Siemens and Procter & Gamble have announced they are scaling AI-based quality inspection across global production, with scrap down by up to 20% where the systems run. Serious companies, serious engineering, and I have no reason to doubt the number. Here is my uncomfortable observation after twenty years of standing over red bins: scrap is the cheapest metric in a plant to move, and the announcement everyone celebrated quietly relocated the accept/reject decision into a software setting that no control plan I have ever audited contains. The dial came preset. Nobody asked who set it.

Most scrap is fear, not defect

Decompose any scrap rate and you find two components of very different character. The first is true defect cost – parts that genuinely fail specification. The second is over-rejection: parts killed because nobody could defend letting them pass. In most plants I have assessed, the second is embarrassingly large. It is pure fear.

The asymmetry is structural. A missed defect lands on a name – the operator, the quality engineer, the shift that signed the release. An over-scrapped part lands nowhere.

Nobody gets called at 22:00 over a good part they killed. So a rational operator under a customer escalation scraps everything marginal, and everyone downstream of the red bin does the same. At Witte Automotive, running QRQC and A3 discipline on the bins, I learned the rule early: operators do not scrap what is defective. They scrap what they cannot defend. The plant pays for that defencelessness daily, in perfectly good parts.

That is why the 20% shows up fast. You cannot engineer real defects out of a process in a quarter – that takes PFMEA work, process changes, supplier development. Fear-scrap can be switched off almost immediately, because it was never physically bound to anything. At SNOP, a large share of our 70% defect-cost reduction came from killing over-rejection: tighter measurement, clearer boundaries, parts we could stand behind. The AI inspection wave monetises the same euros we recovered with discipline and arguments. The machine has no fear at 03:00, and it never gets tired.

Every model ships with a dial

An inspection classifier is a control plan wearing software. Beneath the network sits a threshold: the score above which a part is rejected. Move it one way and false rejects fall; scrap drops, savings appear, someone gets a slide at the quarterly review. Move it the other way and false accepts fall; escapes drop and scrap rises. Producer risk traded against consumer risk, exactly as in any sampling plan – except the textbook assumed a human set the threshold deliberately, after a validation study. In most deployments an integrator set it to a default before handover, and nobody has touched it since.

A threshold nobody owns is an engineering change nobody approved.

When I built the QA/QC function for a 900-plus-employee greenfield plant, I learned where acceptance criteria actually live: not in an operator's judgement and not in a camera, but in the control plan, the PFMEA and the reaction plan – documents with owners, revisions, signatures. Every criterion was defended in review, because moving it changes the product the customer receives. A software threshold deserves the same ceremony. Every re-tuning is an engineering change with re-validation attached. Otherwise it is nothing.

I say that as someone who builds these systems. I designed and run MultiPS, a platform that runs more than sixty AI models in parallel and reconciles their outputs. I trust the engineering more than most people in this debate – and I still would not let one of my own classifiers set acceptance criteria unowned. A model is an instrument. Instruments do not get to decide tolerances.

Pair the scrap number with the escape number

The governance fix is cheap and boring, which is why it works. Publish the escape rate and the customer-complaint trend beside the savings, on the same slide, every month. A scrap figure alone is marketing; paired with escapes, it becomes management information.

The cost asymmetry justifies the pairing. One scrapped part costs its value once – say €12 for a machined component. One escaped lot pays for the 8D, the containment, the sorting at the customer's dock and, in automotive and aerospace, sometimes a stopped line. I have led enough escalations to price the cheap ones at €100,000-plus before anyone counts the meeting hours. A threshold optimised for weekly scrap quietly reprices that tail risk, and the repricing never appears in the savings report.

Two further moves. Add PFMEA lines for model drift and for part or surface changes – a new coating, a new alloy, a new supplier shifts the feature distribution the model was validated on, and yesterday's confident classifier becomes a coin flip. Then take three questions to management review: who owns the threshold; when it was last moved, and against what validation; and whether a vendor's model update can change our acceptance criteria without an engineering change note. The debate about trusting autonomous systems can continue in the background. Inspection systems are not waiting for it. They are already accepting and rejecting parts.

Take the savings – name the owner

The 20% is real and the systems are good. Recovered fear-scrap is still recovered money; I have banked it myself, one defended part at a time. But scrap is the metric plants love because it is visible, bounded and paid in-house. Escapes are paid by the customer and invoiced back later, with interest.

Key takeaways

  • Put the acceptance threshold in the control plan with a named owner, revision history and validation evidence; treat every re-tuning as an engineering change with re-validation attached.
  • Never report scrap savings alone – publish escape rate and complaint trend on the same slide, every review cycle.
  • Add PFMEA lines for model drift and for part, coating or supplier changes that shift the distribution the model was validated on.
  • Ask who last moved the dial, when, and whether vendor model updates can change acceptance criteria without a change note.

So take the savings – they are earned. Then walk to the engineering office and ask one question: who owns the threshold? If nobody in the room can tell you who last turned the dial, the 20% is not a saving. It is a deferred payment on your next customer escalation, and the interest rate is set by your customer, not by you.