How AI Visual Inspection Works on the Factory Line, and What Inspectors Still Do
Cameras that judge parts and paint finishes are among the longest-running uses of computer vision in factories. This explains how a vision inspection system sees, decides and acts, what BMW and Schneider Electric report, and the steps people still handle.
Checked against BMW's Regensburg video and Spartanburg release, Cognex's May 2026 release and its Schneider Electric story, and the chocolate-mould experiments
Role scorecard
Inspectors, Testers, Sorters, Samplers, and Weighers SOC 51-9061.00
- Core tasks AI does in production2 of 4
- Most mature evidenceIn production
- CostNot disclosed; Cognex says its cloud platform cuts scaling costs by up to 50% (vendor-stated)
- BenefitSchneider Electric: yield doubled, false rejects down 70 times (operator quote and vendor-written story)
As of · Tasks: O*NET 31.01
| Core task (O*NET) | What AI or robots do | Extent | Evidence grade | Example case | Source |
|---|---|---|---|---|---|
| Inspect, test, or measure materials, products, installations, or work for conformance to specifications. | AI checks every painted body surface for deviations without a person looking; staff check if necessary | Does it | In production | BMW Group plant Regensburg paint shop: painted surfaces "checked fully automatically using AI methods" (BMW-stated, May 2023) | 2, 3 |
| Record inspection or test data, such as weights, temperatures, grades, or moisture content, and quantities inspected or graded. | Each scanned body gets its own stored profile of deviation locations and classes, which drives the rework | Does it | In production | BMW Regensburg: a digital 3D image of each deviation is created, classified and stored (trade press reporting BMW) | 3 |
| Discard or reject products, materials, or equipment not meeting specifications. | The system flags parts that fail; the sources do not say who removes them, so we count it as assisting | Assists | In production | Schneider Electric reports false bad rates down 70 times on contactor lines with Cognex OneVision (vendor-written story; quote from the operator) | 4, 5 |
| Remove defects, such as chips, burrs, or lap corroded or pitted surfaces. | Four robots sand and polish the flagged spots; people finish body edges, door joints and the fuel filler cap | Assists | In production | BMW Regensburg: robots rework in the order an AI assessment sets; staff handle what robots cannot (trade press reporting BMW) | 2, 3 |
Extent and grade follow our evidence grades. The two "Does it" rows describe the paint shop only: BMW says its AI checks the painted surface fully automatically, but a person may still check a body if necessary, and the occupation covers many other products and checks. We show four of the 31 core tasks O*NET lists for the occupation; the rest (measuring with calipers, reading blueprints, sampling, weighing, stacking and similar) have no AI case in the sources we opened.1
Costs and benefits
| Price | Not disclosed in any source we opened; Cognex says its cloud-to-edge platform lowers scaling costs by up to 50% (Cognex-stated)4 |
|---|---|
| Yield and false rejects | Schneider Electric: yield doubled and false bad rates reduced by a factor of 70, inspection coverage widened from 5 to 17 control areas, cycle time 400 ms to 200 ms (a story written by Cognex and republished by a trade site; the operator's own quote says "double yield, dramatically reduce false rejects")45 |
| Development time | Essity: a sealing inspection that took more than a year to tune with its previous approach was built and demonstrated in less than a day (Essity quote in a Cognex release)4 |
| BMW | BMW reports more consistent surface quality; we found no published defect-escape rate or baseline to compare with23 |
As of · Sources: BMW, Cognex, trade press and the cited research234567
| Case | Who states it | Evidence grade | What the system does | Result as stated | Source |
|---|---|---|---|---|---|
| BMW Group Regensburg paint shop | Operator | In production | AI checks painted surfaces, robots rework them, staff handle edges and joints | Up to 1,000 vehicles a day pass through finishing; no defect rate published | 2, 3 |
| BMW Group Spartanburg (AIQX) | Operator | In production | Visual and acoustic quality inspection with camera systems and sensors; feedback to line staff on smart devices | BMW says it has made AIQX a standard and is assessing offering it to suppliers | 6 |
| Schneider Electric with Cognex OneVision | Operator quote, vendor-written story | In production | Learned and rule-based inspection of contactor assembly in 17 control areas | Yield doubled; false bad rates down 70 times (vendor-written story) | 4, 5 |
| Essity with Cognex OneVision | Operator quote in vendor release | Prototype | Sealing inspection application built and demonstrated | Built in less than a day against more than a year before; no production use stated | 4 |
| Cognex OneVision adoption | Vendor | Vendor claim | Customers using the platform since its June 2025 beta | More than 100 customers, many moving from single-line to multi-site rollouts (Cognex-stated, May 2026) | 4 |
| Human and AI together (chocolate-mould experiments) | Research | Prototype | Two lab experiments on how highlighting the AI's reason affects people deciding if a mould holds faulty bars | Highlights sped people up; misplaced highlights raised errors because people searched less | 7 |
The results are not comparable: they cover different products, checks and measures, and none comes from an independent audit. Vendor-written figures are labelled and kept apart from operator statements. Search results for this topic are dominated by vendor blogs quoting unsourced human-versus-AI accuracy figures; we did not use any we could not trace to a primary page.
How it works
Seeing: light first, then the camera
Most of an inspection system's accuracy is set before any model runs, by how the part is lit and imaged. BMW's Regensburg paint shop uses deflectometry: black and white stripe patterns are projected onto the vehicle, and cameras look for places where a flaw disturbs the stripes on the reflective paint.3 The cameras register where each flaw is and the data goes to a computer system, which builds a digital 3D image of the deviation, classifies it and stores it.3 Computer vision, the field behind this, is covered in the glossary.
Deciding: rules and learned models together
Schneider Electric's contactor lines combine traditional rule-based vision tools with AI tools in one environment, according to the story Cognex wrote about it.5 Engineers label images in detail, mark the region of interest and retrain the model, and the story says models can be trained from small image sets and refined step by step.5 The inspections include height control, screw position and rotation, sorting, orientation and detecting shortages at the base or head of a contactor.5 The labelling is where people's judgment goes in: BMW's planner says the paint system's algorithm was built on the knowledge of the whole team, and now "decides which features need post-processing".3
Where it runs: cloud to train, edge to inspect
Cognex describes a cloud-to-edge split: models are trained, managed and governed in the cloud, while inspection runs on the vision system at the line, so that once a model is deployed no cloud connection is needed and production images stay local.4 Schneider Electric's story says this removed the need for PCs with high-performance graphics cards.5
Acting: reject, mark or rework
A finding must lead somewhere. At Regensburg, the scan data sends four robots to the flagged spots to sand, apply polishing compound and polish, in an order set by an AI assessment, so each body is worked on only where needed.3 The press video BMW published says the surface treatment "is also processed fully automatically", with employees checking bodies if necessary.2 Robots cannot reach body edges or the area next to door joints and the fuel filler cap is too fragile, so staff finish those places, guided by a laser projector that marks the issues on the sheet metal.3
What gets measured, and what is missing
The figure vendors quote most is the false reject, a good part flagged as bad; Cognex's Schneider Electric story calls it the "false bad rate" and reports a 70 times reduction.5 A false reject wastes good product; the opposite error, a bad part passed, is what quality systems exist to prevent, and the sources we opened give no escape rate for any of these systems. BMW reports consistent quality but no baseline.23 Claims of 99% AI accuracy against 80 to 87% for people circulate widely; we found them only on vendor and aggregator pages without a traceable study, so we leave them out.
Humans and machines together
Whether people work better with an AI's help depends on the AI's reasons being right. In two experiments simulating chocolate-factory quality control, participants decided whether moulds held faulty bars, always told what the AI had classified. Highlighting the image areas behind the AI's decision speeded people up, but when a highlight marked an intact area while the fault was elsewhere, error rates rose, and eye tracking showed people searched the rest of the mould less.7 The authors conclude that such support can discourage people from investing effort in their own analysis.7
What people still do
In the cases above a person still sits at the edges of the process: BMW staff finish what robots cannot reach and check bodies if necessary,23 and BMW says Spartanburg's AIQX gives line employees immediate feedback rather than replacing them.6 For how we grade evidence, see how we measure AI in roles.
Sources
- O*NET OnLine, 51-9061.00 – Inspectors, Testers, Sorters, Samplers, and Weighers (O*NET 31.0 tasks; 2025 employment 602,000 and median annual wage $48,570 per its summary), retrieved 2026-10-26.
- BMW Group PressClub, Automatic surface inspection and rework in the paint shop, BMW Group plant Regensburg, video, May 2, 2023.
- Vasilash G. S., CompositesWorld (Gardner Business Media), Robots, AI and Superb BMW Surfaces, 2023 (a trade-press report quoting BMW staff; cited as that source because bmwgroup.com did not respond).
- Cognex Corporation, Cognex OneVision Adoption Ramps as Manufacturers Scale AI Vision Globally, press release, May 13, 2026 (PR Newswire copy).
- Cognex, as republished by Metrology and Quality News, Schneider Electric Sets a New Benchmark for AI-Driven Inspection, June 18, 2026 (a customer story written by Cognex; the 70 times figure is the vendor's).
- BMW Group PressClub, BMW Group advances the use of Physical AI in production with Figure 03 project in Spartanburg, press release, June 25, 2026 (AIQX paragraph).
- Human Factors and Ergonomics in Manufacturing & Service Industries 34(5), The benefits and costs of explainable artificial intelligence in visual quality control: Evidence from fault detection performance and eye movements, arXiv:2310.01220 (v2).