A representative body-in-white line runs at sixty vehicles per hour: twelve robotic welding cells laying down roughly 540,000 spot welds per shift, with an AGV fleet moving sub-assemblies between stations.
Weld quality is monitored the way it is monitored almost everywhere — manual destructive sampling, one part per shift. That is about 1,800 welds inspected out of 540,000 laid down: 0.3% coverage, with the verdict arriving hours after the welds were made.
Sampling catches drift eventually; it cannot catch a defect population that lives in the 99.7% of welds nobody looks at. In a representative late-2025 field campaign, an under-penetration defect on a single weld type made it through sampling for months — and surfaced as a recall of 12,000 vehicles, at a cost of €38M in warranty.
The bitter part is that the information existed at the moment each weld was made: weld current and voltage signatures, electrode force, the visual state of the seam. Nothing was watching at line rate, and the decision window is brutal — at sixty vehicles an hour there are about 50 milliseconds to disposition a weld before the next part arrives at the cell.
We were inspecting 0.3% of the welds we laid.
This is a problem brief, not a customer case study. It describes how Vertex-edge is designed to work on this problem — not a deployment that has already happened.
Modelled on the reference line above, where weld-attributable warranty ran €6.3M a year before the recall. The model is dominated by warranty-event avoidance, not labour savings — with full inspection and closed-loop feedback to the welding controller:
Modelled figures from a scoped and costed reference architecture — not a completed customer deployment. One avoided recall pays for the engagement many times over.
