At a major car factory, the quality team spotted a troubling pattern. Even with hundreds of cameras and layers of manual checks, paint flaws were slipping through unnoticed until the very end of the line. By then, fixing them meant extra work, delays, and mounting pressure on the team to put things right quickly. The obvious question emerged: could AI help out by using data from earlier in the process to flag these problems before they spread?
At the time, our data expert Erik Boom was working at the factory in a hybrid role that combined data engineering and data science. He helped set up the project in its early stages, which focused on testing whether AI could support the quality control team in reacting faster.
“The paint process has many layers. Each one can go wrong in a different way,” Erik explained, adding, “We wanted to see if data could make that easier to manage.”
The project began as a small-scale test, part of the factory’s new model for trying out innovation: small budget, short timeline, clear scope. The goal wasn’t to build a full product. It was to learn fast and see if data could make a clear difference.
The paint shop followed a strict inspection routine. Each car went through camera checks, followed by manual reviews at several points. Still, some paint flaws were only detected after multiple cars had already moved down the line. The aim wasn’t to change where problems were found, but to reduce the delay, catching an issue after five cars instead of ten, for example.
The quality control department raised the concern. They were seeing recurring issues but lacked a way to respond quickly. Once a defect was found, it was hard to tell how far the problem had spread.
The team wanted to narrow the focus. Instead of checking every car the same way, they wanted to find patterns. If certain types of cars or colors had more flaws, that could help target inspections earlier and reduce waste.
The team decided to narrow their focus. Rather than inspecting every car in the same way, they looked for patterns. If certain models or paint colours showed more defects, inspections could be targeted earlier, helping to reduce waste.
“The idea was to highlight cars with higher risk of paint issues, so checks could be more focused,” Erik explains.
“At first, the team turned to camera data, but the volume was too large to process in real time. They decided instead to rely on manual quality checks. Some process data was also missing because of broken sensors, which were repaired once identified.”
Once the basics were in place, the team was ready to move from reviewing data to trying out a real test in practice.
The factory had an experimental setup for testing new ideas, designed to keep projects simple, low-risk, and fast. The spray prediction project fit neatly into that approach.
From the beginning, the team worked to manage expectations. A full root cause analysis wasn’t realistic as there weren’t enough time, budget, or complete data to trace issues back to a single robot or process step. Instead, the focus shifted to finding patterns.
“We weren’t trying to prove one clear cause,” Erik added. “We wanted to show that certain combinations of conditions tend to show up when defects appear. That alone can be very useful.”
To explore the problem, the team used a method called associative rule mining. The aim wasn’t to pinpoint a single cause, like “machine 12 created the flaw,” but to spot conditions such as “this type of flaw is more likely when machine 12 runs slowly and this model is on the line.”
They paired this with a complete process map built from trace data, which showed how each car moved through the paint shop and where data points were missing.
“We built a complete animated map of the paint process based on the actual flow data,” Erik says.
“It helped people see what was really happening, not just what was supposed to happen.”
This work had an immediate effect. At one point, five cars disappeared from the data between two process steps. The system flagged the gap, and when the team checked the shop floor, they found the cars parked off to the side. They had been pulled from the line during a machine stop but never put back. The process map made that problem visible.
For alerts, the team avoided dashboards and chose simple, direct signals instead. The system raised a flag whenever the same type of flaw appeared several times in quick succession, since that often pointed to a bigger issue.
As Erik put it, ‘if you see the same kind of flaw three times in a row, it’s probably not a one-time issue.’ These alerts helped operators act faster without needing to dig through reports or dashboards.
The experiment produced two key tools. First, a working model that flagged early signs of repeating defects. Second, a process map built from live data that showed how the paint shop actually operated. Both uncovered gaps that had previously gone unnoticed.
The early warning system proved especially useful. It quickly flagged recurring issues, giving teams the chance to act before the same problem spread further down the line.
“It showed that in many situations, the system could have caught problems sooner than the manual checks.”
Every car was still inspected manually, and that step remained part of the process. The system didn’t replace manual checks; it simply helped teams focus their attention earlier.
The process map added a different kind of value. It gave the process department a clear picture of what was actually happening on the floor, not just what the paperwork suggested. It showed how cars moved through the shop, where data was missing, and where the flow didn’t match expectations.
The results showed that even with limited time and tools, data could help the car manufacturer make faster decisions. It eased pressure on quality control, built trust in digital tools, and delivered clear value without creating extra work.
The experiment showed that even small changes can deliver strong results. By giving operators straightforward alerts and giving managers a realistic view of how the process worked in practice, the car manufacturer was able to catch issues earlier and cut down on waste.
The main benefits were:
This experiment showed how data and AI can support day-to-day work on the factory floor. By focusing on patterns instead of perfection, the team helped operators respond sooner and gave managers a clearer picture of what was happening in the paint shop.
The approach didn’t rely on complex systems or lengthy training. It worked because it was simple and addressed a real problem.
“It gave people something useful without asking them to change how they work.”
For factories facing late-stage quality issues, the takeaway is clear: even a small experiment can deliver real value if the data is reliable, the goals are well-defined, and the outcome helps people act more quickly.
Ready to explore how AI and data can create value in your business?
Start the conversation with COMPUTD and discover the possibilities in a free consultation.
