01Executive summary
Two milestones eleven years apart complete the transformation the American System began: first make the work flow, then make the process predictable.
The preceding series covered interchangeable manufacture, in which the engineering effort moved from the product to the system that makes it. This part covers what happened next. Ford’s Highland Park plant put the work on a moving line in 1913, cutting chassis assembly from many hours to a small fraction of that. Walter Shewhart at Bell Labs sketched the control chart in 1924, giving quality a mathematical basis for the first time. Together they define how physical goods have been made ever since.
02The moving line and its real mechanism
The common explanation is that moving the work saved the walking time of the assembler. That is true but minor. The substantial effect is that a moving line imposes a fixed cycle time on every station simultaneously, which converts an unmeasured, self-paced process into a synchronised one where imbalance is immediately visible.
Once the cycle is fixed, three things follow automatically. Work content must be divided into increments that fit within it, which forces detailed method analysis. Any station that cannot keep up stops the whole line, which makes local problems everybody’s problem within seconds. And inventory between stations is squeezed to almost nothing, which removes the buffer that previously hid variability.
Line balancing becomes the design task
Total work content divided by takt time gives the theoretical minimum number of stations. Real allocation is constrained by task precedence and by tasks that cannot be split. The gap between theoretical and achieved station count is the balance loss, and reducing it is where industrial engineering effort concentrates.
The slowest station sets output
A line runs at the rate of its bottleneck, so improvement anywhere else adds nothing. This is the origin of constraint-based thinking, and it applies unchanged to software delivery pipelines, approval workflows and hospital patient flow.
Buffers hide problems
Inventory between stations lets an unreliable station keep the line fed and therefore keeps its unreliability invisible. Deliberately reducing buffers to expose problems is the central idea of just-in-time production — and it is an information strategy, not a cost strategy.
Flexibility must be designed in
A line optimised for one product at one rate is brittle. Changeover time, mixed-model sequencing and the ability to re-balance for a different takt are design requirements, and were the weakness of the original Ford system.
The Highland Park system delivered extraordinary productivity on a single unchanging product, and was poor at anything else. Retooling for a model change effectively meant rebuilding the plant. It also imposed severe, well-documented conditions on the people working the line, and high labour turnover was an early and persistent problem. The engineering achievement is real; treating it as an unqualified template is not supportable.
03Shewhart and the two kinds of variation
Shewhart’s contribution is easy to state and surprisingly hard to internalise. Every process varies. That variation has two sources, and they demand opposite responses.
The control chart exists to tell the two apart. Limits are set from the process’s own observed variation, conventionally at three standard deviations from the centre. A point outside the limits, or a non-random pattern within them, signals a special cause worth investigating. Everything else is the process being itself.
If an operator adjusts a machine every time a measurement drifts from nominal, the adjustment adds its own error to the process’s natural variation. The output variance increases — often substantially. This is tampering, and it is the single most common misuse of measurement data in industry. It is also why control limits must never be confused with specification limits: control limits describe what the process does, specification limits describe what the customer requires, and the relationship between them is the entire subject of capability.
| Property | Control limits | Specification limits |
|---|---|---|
| Source | Calculated from the process’s measured variation | Set by design intent, function or contract |
| Question answered | Is the process stable and behaving as itself? | Is this part acceptable to the customer? |
| Changes when | The process genuinely changes | The requirement changes |
| Plotted on | Control charts of process output over time | Capability studies and inspection records |
| Common error | Drawing specification limits on a control chart | Widening them to make a process appear capable |
04Capability: stability first, then conformance
A process must be stable before capability means anything. Capability indices compare the spread of a stable process against the tolerance band, and compare the process centre against the tolerance centre. If the process is not stable, its spread is not a fixed quantity and the index is a number without a referent.
Relationship between spread, centring and yield
The middle case is the instructive one. A process with excellent repeatability but a systematic offset produces defects at one limit while wasting margin at the other. It is usually the easiest problem to fix, because a centring offset is a single adjustment, whereas reducing spread requires changing the process. Diagnosing which of the two is dominant before acting is the practical value of the whole method.
Observed variation is the sum of process variation and measurement variation. If the gauge contributes a substantial share, the control chart is partly plotting the gauge. Assessing repeatability and reproducibility of the measurement system before drawing conclusions about the process is not optional, and it is skipped constantly. Related guidance sits in the AS/NZS ISO 9001 family and in the ISO 3534 and ISO 7870 series on statistical methods and control charts — cited by number only, verify currency.
05What survived, and what did not
Both milestones have been substantially revised by later practice, and it is worth being clear about which parts held.
Flow, takt and the bottleneck
Synchronising work to demand rate, exposing imbalance, and concentrating improvement at the constraint remain correct and have generalised far beyond manufacturing.
Common and special cause
The distinction, and the discipline of not reacting to noise, is as valid in service metrics, safety statistics and financial variance analysis as on a production line.
Rigid single-product lines
Mixed-model production, rapid changeover and modular platforms replaced the fixed line. Flexibility turned out to be worth more than the last increment of throughput.
Deep task fragmentation
Splitting work into the smallest possible increments maximised short-run efficiency but produced quality, turnover and improvement problems. Broader roles with authority to stop the line proved more productive overall.
06Takeaways for current practice
- Improve the constraint or do not bother. Effort spent anywhere but the bottleneck changes cost without changing output.
- Do not adjust in response to common-cause variation. Tampering provably increases variance. Establish stability before attributing meaning to any single result.
- Never plot specification limits on a control chart. They answer different questions and mixing them is the most common analytical error in quality work.
- Qualify the measurement system before trusting the data. A substantial share of apparent process variation is frequently the gauge.
- Buffers are an information decision. Inventory between stages buys stability at the price of concealing the problems that make it necessary.