You costed poor quality: how do you analyse it to actually cut it?

You finally have a figure. The cost of poor quality for your site fits on a single line, defensible, presented to the management committee. And then, nothing. The amount is impressive, but nobody knows where to start on Monday morning.

That is the paradox of the total. It reassures management and leaves the shop floor without instructions. A global amount says neither where the waste concentrates, nor which cause to pull first to bring it down.

Costing poor quality was the subject of an earlier piece of work: aggregating scattered costs into a credible total. That figure is now in place. What remains is the part that makes it actionable: slicing it up, spotting the few causes that concentrate the cost, then acting at the root.

Why does a costed figure not reduce itself?

A total is a photograph, not a map. It says how much poor quality costs, without saying where it lives. Two sites with the same amount can have opposite problems: one accumulates micro-scrap everywhere, the other concentrates everything on a rare, expensive defect.

As long as the figure stays global, everyone projects their own hypothesis. The quality manager suspects a supplier, production blames a setting, management asks for a plan. Nobody is wrong, nobody has the proof.

Worse, a global total can fall for the wrong reason. If production volume drops, the poor-quality bill drops with it, without anything having improved. Without slicing, it is impossible to tell real progress from a mere effect of the business cycle.

The total serves to measure the stakes and to decide whether it is worth acting. It does not serve to choose the first action. That shift, from measurement to decision, hinges on a single thing: slicing.

Along which axes should you slice the cost?

To analyse is to answer precise questions with the same figure, looked at from several angles. Each axis lights up a family of possible causes. None is enough on its own, but together they pin down the problem.

Table scrolls horizontally on small screens.

The defect-type axis deserves a place of its own. It is often the one that links the others: the same defect can point to a supplier, a machine or a step. Its recurrence is a subject in itself, covered in the analysis of recurring nonconformities.

Clean slicing assumes data recorded at the useful level of detail. If the scrap reason is not captured, or if the supplier batch does not flow through to the control station, the axis stays blind. The quality of the analysis depends first on the quality of the traceability upstream.

How do you find the priorities within this slicing?

One rule of thumb holds almost everywhere: a few defects concentrate most of the cost. The rest is noise, costly to handle and yielding little to recover. Analysis serves to separate the two.

To rank, you have to cross two quantities that are often confused:

• Frequency. How often the defect occurs.
• Unit cost. What one occurrence costs: scrap, rework, sorting, customer return.

The product of the two gives the real weight. A very frequent but cheap defect can weigh less than a rare, heavy one, the kind that triggers a customer return or ties up a whole batch.

The exercise also has an upper limit. Trying to handle the long tail of small defects costs more in analysis time and meetings than it brings back. Concentrating on the few heaviest pairs is not laziness, it is the best return on the analysis effort.

What does AI change in this analysis?

A human analyses the axes one by one. They look at defects by line, then by supplier, then by team. The common factor that appears only at the crossing of several axes escapes them, not for lack of rigour, but because the table has too many dimensions for the eye.

AI crosses all the axes at once. It does not look for a cause, it spots where costs cluster abnormally, even when the clustering rests on three simultaneous conditions.

You have to stay honest about what the tool does here. It offers a correlation, not a cause. The fact that the cost concentrates on this triplet does not say why. Perhaps the batch reacts to a lower workshop temperature at night on that line. The correlation steers the investigation, the root cause is confirmed on the ground. The tool links and detects, the human decides and verifies.

This honesty has a practical corollary. The tool creates no data: it links and groups what already exists, scattered. If the scrap capture is wrong or incomplete, the crossing will point to false priorities with the same confidence. A suspect pattern always deserves a check against the evidence before launching a project.

From analysis to reduction: how do you close the loop?

Finding the priority reduces nothing. The drop comes from the action, and the action must target the cause, not the symptom. Sorting harder at the end of the line masks the problem and adds a cost. Correcting the setting, the batch or the upstream step removes it.

Before acting, it is worth freezing the starting state: which cost, over which scope, measured how. Without that documented reference point, the later comparison will be contestable, and the gain will become a matter of opinion rather than measurement.

• Handling the most frequent defect rather than the most costly one.Frequency catches the eye, unit cost makes the bill. Cross the two before choosing.
• Confusing correlation with cause.A crossing shows where to look, not why. Confirm the root cause on the ground before acting.
• Analysing a single axis at a time.The most costly causes often hide at the crossing of several axes, invisible in isolation.
• Not re-measuring after the action.Without a new measurement, you do not know whether the cost has fallen, nor can you defend the next project.

Once the action is carried out, a final step closes the loop: re-measure the same cost, over the same scope, with the same method. It is the best proof that the effort paid off, and the argument that unlocks the next project.

That is also where the credibility of a data-driven quality approach is decided, and AI in the service of industrial quality gives its wider frame. Measure, slice, act, re-measure: the cycle is worth more than any isolated tool.

This kind of multi-axis analysis, on data that lives in production more than in documents, benefits from purpose-built tooling. That is the ground of AI systems built around your data, calibrated on your axes and your defects rather than on a generic model. This logic fits within a broader industrial AI approach than quality alone.

What should you take away?

A costed cost of poor quality is a starting point, not a result. Its value is born from slicing: by product, line, team, supplier, equipment, process, period and defect type. Priorities come out of crossing frequency with unit cost, never frequency alone. AI accelerates this crossing and surfaces common factors that are invisible axis by axis, but it shows a correlation, not a cause. Reduction, for its part, is played out at the root, then verified by a new measurement.