what the process does, and the correct action is to leave it alone. Outside them, or in one of the run patterns, something has happened worth looking for.
This is the tool that stops improvement work being a treadmill: adjusting a machine in response to ordinary variation measurably increases the variation — see common cause and special cause variation, and what a Shewhart chart is for which chart in the family to use.
When it is the wrong tool: it says nothing about whether the process meets the tolerance. Control limits come from the process, specification limits come from the drawing, and a process can be in perfect control and entirely incapable. That second question is capability, and the order is not negotiable — an index computed on an unstable process describes something that does not exist. Paste a column into the control chart generator to see it on your own numbers.
Stratification: is this one population, or several?
Splitting the data by something recorded alongside it — machine, cavity, shift, operator, supplier, lot — and looking at the pieces separately. It is less a chart than a habit, and it is the one that most often turns an intractable problem into an obvious one.
A flat Pareto is frequently steep inside one stratum. An unremarkable histogram separates into two clean distributions when split by spindle. An out-of-control chart with no assignable event is often two processes plotted as one. Side-by-side box plots show several strata on one axis, though a histogram per stratum shows shape that a box plot deliberately hides.
When it is the wrong tool: when the stratifying variable was never recorded. You cannot split by cavity number if nobody wrote it down. This is why the check sheet comes first.
Which tool for which question
| The question you actually have | Tool | What you need first |
|---|---|---|
| What is going wrong, and how often? | Check sheet | Categories agreed in advance |
| Which problem should we work on? | Pareto chart | Counts, and a chosen weight |
| What might be causing it? | Cause-and-effect diagram | A specific, measurable effect |
| Is that candidate cause real? | Scatter diagram | Both variables, over a real range |
| Is it one machine, shift or lot? | Stratification | The stratifying variable recorded |
| Is this normal for this process? | Control chart | Readings in time order |
| Is the output the right shape and width? | Histogram | A stable process, first |
| Did the fix hold? | Control chart | Limits frozen to the pre-fix baseline |
The order matters more than the drawings
- Check sheet — find out what is actually happening, in categories chosen deliberately, with the context recorded alongside.
- Pareto — decide which of those is worth anyone's month.
- Cause-and-effect — generate hypotheses about the one you picked, and only that one.
- Scatter, stratification, histogram — test the hypotheses against data. This is the step that gets skipped, and the only one that produces evidence.
- Control chart — confirm the fix held, against limits from before the change.
A fishbone drawn before anyone counted anything is a meeting, not an investigation. A Pareto drawn from a log with three categories ranks the form, not the process. An improvement declared on a lower average next month is a coin toss someone has chosen to interpret.
What the seven cannot do
They are deliberately elementary and they stop somewhere. They will not size a small effect against a noisy background, which needs a hypothesis test or a designed experiment, nor disentangle several factors interacting, which needs a factorial design. And they will not tell you whether the numbers coming off your gauge are the part or the gauge, which needs a measurement systems analysis — worth checking early, because a gauge contributing most of the observed variation makes every other tool here describe the instrument.
Ishikawa's argument was never that these are the most powerful tools available. It was that they are powerful enough, and that the alternative — sending the problem to a specialist queue — solves less.