KIE Factory Stories: Why Did the AI System Approve a Defective Part?

“The new vision system has approved all 500 parts.”

The Quality Manager looked relieved.

The factory had installed an AI-powered camera to detect surface defects. It was faster than manual inspection and never complained about repetitive work.

But a customer returned a part with a clear scratch.

The Quality Manager checked the inspection records.

The AI had marked the part as good.

“How did it miss something this obvious?”

The team reviewed the images.

The defect was clearly visible to the human eye.

But most of the images used to train the system contained clean parts and only a few examples of scratches like this one.

The AI had learned to recognize the common patterns.

It was not reliable enough at recognizing this particular defect.

The team collected more examples of scratched parts, retrained the model and tested it against a separate set of known defects.

They also kept human checks in place until the system proved reliable enough.

The results improved.

The lesson was not that AI had failed.

The team had expected more from it than its training and testing could support.

This is a critical principle in industrial AI: a model is only as reliable as its data, validation and real-world performance.

A high accuracy score does not automatically mean a system will catch every defect that matters.

AI can inspect thousands of parts in minutes.

But speed without reliable detection just helps bad parts move faster.