The robot is consistent. So is the defect.

Every wave of automation coverage — the F-35 line is the current example — sells the same virtue: machines that never tire, never vary, never hand a sloppy Friday afternoon to the next shift. Fine. Here is the uncomfortable observation from the shop floor: automation does not eliminate defects. It converts random ones into systematic ones.

I have stood at the end of a line on a Monday morning where an automated cell had run from Saturday 06:00 to Sunday night without a single alarm. The first pallet held 1,900 parts, all of them wrong in exactly the same way. No operator variation to break the pattern. No handover to catch the change. The robot worked the weekend without complaint. That was the problem.

Random defects protect you, systematic ones bury you

Human scatter is an accidental detection system. A tired operator produces one bad part in ten, then a good one, and the noise flags itself — on an SPC chart, in an audit sample, in the feel of a torque wrench that did not pull right. Sampling plans and control limits were quietly designed around that scatter. A drifted robot produces the same defect every cycle, at rate, with variance so tight your chart reads it as capability. I have seen Cpk improve while every part on the pallet was bad. The chart was measuring consistency. Not correctness.

Now do the maths. A cell at 60 parts an hour, 32 unsupervised hours, drift starting at 07:00 on Saturday after a sensor glitch — call it 1,900 suspect parts. The containment bill writes itself. Internal sorting at €3 a part, €6,000. Stop-ship and 100% re-inspection across two shifts. Expedited freight, €6,000–8,000. The customer's line stop, billed at €10,000–15,000 an hour if you are tier one. Then the 8D team, the sorting at the customer's dock, the penalty. Sixty to a hundred and twenty thousand euros before anyone mentions reputation. The human version of the same failure costs three parts and a conversation. Same defect, different physics.

The robot does not make mistakes. It makes copies.

Your PFMEA assumed a human

Pull the PFMEA for any automated cell and check what disappeared against what arrived. Gone: fatigue, the skipped torque step, shift-to-shift variation, the misread drawing. Arrived: the wrong program revision loaded after a maintenance window, calibration drift on weld current, the silent parameter edit by someone chasing cycle time, sensor lag, the slowly dirtying lens on a vision system that starts passing everything.

Occurrence ratings drop for the old failure modes. Detection ratings get worse for the new ones, because every detection method in the document was calibrated to human noise — audit sampling, visual checks, limits set wide enough to absorb a person. The two numbers move in opposite directions, and the RPN arithmetic flatters you.

When I built the QA/QC department for a 900-plus-employee greenfield plant at SNOP, from a blank sheet, the control plans for the automated cells were written before the robots were bolted down. That discipline, with QRQC run at full rigour, took 70% out of defect cost while customer satisfaction sat at 98%. Not one euro of that came from automation being safe by default. It came from assuming the machines would fail systematically — and watching them accordingly.

Reaction plans built for rate

When drift is systematic, containment is a batch decision, not a part decision. Three disciplines decide whether you recall 300 parts or 1,900.

• First-article validation after every program or parameter change. Full layout, documented, against a named revision — not a glance at the first part out of the cell.
• Traceability granularity. Records that tie parts to shifts leave you owning a shift of escapes. Hour-stamps and program-revision logs bound to serial numbers are the only lever that shrinks the batch.
• QRQC when detection lags the drift. The clock starts when the drift started, not when you found it. Quick response means reconstructing the timeline from program logs, not arguing about whose alarm should have fired.

At WITTE Automotive the QRQC and A3 work cut our failure costs hard, and the distance between one bad part and one bad pallet was repeatedly one discipline. A maintenance reload put the wrong weld schedule into a cell over a Sunday. First-article validation at 06:05 on Monday turned a 1,900-part escape into one scrapped part and a revision-control fix.

A quarter with zero critical customer escalations did not arrive by luck. It was a trained reflex.

I design and deploy autonomous AI systems myself, and the failure modes that keep me awake are never effort or concentration. They are design gaps in detection. The current wave of AI inspection helps only if the system knows what your drift looks like. A model that has never seen a specific failure mode will pass it, at full confidence, every cycle.

Automation is a quality-system decision before it is a capex decision. Before the next cell is ordered, reread the control plans for the hidden assumption of human variability — the sampling frequencies, the limits, the reaction plans that bet on people being inconsistent. Strip that bet out deliberately. The robot is consistent. So is the defect.