PAN Lab example
Audi press-shop inspection
The inspection the model inherited
Audi's software looks for hairline cracks in pressed sheet-metal parts. Audi said it would replace the earlier check by people and cameras.
See more
Audi's crack-detection software learns from example images to find hairline cracks in pressed sheet-metal parts. Audi IT's innovation department has developed it in-house with Audi's production technology division since mid-2016. It was trained on several terabytes of images from seven presses at Audi's Ingolstadt plant and from several Volkswagen plants.
What it is to replace
In 2018 Audi inspected every part in the press shop right after it was made. People look at the parts, and cameras in the presses check them with image-recognition software. Audi said this process would soon be replaced by the new software.
So no person would check each part alongside the software. A part it rejects is in hand and easy to check again. A part it passes goes on down the line.
Audi also said the method supports its employees. The sources do not say what inspection work people keep.
How far it has gone
In 2018 Audi said it was testing the software for series production, its regular car production, at its Ingolstadt press shop. The sources behind this case do not say whether, or when, series use began.
Why the pooled images matter
Images from seven presses and several plants make the software general. They also make its blind spots general. A kind of crack it cannot see, it misses on every line it inspects.
Seven human inspectors would fail in seven different ways. One program fails in one shared way across every press.
What would stand in for the people who left
Two ordinary quality controls carry the weight here. One is a drift watch. Drift means the software gets worse as the dies, the press tools that shape the parts, wear and part designs change. The watch checks the software's calls against freshly confirmed findings on the same parts.
The other is re-inspecting a sample of the parts the software passed. A rejected part gets looked at anyway. A crack in a passed part stays hidden unless someone deliberately repeats part of the retired inspection.
The case file says a differently trained model or a periodic human audit can run the drift watch. The sources do not say whether Audi runs either control.
What the sources do and do not show
The source is a Just Auto news article of 17 October 2018 that reprints Audi's own press material. It quotes two Audi managers, one of them its chief information officer. It documents how the software was developed, how many images trained it, and what it is to replace.
It publishes no accuracy, no rate of missed cracks, and no before-and-after comparison. So how much the software helps is Audi's claim, not an audited measurement.
Research behind the Lab's industrial quality inspection cases found no manufacturer, Audi included, that has publicly blamed a shipped defect or a product recall on its AI inspection. So the risks here come from how such systems fail in general. Two stand out: the software drifting as dies wear and part designs change, and complacency over an inspection nobody performs again.
What the available tools can and cannot address
A failure pathway is a link between two parts of the network, where a mistake made by one part can be passed on to the other. A pathway is closed when mistakes stop passing along it. The parts and records still move.
This case has a budget of 9 units. Explore (No Targets) sets no targets. There, and under Service Targets Only, three tools costing 7 of the 9 units are enough to keep the mistakes on this network contained. The only set at that price is Check with a second model, Review on schedule, and Escalate checks. Under Service Targets Only, the same three also meet the service target, a floor on how much the software helps the work.
Under Service Targets Only, more tools are not better. With every offered tool at its highest setting, ignoring the budget, the software's net help to the work falls below the service target.
Under Service and Safety Targets and All Governance Targets, the targets are not fully addressable with the available tools. Those levels require every failure pathway closed. Seven stay open in every combination, even with every tool at its highest setting.
Three of the seven involve the operators: their decisions on rejected parts, the recording of those decisions, and their reading of the record at the press. Two are training: Audi IT training the software, and inspection history used to retrain it. The last two are the pooled images used for training and the software logging its own results.
No tool offered in this case closes any of those seven pathways. None adds the sampled re-inspection of passed parts either. So what stands in the way is the shape of this deployment, not a shortfall of budget.
Open this example in PAN Lab v0.1 to apply pressures and levers and watch what the system does.
What this models
This example runs on the Press-inspection-class that inherited the whole duty network: 6 components and 12 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.
Evidence base: 3 assumed · 1 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.
Show all 4 assumptions
- assumed
This example treats replacement as the deployment's defining fact. Audi said its software would replace crack inspection by people and by cameras with image-recognition software. So no person would check each part alongside it, and the operators see only the parts it rejects. The example shows the pooled images as their own source, because the sources document them as the software's defining input. They come from seven presses and several plants, which makes the software general and its blind spots shared. The example shows Audi IT's innovation department as its own group, because the sources document an in-house build since mid-2016. So upgrading the software is Audi's own choice. The example assumes a heavy inspection workload, because parts arrive at line speed. It also assumes Audi's quality organization is well resourced. What the replacement changes is where people look, at rejected parts alone, not how many people there are.
- assumed
The example includes two checks the case file names as the checks still possible after the replacement. The first is a drift watch on the software. The dies are the press tools that shape the parts. As they wear and part designs change, software trained on yesterday's presses can get worse, with no inspector left to notice first. The second is re-inspecting a sample of passed parts. A rejected part is examined anyway. A missed crack stays hidden unless the quality system deliberately repeats part of the inspection it retired. Both are ordinary quality-system work. The replacement is what makes them the only remaining coverage of this kind of crack.
- baseline
The source is a news article that reprints Audi's own press material. It documents how the software was developed, the scale of its training images, and what it is to replace. It publishes no defect-rate figures, so how much the software helps is Audi's claim, not an audited measurement. The research behind the Lab's industrial quality inspection cases found no manufacturer that has publicly blamed a shipped defect on its AI inspection. So the failures this example shows are how such systems can fail in general. It never claims that Audi's software let a defect ship.
- assumed
This example computes no product-safety outcome and no missed-defect outcome. It shows how mistakes can be passed on between Audi's software, its people, and its inspection record. The parts, and the people who ride in the vehicles, are outside it. The development history, the scale of the training images, the replacement, and the unpublished figures come from the case file. Nothing in this example computes them.
What this example does not show
Show all 2 limitations
- This example computes no product-safety or missed-defect outcome. It shows how mistakes can be passed on inside Audi's inspection work. The parts and the people who ride in the vehicles are outside it. The development history, the scale of the training images, and the replacement come from the case file, not from this network.
- The source is a reprint of Audi's own press material and gives no defect figures. So this example treats how much the software helps as Audi's claim. Research behind the Lab's industrial quality inspection cases found no manufacturer that has publicly tied a missed defect to its AI inspection. Nothing here claims Audi's software let one through.
Sources and evidence
What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.
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 automaker said the system would take 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.
empirical- Trade press Just Auto (2018, October 17). Audi develops AI software for quality inspections in press shops. https://www.just-auto.com/news/audi-develops-ai-software-for-quality-inspections-in-press-shops/
The governance shape of this deployment is inheritance rather than assistance: by being built to replace 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.
empirical- Trade press Just Auto (2018, October 17). Audi develops AI software for quality inspections in press shops. https://www.just-auto.com/news/audi-develops-ai-software-for-quality-inspections-in-press-shops/
Where this connects
Institutional pressures in this domain
- Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.
- Austerity & recovery incentives — Cost-cutting and overpayment-recovery targets tilt the system toward denial and enforcement errors.
- Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).
- Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
- Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.
All of them in context on the Industrial QA & operations AI domain page.
Levers available here and the patterns behind them
- Upgrade model — Improve the model
- Check with a second model — Cross-model verification
- Review on schedule — Oversight cadence & retrospectives
- Review the riskiest first — Risk-tiered oversight
- Check copied records — Reconcile copied records
- Keep skills sharp — Deskilling-arrest mandate
- Train the staff — AI literacy & boundary rules
- Escalate checks — State-feedback vigilance
- Pause AI on alarms — Deployment circuit-breaker