The zeros are rented. The defects are owned.

Here is an uncomfortable observation: every decade the industry re-packages detection technology as the end of defects, and every decade the pitch finds fresh buyers. This week's trade coverage of zero-defects technology – machine vision with AI bolted on, presented as the moment inspection finally abolishes the defect – is the oldest promise in quality, wearing better cameras. I fell for a version of it once, early on. The demo was flawless. It ran on good parts. My line was not producing good parts at anything like a stable rate. The shop floor corrected me within a quarter: a warehouse of sorted parts, plus a piece of arithmetic I now run before anyone opens a capex file.

What the scanners actually buy

Run the numbers honestly. Say your process generates 1,200 ppm on a recurring mode and you install a €400,000 vision gate at 98% effectiveness – generous, on a floor with lighting drift, part presentation variation and three shifts. You have bought 24 ppm of guaranteed escapes, plus a permanent handling cost for every defect you do catch: sorting, rework, disposition, records.

The defect did not get cheaper. It got a fixed price.

Worse, you have converted an engineering problem into an operating cost that scales with your own failure. If the process drifts – and an unstable process is exactly the one people buy walls for – the scanner catches more, the cost line grows, and the generation rate sits exactly where it was. Then come the licences, the calibration contracts, the spare cameras and the false rejects: turn sensitivity up to chase the last ppm and you start scrapping good parts. The vendor's deck never shows the year-four renewal slide.

Zero defects found is not zero defects made.

That sentence belongs on every purchase order for detection hardware. Compare the two prices that matter: € per escape caught, paid forever, versus € per failure mode eliminated, paid once. They are rarely within an order of magnitude of each other.

The audit was never fooled

Nothing in the standards rewards a control plan made entirely of end-of-line gates. In a PFMEA, detection is the weakest lever you hold. A wall at the end scores poorly no matter how many megapixels it carries; a process change that cuts occurrence is what actually moves the assessment. A VDA 6.3 auditor walks your line backwards from the gate and asks one quiet question: where does the control sit relative to the cause? If every answer is "at the end", you will bleed points on process control. You will deserve to.

Customer SQEs discount your vision wall for precisely this reason. A screen at the end of the line tells them you expect escapes – you have industrialised the expectation. I have sat on both sides of that table. The supplier with boring upstream controls and a modest gate routinely out-scores the one with an impressive wall and nothing behind it.

Where the real zeros come from

The zeros worth having come off the generation rate, and nothing in a camera touches that. At the SNOP greenfield – 900-plus people, a QA/QC department I built from an empty office – the make-or-buy question arrived in the first capex round: a full detection line on day one, or measurement at the source with the discipline around it. I held the scanners back. The money went on capability at the point of manufacture: SPC where the part was made, layered process audits, poka-yoke where the physics allowed, and a QRQC loop that put a named owner on any defect within 24 hours. The 70% defect-cost reduction we delivered came out of that loop and the upstream fixes it forced, not out of filtration. We ran a quarter with zero critical customer escalations on the strength of it.

Witte Automotive taught me the containment side of the ledger. The Q-Wall caught what the A3 teams had not yet killed, and it earned its keep as a safety net. The metric we watched was the wall's catch rate going down. Every recurring mode the A3 work eliminated meant fewer catches. Fewer catches was the win.

I deploy this class of system myself – I run a platform orchestrating more than sixty AI models in parallel – so I can say precisely what pattern recognition cannot do: it cannot change why the part is wrong. The right place for the technology is as the sensor feeding the improvement loop – instant, contextual data on failure modes at the station, in the shift it happens, not three weeks later in a Pareto chart. The moment it becomes the wall that hides the loop, you have bought the subscription.

2027 capex planning is opening now, and the zero-defects decks are already in circulation. Apply the rule I used at the greenfield build-out and have never had to apologise for: every euro on detection paired with a euro aimed at the cause, in the same budget line, on the same page – approve them together or reject them together. Otherwise you are not buying zero defects. You are renting zeros at a rising price from a process you already own, and calling the arrangement transformation.