Should you build or buy Industrial Machine Vision Software?
Machine vision software runs the cameras, algorithms and integration logic that inspect parts on a production line: defect detection, dimensional measurement, code reading and guiding robots. Manufacturers buy it to catch quality problems in real time and to connect inspection results into PLCs and plant systems.
Buy the toolkit for stations that need deterministic measurement, OCR, code verification and fast PLC integration; that is where Cognex and HALCON earn their price. If your bottleneck is AI defect detection across many similar lines, BMW and P&G show an in-house platform is feasible, but budget for the commissioning and validation work the licence would otherwise cover.
Build it, buy it, or bridge?
When building makes sense
Building is a real option for the AI defect-detection core specifically, and large manufacturers have proven it out. BMW Group's Regensburg plant, working with the Munich startup Datagon AI, developed an in-house AI system that generates tailored inspection checklists for each of the roughly 1,400 vehicles it assembles daily, combining vehicle specifications with real-time production data. That kind of project needs a manufacturer with real AI engineering capacity and a defect-detection problem specific enough to be worth the investment. What building does not cover is everything around the detection model: camera calibration, code reading, PLC integration and the validation documentation a quality audit expects. Those are the parts that took Cognex, MVTec and Keyence years to harden across thousands of installations, and reproducing them in-house adds cost that erases much of what building the model itself saved.
When buying makes sense
Buy the toolkit when your bottleneck is deterministic measurement, OCR, code verification or fast integration with existing PLCs, which is where commercial platforms earn their price through years of hardened calibration and validation tooling. That covers most stations on most lines. The BMW example shows that AI-based defect detection is a viable build project for a manufacturer with the engineering depth to run it, but it does not show that the surrounding commissioning and validation work goes away: BMW still had to integrate its system with existing production data and camera infrastructure. For a typical plant without dedicated computer-vision engineers, buying gets a station running and validated in days rather than months, and the vendor absorbs the ongoing work of tracking new camera hardware and defect types.
The desk read
Build-versus-buy analysis for Industrial Machine Vision Software is being written. In the meantime, the framework that drives every B4 call is on the B4 Index page.
Frequently asked
Can a manufacturer build its own AI inspection system?
BMW Group did, developing an in-house system at its Regensburg plant with startup partner Datagon AI that generates customized inspection checklists for roughly 1,400 vehicles a day using real-time production data. It required dedicated AI engineering capacity most plants don't have.
What does a commercial vision toolkit like Cognex or HALCON actually provide?
Deterministic measurement, OCR, barcode and code verification, camera calibration tools and validated PLC integration, refined across thousands of installations. Those pieces are commodity functions that a self-built system would need to reproduce from scratch.
Is the detection model itself worth building in-house?
For a high-volume line with a specific, well-understood defect pattern, a trained model can be a real proprietary asset. For most stations, the model is a smaller part of the total system than the calibration and validation work around it.
How fast can a station go live with a commercial toolkit?
Once cameras are mounted, commissioning a standard inspection station can take days rather than the months a custom-built system typically needs to reach production-grade reliability, since the calibration and validation tooling is already built and tested.