10 to the 23 AI logo

Domain Atlas / Industrial QA & operations AI

Case fileGermany (an automaker's press-shop inspection; reprinted corporate press material)giant deployment

The inspection the model inherited

Explore this deployment in the PAN Lab ↗

A large automaker developed an in-house deep-learning system to detect hairline cracks in pressed sheet-metal parts, trained on several terabytes of images drawn from seven presses at its home plant plus several sister plants, in development since mid-2016 and tested for series deployment. The documented change is a generational replacement: the system takes over an inspection duty previously performed by manual visual checks plus fixed-rule camera systems, rather than augmenting a human inspector's judgment on each part. The record — a reprint of the manufacturer's own press material with its CIO quoted — documents the development lineage, the data scale, and what the system replaced; it publishes no quantitative defect-rate figures, so the deployment's benefit magnitude is a corporate claim, not an audited measurement.[]

What happened

A large automaker built its own deep-learning system to find hairline cracks in pressed sheet-metal parts. The training corpus is unusually well documented for this domain: several terabytes of images, drawn from seven presses at the automaker's home plant and from several sister plants in the wider group, with development running since mid-2016 and the system tested for series deployment. This is an in-house build, not a vendor purchase — the automaker's own IT function developed it, and its CIO is the voice quoted in the record.

The change that matters for governance is what the system replaced. Press-shop crack inspection had two prior generations: a person looking at parts, and fixed-rule camera systems checking them against hand-written criteria. The deep-learning system replaces both. That is a different relationship to human judgment than this domain's other named deployment, where the AI flags and a worker on the line responds to each flag. Here the learned system inherits the inspection duty itself for the defect class it covers — there is no per-part human check running alongside it, because the per-part human check is the thing it replaced.

Inheritance has consequences the record lets us name precisely. First, the pooled training data: images from seven presses and several plants make the model general, and they also make its blind spots general — whatever this one model cannot see, it cannot see on every line it inspects, which is a different risk shape from seven human inspectors with seven different sets of habits. Second, the verification asymmetry: a flagged part is physically in hand and easy to re-inspect, but a passed part flows on, and the discipline of re-checking a sample of passes belongs to the plant's quality system now, not to an inspector whose job was exactly that.

What the record does not contain is any quantitative defect-rate figure — no accuracy, no escape rate, no before-and-after delta. The source is a reprint of the manufacturer's own press material, so the deployment's benefit is documented as real enough to develop for years and test for series use, and its magnitude is a corporate claim. And as everywhere in this domain, no named manufacturer has publicly attributed a shipped defect or a recall (the share of real cases the model actually found) to its AI inspection, so the failure modes are understood at the mechanism level: the model drifting as dies wear and part designs change, and the organizational complacency that grows over an inspection nobody re-performs.

The honest reading is that this is a serious, multi-year, in-house industrial deployment whose defining governance fact is replacement — a learned system inheriting an inspection duty from both a person and a rule-based machine — with the benefit magnitude unpublished and the risks living in drift and in the quiet disappearance of second looks.

The sociotechnical reading

This case is the industrial-QA domain's replacement portrait, and it completes a pair with the domain's response-loop anchor. There, the AI flags and a resourced human responds to each flag; here, the learned system inherits the whole inspection duty from the manual check and the rule-based camera generation it replaced. The map's instruction is to read those as two different governance shapes, not two intensities of the same one: an assist changes what a person attends to, a replacement changes whether anyone attends at all, and the controls that keep each honest are different.

For a replacement, the load-bearing controls are the ones that stand in for the judgment that left. Drift monitoring carries more weight here than anywhere in the domain, because dies wear, part designs change, and a model trained on yesterday's presses degrades quietly — and there is no per-part human check left to notice. Sampled re-inspection of PASSED parts is the second control: a flagged part gets looked at by construction, but an escape is invisible unless the quality system deliberately re-performs a fraction of the inspection it retired. Both controls are ordinary quality-system machinery; the point is that the replacement makes them the only human eyes remaining on this defect class.

The pooled training data adds the domain's correlation lesson. One model trained on images from seven presses and several plants inspects them all with the same learned blind spots — a defect class it systematically misses, it misses everywhere at once. Seven human inspectors fail in seven uncorrelated ways; one model fails in one correlated way at fleet scale. The map reads this as the industrial version of the monoculture pathway it draws elsewhere, and the answer is the same: an independent check — a second model trained differently, a periodic human audit — whose errors do not correlate with the first.

The Lab network models only the deploying organization: its detection model, its press-shop operators, and its inspection records. No product-safety or defect-escape outcome is computed on any diagram. The parts and the people who eventually ride in the vehicles are boundary-only; the development lineage, the data scale, the replacement, and the unpublished magnitudes are institutional signals that live in this case file, never on any network. The map's instruction is to treat replacement as a distinct governance shape, to resource the drift monitoring and sampled re-inspection that stand in for the judgment that left, and to keep the benefit's unpublished magnitude explicit.

The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.

Grounding sources for this case

The same sources that ground this model organization in the PAN library: evaluations, government documents, investigative reporting, and advocacy documentation, each labeled by tier.

leanenterpriseinstituteGroundingReferenceSave

Lean Enterprise Institute. Automatic Line Stop (Lean Lexicon). https://www.lean.org/lexicon-terms/automatic-line-stop/

https://www.lean.org/lexicon-terms/automatic-line-stop/

Appears in: PAN framework development

Grounds: domain grounding: industrial operations and QA (visual inspection, predictive maintenance); model org: audi_press_shop_inspection; model org: bmw_aiqx_inspection

Seeing your organization in this case file?

The histories here are documented after the harm. Mapping a live deployment's pathways and pressures, before the incident report, is engagement work: intake, diagnosis, prescription, and monitoring, with every limitation stated.

Sources & Evidence

Claims made on this page and what supports them. The full registry lives in Evidence.

EmpiricalA large automaker developed an in-house deep-learning system to detect hairline cracks in pressed sheet-metal …

A large automaker developed an in-house deep-learning system to detect hairline cracks in pressed sheet-metal parts, trained on several terabytes of images drawn from seven presses at its home plant plus several sister plants, in development since mid-2016 and tested for series deployment. The documented change is a generational replacement: the system takes over an inspection duty previously performed by manual visual checks plus fixed-rule camera systems, rather than augmenting a human inspector's judgment on each part. The record — a reprint of the manufacturer's own press material with its CIO quoted — documents the development lineage, the data scale, and what the system replaced; it publishes no quantitative defect-rate figures, so the deployment's benefit magnitude is a corporate claim, not an audited measurement.

EmpiricalThe governance shape of this deployment is inheritance rather than assistance: by replacing the manual visual …

The governance shape of this deployment is inheritance rather than assistance: by replacing the manual visual check and the fixed-rule camera generation, the learned system inherits the whole inspection duty for the defect class it covers, so there is no per-part human judgment running alongside it to catch what it misses. Its training data is pooled across presses and plants, which means one model's blind spots are correlated across every line it inspects. The failure regime is mechanism-level — drift as dies wear and parts change, complacency over an inspection nobody re-performs — because no named manufacturer, including this one, has publicly attributed a shipped-defect escape to its AI inspection.