The vision system detects defects in real time. Real time is still downstream.

Here is an uncomfortable observation. In one week, three machine-vision announcements landed in my feeds: a plastics-industry piece celebrating systems that detect defects in real time, a roundup of ten vision companies said to be transforming manufacturing, and an automation heavyweight showing inspection systems to the semiconductor crowd in Taiwan. Every headline sells the same product. Faster eyes. Not one mentions where the defect was born. Two decades in automotive and aerospace quality, and I have yet to see a camera stop a tool from wearing. Faster eyes find failure sooner. They do not make it rarer, and the difference between those two sentences is measured in euros.

The arithmetic of an escape

Run the numbers the press release skips. A short shot scrapped at the press costs you the material and thirty seconds of an operator's attention: call it €0.50. The same defect caught at end-of-line costs rework, handling, re-inspection and a buffer full of suspect stock, ten to fifty times as much, having eaten cycle time you never get back. Escape to the customer and the curve goes vertical. Containment sorting, 8D documentation, expedited freight, quarantined lots, engineers travelling to explain themselves. I have watched single escapes consume more money than a year of vision-system maintenance, on systems installed to prevent exactly that escape.

Real time does not change where on the curve you sit. It changes how fast you sit there. Milliseconds of detection latency are irrelevant when the defect was created a press cycle, or a tool-maintenance interval, earlier. Detection speed truncates the tail of a bad run: two hundred suspect parts become twenty. That is worth something, and I will not pretend otherwise. But twenty parts is still twenty you paid materials, energy and labour to manufacture, and will now pay again to scrap.

Real time is a property of the loop, not the camera — and a loop that closes after the part exists is just a museum with an alarm system.

What the camera cannot see

Cavity imbalance. Progressive tool wear. A material lot that behaves differently at the gate. Cycle drift that looked innocent shift by shift. These are the origins of most defects I have chased, and none of them live in a pixel. They live in cavity-pressure curves, clamp tonnage, shot weight, hold time, the force-displacement signature of an assembly station.

Process signals lead. Pixels confirm. By the time the camera sees anything, the process has been telling the story for hours or days, to anyone with instrumentation wired to a control chart.

This is textbook PFMEA discipline, and it is rarely applied. Prevention controls reduce occurrence; detection controls only reduce the odds of shipping the failure. A plant that stacks detection controls on a characteristic whose occurrence is driven by an unmonitored process variable is renting peace of mind. At WITTE Automotive, the substantial failure-cost reduction we delivered through QRQC, A3 and the Q-Wall came almost entirely from moving control upstream, onto tool-side and process-side signals such as tonnage trends and force curves, not from making end-of-line inspection faster. The end-of-line cameras kept doing their job. They simply had less to do.

Put the camera on the process, not just the part

The strongest case for vision is the one the vendor demo never makes: use it as a drift sensor. A system that measures a critical dimension on every shot and feeds the result into an SPC chart is no longer a gatekeeper. It is a high-frequency process sensor that alarms drift before tolerance is breached and points maintenance at the offending cavity or station. That is where 100% automated inspection genuinely belongs, on drifting characteristics you are actively managing, wired into the control loop rather than bolted to the end of it.

I ran into this exact trade building the greenfield QA/QC department for a 900-plus-employee plant. We designed the entire inspection strategy before a single camera was ordered, and the hardest, highest-value decision was what not to run at 100%. Every automated gate you add downstream is a standing invitation for the process to keep drifting. The safety net quietly becomes the plan. We pushed containment upstream, keyed to process signals, and the 70% defect-cost reduction followed from that move. We ended up with more cameras than a zero-based design would predict, but they were pointed at the process, measuring it — not at its output, sorting it.

Two questions for the vendor

Vision AI has entered procurement cycles, and this quarter's capital requests are being written. So before the demo, before the ROI slide, I ask two things. First: what is your escape rate, the probability a real defect passes your system undetected, and how was it validated — on whose parts, at what severity? Second: what does a false reject cost me per week in €, in good parts scrapped, re-inspection labour and stoppage, at my throughput and my part mix? Catch rate is a brochure number. Escape rate and false-reject cost are plant numbers. A system catching 99.5% of defects while rejecting 2% of good parts on a high-volume line can quietly cost more than the defects it finds.

Celebrating real-time detection is celebrating your reflexes in a crash you caused. The plants I respect buy instrumentation first, SPC discipline second and cameras third: measurement for the loop, not absolution at the end of it. When this quarter's capital request lands on your desk with a camera line and no line for cavity-pressure sensors, you will know exactly what is being purchased. Sharper pictures of a process nobody is listening to. The defect is born at the press. Buy the things that hear it there.