AI Quality Control with Computer Vision: Pilot Guide

AI quality control: 100% inspection does not mean 100% certainty

The worst computer vision project starts with one sentence: „AI will check every product and find every defect.” A slogan like that ignores the light on the line, the dust on the lens, new product variants and the way errors are counted. When 99.5% of production is good, a model that always says „OK” has 99.5% accuracy and detects no defect at all. So the pilot ends with a decision table that holds two numbers: how many defects the system let through, and how many good items it stopped for no reason.

The system assesses the photo, marks the place of the suspected defect and sends the item to an inspector. The inspector sees the photo with the marking and approves or rejects it. Sampling does not disappear after the rollout: a person still checks random items the system judged good, so you know how many defects the model lets through.

The quality department sets the definition of a defect with examples: what counts as a scratch and what is a tool mark. Without that definition the model learns from inconsistent labels, and an argument about the pilot result turns into an argument about the definition.

What computer vision quality control costs and when it pays back

The count covers the work around each case: searching, reading, writing and handing over. What remains is the time to check the result, 35% of today's work in the scenario. Before the quote we measure both times on 20 to 100 cases the team has already handled, in a free test that takes 5 working days.

When not to roll out computer vision

When defects are rare and cheap and a visual check takes seconds, the system will not pay back. When conditions on the line change from shift to shift, first stabilise the lighting and the camera position. When a defect only shows after cutting or measuring, a camera will not see it. Photos from the line can include employees, so you set the frame and the storage period with the GDPR in mind, and Article 14 of the AI Act describes human oversight of high-risk systems.

Where to start

The first scope covers one line, one product and two or three defect types, with a sample labelled by the quality department before the start. We build this kind of system as an AI app. For the free process review, a description of one case from the last few weeks and the tools it went through is enough. You do not need to send confidential documents, and you receive a written plan within 2 working days.