The visibility problem
While machines and sensors capture temperatures, speeds, and outputs, the human-led processes that sit between them remain largely invisible. Manual assembly, inspection workflows, and operator activity rarely appear in any dashboard or report. For manufacturers running large orders with a mix of automated lines and manual steps, this blind spot shows up as missed deadlines, inaccurate capacity planning, and no clear path to improvement.
What computer vision actually is
Computer vision is a measurement tool, not a magical solution. It turns visual signals into structured data: counts, times, and classifications. It only sees what it has been trained to see, and it will not fix a broken process. What it will do is quantify that process with a precision that was previously impossible. The difficulty rarely lies in the modelling itself. It lies in defining the problem, controlling the environment, and being clear about what decision the output is meant to inform.
Where it fits
Computer vision has three natural landing zones in manufacturing:
- Automated quality inspection, where high-speed cameras identify defects that human inspectors either miss or cannot verify at production speed.
- Process and operations monitoring, where manual assembly steps can be measured for productive time, cycle variability, and bottleneck identification, effectively delivering an OEE metric for work that was never measurable before.
- Safety and compliance, where the goal is risk reduction and early intervention, not surveillance. Bringing operators into the conversation early and being transparent about what is being measured — and what is not — is essential to successful deployment.
"If you can see a process but can't measure it, we are happy to help."
A real world example
In one deployment, DataQI used a single fixed camera above a manual assembly jig to track productive time and identify which step in the process was being performed. Two lightweight models running simultaneously delivered a working prototype in weeks. The constrained environment was a feature, not a compromise. Fewer variables meant a simpler model, faster build time, and fewer failure modes.
Making it stick
The difference between a computer vision project that delivers long-term value and one that quietly dies as a pilot comes down to design. Solutions built around actual workflows have defined owners, clear exception handling, and assigned stakeholders. Those built purely to prove technical feasibility rarely survive past the demonstration stage. Operators are not just end users of these systems. They are the most qualified people to identify edge cases, contextualise anomalies, and validate that what the model is seeing reflects what is actually happening on the floor.
