Every AI vision inspection pitch I have sat through includes the same slide: a human inspector with a magnifying glass, crossed out, next to a robot with a checkmark. The implication is clear. AI will replace your inspectors. It is faster, more consistent, never gets tired, never takes a break. Why would you keep humans?
After deploying AI vision systems across three aerospace manufacturing plants in the last three years, I can tell you that slide is wrong. AI does not replace inspectors. It exposes them. It exposes which inspectors were actually doing their job, which were rubber-stamping, and which inspection processes were theatre all along. The AI does not eliminate the human. It eliminates the illusion that the human was effective.
The first deployment: where we learned humility
Our first AI vision system was deployed on a surface finish inspection line for machined aluminium components. The line had four inspectors working in two-hour rotations — a standard fatigue management practice. Their first-pass agreement rate — the percentage of times two inspectors looking at the same part agreed on accept/reject — was reported at 94%. Excellent, we thought. The human baseline was strong.
Then we ran the AI system in shadow mode for thirty days — inspecting every part, logging its decision, but not acting on it. The AI's agreement rate with the final disposition was 81%. Thirteen points lower than the human self-reported agreement rate. We investigated.
It turned out that the 94% agreement rate was measured during audit conditions — when inspectors knew they were being observed. During normal production, the actual agreement rate, reconstructed from the AI shadow data and a review of retained samples, was closer to 76%. The inspectors were not dishonest. They were human. They adapted their behaviour when watched, and they drifted when not watched. The 94% number was real, but only for the parts inspected during audits, which was about 5% of production.
AI inspection does not fail at what humans do well. It fails at what humans do poorly: sustaining attention on the ten-thousandth part of the shift.
What the AI actually caught
Over the thirty-day shadow period, the AI flagged 247 parts that passed human inspection. We reviewed every one. Of those 247:
— 189 were genuine defects that the inspectors missed. Most were minor surface anomalies — scratch patterns, tool mark variations, texture deviations at the edge of the acceptance criteria.
— 38 were defects that the inspectors saw but judged acceptable, disagreeing with the written standard. This was the most interesting category. The inspectors had developed an informal "practical tolerance" that was wider than the engineering specification. They were not missing the defects. They were redefining the criteria based on experience. In some cases, they were right — the engineering specification was overly conservative. In other cases, they were wrong, and the parts would have failed at the customer.
— 20 were false positives from the AI.
The 189 missed defects were the expected finding. Fatigue, repetition, time pressure — these are known human limitations. The 38 "judgement overrides" were the unexpected finding. They revealed that the inspection process had drifted from the specification, and nobody knew it. The AI did not create this drift. It had been happening for years, invisible, because nobody had a tool that could measure inspector behaviour at scale.
The inspector response — and why it matters
When we shared the findings with the inspection team, the reaction was exactly what you would expect: defensiveness, then denial, then grudging acceptance. But the most revealing comment came from the senior inspector, a woman with twenty-two years of experience: "Of course we miss things. You try staring at aluminium surfaces for eight hours and see how sharp you are at hour seven."
She was right. And that is the point. The problem was never the inspectors. The problem was a system that asked humans to do something humans are fundamentally bad at — sustained, high-accuracy visual discrimination under time pressure — and then measured them only during audits.
The second deployment: where AI changed the process
In our second deployment — a composite layup inspection — we changed the approach. Instead of running AI in shadow mode, we ran it as a first-pass filter. The AI inspected every part. Parts that the AI flagged as borderline — confidence between 40% and 85% — were routed to a human inspector for final disposition. Parts above 85% confidence were auto-accepted. Parts below 40% were auto-rejected.
This reduced the human inspection volume by 70%. But here is the critical part: the humans were not removed. They were concentrated on the cases where their judgement added value — the borderline, the ambiguous, the cases where contextual knowledge and experience outperform a pattern-matching algorithm.
The inspectors' job changed from "stare at every part" to "investigate the parts that the system cannot resolve." Their engagement went up. Their skill utilisation went up. And the overall defect detection rate — AI plus human — improved by 34% compared to human-only inspection.
What AI inspection cannot do
AI vision systems are excellent at detecting patterns they have been trained on. They are terrible at detecting patterns they have not been trained on. When we introduced a new composite material variant that produced a different surface texture, the AI's false positive rate spiked to 18% because it had never seen the new texture and flagged normal variation as anomalous.
AI cannot negotiate acceptance criteria with engineering. It cannot pick up a part and feel the weight, flex it, listen to it tap. It cannot look at a defect and say, "This is the same failure mode I saw three years ago on the old machine — has anyone checked the coolant?" It cannot connect a surface anomaly to a root cause in the process upstream. It can only see the surface. The inspector sees the surface, the context, the history, and the implications.
The honest truth about cost
AI inspection systems are expensive. The hardware, the integration, the training data, the validation, the ongoing maintenance of the model — all of it costs more than most plants are willing to admit in their ROI calculations. I have seen business cases that assume 90% inspector reduction. In my experience, the realistic reduction is 30-50% of inspection hours, with the remaining hours redirected to higher-value tasks. The business case works, but not at the numbers the vendors project.
And there is a cost that does not appear in any spreadsheet: the loss of institutional knowledge that flows from experienced inspectors. When you reduce the inspection team, you reduce the number of people who have seen ten thousand defect patterns and can tell you not just that something is wrong, but why it is wrong and what to do about it. That knowledge is not captured by the AI. It walks out the door when the inspector retires.
What I tell my peers
Deploy AI inspection. It works. It will improve your defect detection rate, reduce escape rates, and free up inspector time for higher-value work. But do not deploy it as a replacement for inspectors. Deploy it as a tool that makes your inspectors more effective, more focused, and more engaged. The plants that get the best results from AI inspection are the ones that redesign the inspection process around the AI's strengths and the human's strengths, not the ones that try to eliminate the human.
The AI is a very fast, very consistent, very limited inspector. It needs a human partner who can do what it cannot. If you remove that partner, you get speed without wisdom. And in aerospace, speed without wisdom kills people.
