Here is an uncomfortable observation. When Siemens and Procter & Gamble announced they would scale AI-based quality inspection across global production, the market read a software milestone. I led quality across multiple sites – more than 2,000 people, under IATF 16949 and VDA 6.3 – and I read something different: dozens of new measurement systems going live at once. Not a software rollout. A gage rollout. In a century of automotive quality, no gage has ever been validated by a license key.
An inspection model is not an app. It is a measurement device – a camera with an opinion, wired straight into your accept/reject decision. The model does not care that procurement bought forty seats. Each line it inspects is a new measurement system, and measurement systems earn trust one attribute study at a time, on the line, with that line's parts. Scale the software without the studies and you have not deployed one capability forty times. You have deployed one validated pilot and thirty-nine unvalidated gages.
A license buys capability, a study buys evidence
Under IATF 16949, every measurement device earns its study before it touches product. Calipers get a gage R&R. CMMs get correlation. Attribute systems, human eyes included, get an attribute agreement analysis: known parts with reference verdicts, repeated trials, and the two numbers that actually matter – false-accept rate and false-reject rate. False accepts are escapes: a €500 containment exercise at the customer, or an eight-figure recall if the characteristic is safety-related. False rejects are scrap and sorting. You cannot trade one for the other on a spreadsheet. You have to know both, per line, in writing.
A vendor's validation report validates an algorithm – their images, their lighting rig, their defect mix. It is not a gage study on your line. When I built the QA/QC department at a 900-person greenfield plant, every caliper, projector and CMM ran its study before release. So did every inspector, who had to pass attribute agreement on boundary samples before signing off a single part. A camera deserves the same ceremony. Swapping human inspection for an AI station is, in IATF terms, a change to the measurement system on the control plan, and for customer-notified characteristics that often means PPAP notification. The audit trail does not care how clever the model is. It cares whether a study exists for that line.
Plant forty is not plant one
The pilot worked at plant one. Of course it did. Pilots are chosen, staffed and pampered, and plant forty inherits none of that. A new hall with LED lighting is a different machine-vision problem to an old hall with sodium lamps and glare off a skylight; the pixel distribution shifts and the model shifts with it. Fixture tolerances, cleanliness, coolant film, the angle a part lands in the nest – all of it ends up in the image, and none of it was in the vendor's test set.
Then the inputs move underneath you. A new steel supplier, a tool change, a worn punch: the defect signature changes before the inspection model ever hears about it. And the defect mix itself is the quiet killer. The failures the pilot never saw are the dangerous ones, because a model confronted with a novel defect mode does not raise its hand. It calls the part good. At line rate.
I learned the input-distribution lesson twice. Once on plants: the same nominal process at two sites produces recognisably different defect populations, which is why my teams validated gages per site, not per commodity. Once in my own deployment work: I designed and built MultiPS, an orchestration platform running 63+ AI models in parallel, and I have watched the same model on a shifted input distribution behave like a different model. Not slightly worse. Differently wrong. That is why deployment, not training, is where quality is won. The vendor trained it. You run it, at takt, on your distribution.
A flag without containment is noise
Every control plan has a column people skip when they are excited: the reaction plan. What does the operator do when this gage says stop? An AI station that flags defects with no defined response is telemetry, not detection.
So the real rollout deliverables are not dashboards. Per line: a PFMEA update that treats the station itself as a failure-mode source – camera down, model version reverted, image cache full, threshold drifted. A control-plan revision that names the station as the detection control. A QRQC trigger with a named owner, so a reject spike at 03:00 starts containment instead of a discussion. An escalation path with quarantine rules for everything inspected during a suspected degradation window. If nobody can say who contains product in the middle of the night without calling the vendor, the reaction plan does not exist.
The same week the Siemens announcement ran, Volkswagen recalled 2.9 million vehicles over a steering defect. Detection is what stands between ordinary process variation and that kind of headline, and nothing in a regulated supply chain scales on faith.
Software scales in a week. Evidence scales one line at a time.
Key takeaways
- Treat every model–line pair as a new measurement system: run an attribute agreement study on known parts before the station gates product, and record the false-accept and false-reject rates per line.
- Re-run the study whenever the input distribution moves – lighting retrofit, fixture change, new supplier, tool change, defect-mix shift. The distribution is part of the gage.
- Update the paperwork per site: PFMEA, control plan, reaction plan and, for customer-notified characteristics, PPAP notification. Auditors ask per line, not per enterprise agreement.
- Define the degradation response before go-live: QRQC trigger, containment owner, escalation path and a quarantine rule for lots inspected during suspected drift.
The launch deck will count seats. Count validated lines instead – lines with a signed attribute study, an updated control plan and a reaction plan somebody can execute at three in the morning without a support ticket. None of this argues against AI inspection; the return is real, but it is denominated in escapes prevented, not seats enabled. Forty licenses is a procurement event. Forty validated lines is a quality event. One happens in an afternoon. The other takes a year of quiet, unglamorous studies – and somewhere down the road an auditor, or a recall, will ask a question for every line the launch deck skipped.