External-source research synthesis, not a Sander Gansen client result. Evidence reflects the supplied October 2026 research and source-specific dates.
Start at the actual production bottleneck
An automated welding process can still depend on a person presenting loosely stacked parts in the correct orientation. Automating that upstream task is not just image recognition. The system must separate parts, grasp them reliably and deliver the required pose within the production cycle.
What the BUMET test established
The AI-MATTERS account describes sheet-metal components and welded assemblies at BUMET’s approximately 60-person Heeze site. With Brainport Industries and Affix Engineering, the company tested picking loosely stacked components from a pallet for welding. CAD matching was feasible but insufficiently accurate. Deep learning improved recognition without reaching the required 3D accuracy; low-distinctiveness geometry and interlocking parts remained barriers.
A no-go is not a failed business
The proposed high-volume, precise and repeatable workflow was not yet reliable enough. The test identified an intermediate flex-feeder route for separation and orientation. The page does not publish a numerical acceptance threshold, test count, realised saving or current deployment of that redesign. Do not convert a useful feasibility decision into either a general failure of AI or proof that the alternative is already working.
Validate against production conditions
For an illustrative metal-component manufacturer, a test brief needs part families and variation, presentation, lighting, gripper, calibration, cycle time, acceptable failure and recovery. Define false-pick, pose-error and failed-feed measures rather than one vague accuracy number. Include low-distinctiveness and entangled worst cases, not only separated examples under perfect lighting.
Choose the capital gate before the demonstration
A demonstration can show recognition without meeting throughput, quality, downtime or maintenance requirements. A qualified integrator and the safety/quality owner must validate physical constraints. Ask what a failed pick costs, how the cell recovers and whether an operator can safely intervene. The acceptance decision should identify what must be true before capital is committed.
Use a validation route for the right job
A Testing and Experimentation Facility is a way to validate technology, not an ERP or a guaranteed production supplier. Catalogue entries are not necessarily instantly bookable, and funding eligibility must be checked for the actual project. The value of the BUMET case is an honest answer to a bounded question and a more informed next route. Good AI work can prevent an unsuitable investment as well as improve a successful process.
Sources and next reading
Explore OfferSignal · Focused Business & AI Workshop · All research
Hands-on product, business and AI leadership for executives of established companies. Working on agentic commerce through OfferSignal.
About Sander