Quality Control: SPC, Cpk, Cost of Quality & the Inspection Paradox

What is quality control?

Quality control (QC) is the part of quality management that verifies — through inspection, measurement and statistical evaluation — whether a product or process actually conforms to its specification. It is a detection activity. It answers the question "is this part good?" and, when done well, "is this process still capable of making good parts?" It does not, by itself, make anything better. It reveals reality. What happens with that reality is the job of the broader quality system.

After 25 years in manufacturing, including three as a Six Sigma Black Belt in automotive headliner production and more than a decade running global MES and traceability programmes at Johnson Controls and Visteon, I have a strong conviction: most "quality problems" in industry are not really quality problems — they are measurement and data problems. A plant that cannot reliably separate conforming from non-conforming parts in real time is not running quality control; it is running quality theatre. A plant that measures but cannot tell you the Cpk of its top five characteristics on yesterday's shift is running inspection, not control. The distinction matters, because the interventions are different.

This article covers quality control as it actually functions in a modern discrete manufacturing plant in 2026 — the statistical foundations, the inspection strategies, the automotive-specific frameworks, the Cost of Quality model that every plant should know and few actually apply, and the honest view of where AI-based inspection is and is not ready. It is the long-form successor to the original Symestic glossary entry.

Quality Control vs. Quality Assurance vs. Quality Management — the confusion that costs money

Three terms, routinely used interchangeably, with genuinely different meanings. Getting the distinction right is not pedantry; it determines where the improvement budget goes.

The practical consequence: a plant with excellent QC (every defect caught at final inspection) but weak QA (process routinely produces defects that need to be sorted out) is burning money. Every caught defect is a prevented escape to the customer — a good thing — but also evidence of a process that should not have produced it in the first place. The honest improvement target is to make QC find fewer defects, not more, because QA eliminated the source. Organisations that celebrate high QC catch rates without questioning the defect rate are optimising the wrong variable.

The four pillars of operational quality control

Plants that do QC well are consistent about all four. Plants that do QC poorly concentrate on pillar 3 — final inspection — because it is the most visible and the easiest to staff. Relying on final inspection to protect the customer is the classical "inspect in" strategy. It can work, but it is the most expensive way to deliver quality and it is the strategy that hides the largest Cost of Quality.

Statistical Process Control (SPC) — done properly, not cosmetically

Statistical Process Control is the backbone of in-process quality control, and it is probably the most widely misapplied tool in manufacturing. The technique is well over 90 years old — Walter Shewhart developed it at Bell Labs in the 1920s — and the principles have not changed. What changes is how rigorously a plant applies them.

In automotive, the expected capability indices are usually stated as Cpk ≥ 1.33 for ongoing production and ≥ 1.67 for critical characteristics. What I have seen repeatedly in real plants — including top-tier suppliers with official PPAP submissions showing Cpk ≥ 1.67 — is that the field data, six months later, tells a very different story. The PPAP was run under controlled conditions; the daily process is not the PPAP process. The only defence against that drift is continuous, automated SPC running on live data, not charts filled in by hand once per shift.

Inspection strategies — 100 % vs. sampling vs. in-line

The choice of inspection strategy is an economic decision, not a quality decision. Every strategy delivers the same outcome if executed properly; they differ in cost, speed and risk profile.

An uncomfortable number from industrial psychology research: a human inspector performing 100 % visual inspection on a repetitive task typically catches 60–80 % of defects, not 100 %. This is why "we inspect every part" is not a quality strategy — it is a belief about a quality strategy. Machine vision, properly trained and maintained, reliably hits 95 %+, but it requires engineering discipline to keep it there. Lighting changes, seasonal temperature effects, product-mix changes and camera ageing all degrade performance silently.

Cost of Quality — the PAF model every plant should know

The Cost of Quality (CoQ) model, in its prevention-appraisal-failure (PAF) form, is one of the most underused tools in manufacturing finance. It separates the total cost of quality into four buckets and reveals that most plants are spending their quality budget in exactly the wrong place.

The 1:10:100 rule is the operational insight from this model: a defect costs roughly €1 to prevent in design, €10 to detect in-process, and €100 to correct after it has left the plant. The numbers are not literal, but the order of magnitude is well-supported by decades of data from ASQ, Juran, and Crosby. The spending pattern most plants actually exhibit is the inverse: they under-invest in prevention and over-spend on appraisal and internal failure. Shifting €1 from failure to prevention saves €5–€20 in total CoQ over 12–24 months. This is the financial foundation of any serious quality improvement programme.

Automotive-specific frameworks — PPAP, APQP, IATF 16949, 8D

Automotive quality has developed its own vocabulary and its own discipline, and the rigour is higher than almost any other discrete manufacturing sector. If you produce for a Tier-1 supplier or an OEM, you work inside this framework whether you want to or not. The following is the minimum set of terms every production and quality manager in the automotive supply chain needs to know cold.

A number that every quality engineer should have internalised: 10 PPM is 0.001 % defective. That level of quality is not achievable by inspection. At 10 PPM, you cannot inspect your way to quality — you would need to inspect 100,000 parts to reliably find one. Control at 10 PPM requires a process that is inherently capable (Cpk ≥ 1.67 typically) and a measurement system that can resolve the variation. Which brings us to the honest view of why automotive quality is hard: the statistics simply do not allow "inspect harder" as a solution.

Measurement System Analysis (MSA) — the foundation nobody wants to talk about

Before you trust any SPC chart, any Cpk number, any control decision, you need to trust the measurement system that produced the data. Measurement System Analysis — specifically Gauge R&R (Repeatability & Reproducibility) — answers the question: how much of the variation I am seeing is the actual process, and how much is my gauge?

The unpleasant consequence: a plant that has not run MSA on its key gauges does not know whether its Cpk of 1.2 is a real process performance or a measurement artefact. Every serious quality engineer I know can tell a story about chasing a "process problem" for weeks and discovering the root cause was the calibration of a single gauge. MSA is not glamorous. It is also the most leveraged 40 hours anyone in quality will ever spend, and in automotive it is non-negotiable.

Where AI-based inspection is — and is not — ready in 2026

AI-based visual inspection has matured significantly since 2022. In 2026 it is genuinely useful for a well-defined set of problems and genuinely overhyped for a larger set. The honest breakdown:

The ongoing cost that most AI-vision pitches understate is labelling and retraining. A vision system trained on 2024 production will start drifting as tooling wears, materials shift, and defect patterns evolve. Budget for continuous labelling work — typically 0.2–0.5 FTE per deployed system — or the false-positive rate will quietly creep up and operator overrides will turn the system into expensive wallpaper.

The 8D problem-solving discipline — when QC has failed

When a defect escapes QC and reaches the customer, the automotive industry's standard response is the 8D report. It is a structured problem-solving method that forces discipline on a situation where the natural response is panic and blame. It is also the most common context in which the gap between a plant's theoretical QC system and its actual QC system becomes painfully visible.

The step where most 8Ds fail is D4. Under OEM time pressure, teams converge on the first plausible root cause that fits, apply a countermeasure, and file the report. Six months later the defect recurs, because the actual root cause was one level deeper. A D4 that is not independently verified with data — preferably from the MES — is usually a D4 that will cost the plant the same 8D again.

The honest view: why "quality problems" are usually measurement problems

The single most useful observation I carry from 25 years of quality work is this: the majority of problems labelled as "quality" are actually measurement, data or response-time problems in disguise. In my book "OEE: One Number, Many Lies," I wrote it up as a narrative of systematic number-polishing on the shopfloor. The quality version of the same story plays out on every OEM supplier scorecard.

This is why modern quality control in 2026 is inseparable from the data infrastructure that feeds it. SPC without real-time data is shelfware. Cpk without MSA is a decimal number with no meaning. Control plans that are not linked to actual machine signals are wishful thinking. The plants I have seen transform their quality performance in the last decade did not buy better gauges or hire more inspectors — they connected the data.

Where an MES changes quality control

The Meleghy case study is a concrete example of what bidirectional MES–QM integration delivers: the SYMESTIC MES triggers sample inspections in CASQ-it based on cycle counts, captures the result, and writes back to both the quality system and SAP. No paper, no manual transcription, no gap between "the plan says inspect every 50 parts" and what actually happens on the floor. The Neoperl case (PLC-based alarm correlation with quality defects) is the same pattern from the defect side: 15 % scrap reduction came directly from being able to connect defect timestamps to machine alarm events, something that was technically impossible without the integrated data backbone.

FAQ

What is the difference between quality control and quality assurance?
Quality control is detection — it verifies that parts and processes conform to specification. Quality assurance is prevention — it builds processes that are inherently capable of producing conforming parts. A plant with strong QC and weak QA finds lots of defects; a plant with strong QA finds few defects to find. The distinction matters because the investments are different. More inspectors will improve QC; they will not improve QA. The reverse is also true: a great control plan will not help if the measurement system that feeds it has a 40 % Gauge R&R. In practice, effective quality systems require both — but when forced to choose where to invest next euro, prevention beats detection on a 5-to-20x return horizon, confirmed by decades of Cost of Quality data.

Is 100 % inspection always better than sampling?
Statistically no, economically almost never. Human 100 % visual inspection catches 60–80 % of defects, not 100 % — repetition fatigue is well-documented. Automated 100 % inspection can hit 95 %+ but carries significant capital and maintenance cost. AQL sampling, when the process is stable and capable (Cpk ≥ 1.33), delivers defensible quality at a fraction of the cost. The right answer depends on the failure-mode severity. Safety-critical characteristics often warrant 100 % automated inspection with poka-yoke backup. Cosmetic characteristics usually do not. The wrong answer is to apply 100 % manual inspection as a universal insurance policy — it creates false security while consuming resources that would deliver more quality if invested in prevention.

What Cpk should I target?
For ongoing production in most discrete manufacturing: Cpk ≥ 1.33 (short-term) with Ppk ≥ 1.33 (long-term). For critical safety characteristics in automotive: Cpk ≥ 1.67 is the typical OEM requirement. What matters more than the specific target is the relationship between Cpk and Ppk: if your short-term capability is 1.67 but your long-term performance is 1.20, you have process stability problems that inspection will not fix. The warning sign is a Cpk number that has not moved in 12 months — it usually means the study is not being re-run against current data. Capability is not a one-time PPAP exercise; it is a rolling measurement. The best plants track Cpk/Ppk per characteristic per week and treat drift as a trigger for investigation.

Is AI-based visual inspection mature enough to replace human inspectors?
For well-defined, high-volume defect categories with stable visual characteristics: yes, often exceeding human performance. For complex, low-volume, or evolving defects: not yet. The realistic 2026 deployment pattern is hybrid — AI catches the vast majority of well-known defect types, and a reduced human inspection team handles exceptions, novel defects and final judgement calls. The hidden cost of AI vision is ongoing labelling work: training data ages as tooling wears and products evolve. A system installed in 2024 without a labelling-refresh budget will have drifted meaningfully by 2026. Plants that budget for the model-maintenance work get the benefit; plants that treat AI vision as a one-time capex line quietly watch the false-positive rate climb until operators start overriding the system.

How does quality control connect to OEE?
Directly. The third factor of the OEE calculation is Quality Rate — good parts divided by total parts produced. But the connection runs deeper than the formula. Every scrapped part consumed cycle time (Performance loss), often consumed rework time (Availability loss), and in some cases required the line to stop for investigation (further Availability loss). A 2 % scrap rate rarely costs 2 % of OEE — it costs 5–8 % once the knock-on effects are counted. This is why focusing on quality alone in improvement programmes usually underdelivers; the real return comes when QC, OEE and downtime response are treated as one system. That is the thesis of my book and the operational conviction behind how SYMESTIC is built: a number reported in isolation, whether it is OEE or PPM, is always less honest than the same number seen in its full context. Real quality control in 2026 is not a department or a checklist — it is a data discipline that connects measurement, process, response and financial consequence in one loop.

Related: OEE · Six Sigma · Machine Downtime · Predictive Maintenance · Kaizen · Production Costs · Process Data · MES

About the author

Christian Fieg

Head of Sales at SYMESTIC. 25+ years in manufacturing. Previously iTAC, Dürr, Visteon, Johnson Controls. Six Sigma Black Belt, with three years in DMAIC project leadership in automotive headliner production. Former Manager Center of Excellence at Visteon responsible for global MES and traceability programmes across 900+ machines, 750+ users and 30+ processes on four continents. Author of "OEE: One Number, Many Lies" (2025) — on why a low but honest OEE is worth more than a perfect one that lies. · LinkedIn