The inspection is automated. The vigilance isn't.

This week the money said machines will do the looking. Assembler AI closed a $4m seed round for its manufacturing computer-vision platform. Oxipital AI launched a foreign-material detection capability, and Huawei and NISCO reported AI running quality checks inside a steelworks. Three signals, one direction: machine vision is moving into the inspection jobs plants used to staff with people. Here is an uncomfortable observation. I spent two decades staffing and auditing those walls in automotive and aerospace. The industry is now paying to remove its cheapest quality sensor without ever having written down what that sensor did.

What the wall did besides pass or fail

Final inspection looks like one job. It is at least two. The first is conformance: does the part match the defect classes, the gauge, the boundary samples, the surface criteria? High volume, repetitive – and a camera does it better. Machines do not tire in hour nine of a shift. They do not soften their criteria after lunch.

The second job never appeared in any control plan I audited. At Witte Automotive we ran a Q-Wall as the front end of our QRQC discipline, and the operator's real value was not catching the defect in front of her. It was noticing that the whole tray looked different from Tuesday. Zinc-flake plating half a shade colder. A faint witness mark on a radius that had been clean for two years. Every dimension in spec, every control chart flat – and she would still pull the batch. More than once the QRQC she opened with a phone call to the supplier found a changed passivation chemistry or a new steel coil entering the chain a full day before any data would have moved. A meaningful share of our failure-cost reduction came from catches like that. Not one of them appears in the inspection instruction.

When I built the QA/QC department for a 900-plus-employee greenfield plant at SNOP, I learned the lesson from the other side. The specification covered measuring. The tribe covered noticing. Writing the job descriptions forced me to admit how much of the control system had lived nowhere but in people's heads.

Classifiers are not curious

I design and deploy multi-model AI systems myself, so I have no patience with salesmanship or with nostalgia. A classifier is a set of learned boundaries around the defects it has seen. Show it a part and it answers exactly one question: does this resemble a known-bad pattern? An out-of-distribution defect – the kind that arrives with a changed batch, a new supplier, a tweaked chemistry – does not present as an alarm. It presents as nothing-known, and many architectures resolve that ambiguity as a confident OK.

The new foreign-material detection products are excellent across the catalogue they were trained on: the stones, the plastic shards, the wire fragments. A contaminant outside the catalogue is a stranger the system waves through, or confidently mislabels as product. The steelworks deployments fare better, because rolled surface defects form a closed and well-photographed world. A tier-one automotive supply chain does not. I trust any model I deploy exactly as far as its training data and no further. A confidence score tells you how closely the input resembles that data. It says nothing about whether the part is good, and those are the same question only in a factory where nothing ever changes.

A classifier answers yesterday's question. Vigilance is noticing the question changed.

Keep the human, change the job

The answer is not to keep the wall as it was. Redesign the control plan so the undocumented function becomes a documented one, with an owner, a cadence and a budget.

• Adversarial sampling. Seed known-bad parts – including subtle, near-miss ones – through the automated station on a schedule and measure the catch rate. If the audit never finds a miss, you are not sampling hard enough.
• An auditor-of-the-auditor tier. Keep a smaller, sharper human group whose job is hunting what the machine cannot classify. Start with the reject pile: novelty shows up in rejected images first, and a weekly review of them is the cheapest drift radar you will ever install.
• Instrument the vision system itself. Trend reject rates and confidence distributions daily. Challenge it with golden samples every shift. A distribution that tightens suspiciously on a new batch is not good news – the machine recognises the world rather too well.

Then write the removed vigilance into the PFMEA and price it. Name the line item what it is – ambient anomaly detection, removed by automation – and put a euro figure on severity. One escaped batch runs €40,000–€60,000 in sorting, containment, expedited freight and 8D paperwork before anyone mentions a warranty claim. What carries a number gets staffed.

At SNOP, a quarter with zero critical customer escalations was not won by our measuring. It was won by early catches from people whose job title said inspector and whose actual function was pattern recognition across batches. Automate the measuring, by all means. But if the redesign deletes the noticing, you have swapped a scheduled cost for an unscheduled one.

Every inspection automation I have run or audited removed noise and signal together – deletion does not distinguish. The plants that came out ahead kept a smaller, sharper human tier and instrumented the machine as aggressively as the product. A camera has perfect attendance and no hunches, and you are only buying the first of those. It will never ring a supplier because today's parts look wrong. Someone still has to.