The tech crosses over. The sample size doesn't.

Hyundai Rotem has announced it will run its aerospace plant on smart-manufacturing technology imported from Hyundai Motor, and the coverage treats it as a novelty – sensors and dashboards hopping industries, digital quality heading upstream. I read it as something older. It is the crossing every quality leader who has moved from automotive into aerospace has made by hand, without a manual and without applause. The digital layer will install in a quarter. The statistical assumptions underneath will not install at all: they were validated on millions of parts, and aerospace hands you lots of twelve. I have made this crossing carrying QRQC boards into EN 9100 territory. Some of what you carry survives. Some of it lies to you politely for years.

What crosses intact

My career is the case study. At WITTE Automotive, QRQC, A3 and a Q-Wall pulled failure costs down hard – the kind of reduction that shows up in the monthly numbers. At SNOP I built a greenfield QA/QC department for more than nine hundred people and cut defect costs by 70%. Then came the other side of the fence, AS9100 and EN 9100 territory, where the lots shrink and the paperwork grows. What crossed unchanged was never statistical: QRQC, A3, layered process audits, escalation discipline. Those are behavioural systems. They encode who reacts, how fast, and who owns the problem until it is dead. Volume is irrelevant to them.

I once shipped an automotive reaction cadence – containment inside the hour, root cause on the wall inside 48 hours, verified countermeasure inside a week – into an AS9100-certified supplier machining housings in lots of forty. The paperwork moved slower; configuration control and first-article discipline see to that. The behaviour held completely. When an escaped defect surfaced, suspect stock was walled off before lunch. Reaction speed is worth more at low volume, not less, because nothing else is standing guard.

Behaviour scales down. Statistics don't. Buy the behaviour, rebuild the statistics.

What breaks: statistics without volume

Cpk is the honest example, because it is the number executives trust most. Capability is an estimate built from a mean and a sigma, and on a lot of twelve your sigma rests on eleven degrees of freedom. The interval around it is wide enough that a displayed 1.67 is compatible with a process that is barely capable and one that is heroic. The dashboard still shows one decimal place. That is confidence without information.

SPC fails from the other side. Control limits mean something when subgroups come from a stable, common-cause stream; aerospace lots change configuration, material batch and tooling between orders, so drawing limits across that mixture puts a border around noise. The standards admit the difference themselves. IATF 16949 is written around volume, sampling and ppm logic. AS9100 around configuration control, first-article inspection, traceability and concessions.

Then the machine learning. Automotive anomaly detection – torque signatures, press-fit curves, weld currents – earns its keep on millions of baseline cycles. Give the same model three hundred cycles spread across fourteen configurations and it either alarms on everything or learns silence. Neither failure appears in the demo. Gage R&R is quieter about its assumption: the ten sample parts are meant to span the process spread, and in aerospace those ten parts are the production run.

The asymmetry nobody ports

In automotive, a false alarm costs a sorted batch and a line stop – painful in euros, fully recoverable – while a missed signal usually gets caught downstream by volume, so the economics push you to tune alarms until the line keeps flowing. Aerospace inverts the cost function. A false alarm costs an MRB stack and a schedule slip. A missed signal costs a concession at best and an escape at worst – and an escape multiplies by the fleet. The narrow-body defect currently making headlines, some five hundred in-service aircraft, is not a story about missing dashboards. It is a story about what one escaped defect costs when the denominator is a fleet.

So the loop has to be redesigned around irreversibility, not merely connected. In low-volume production you accept a high false-alarm rate on purpose: one confirmed anomaly becomes a containment event even when you suspect the alarm is lying, because scrapping twelve parts is cheap and living with an escape is not. Connected-operations platforms speed up whatever loop you already have. If that loop is tuned for automotive flow, you have built a faster way to make the wrong call.

The question that ends the demo

Here is the question every plant manager should put to any vendor selling an automotive-proven platform into aerospace: what lot size was it validated on? Ask for the validation data – reference volumes, subgroup sizes, baseline cycle counts. Most will struggle, because their reference customers are tier-one automotive plants running million-piece programmes, and nobody inside those plants ever had to ask. The phrase automotive-proven is doing load-bearing work it was never designed for. Ask it and watch the demo run out of time. Then split the contract in two. The behavioural modules – reaction management, escalation, 8D workflow – can port today. Every statistical module gets re-baselined on your first six months of real production before anyone trusts the screen.

Hyundai Rotem will be judged, initially, on whether the screens light up and the plant hums. Wrong metric. The digital layer was never the risk; the risk is statistical machinery quietly running on assumptions that ended at the factory gate. Management systems transfer. Statistical models must be rebuilt. The vendors cannot tell you which is which, because nobody has ever asked them. The people who can tell you are the ones who ran both volume regimes with their own hands. Done honestly, the plant ends up with automotive reflexes and aerospace arithmetic, and that combination wins. The reverse – aerospace reflexes, automotive arithmetic – is how proven technology earns an unproven reputation.