AI quality control: what automated visual inspection would really take
A 3.5-month pilot turned photos the company already collected into a reviewable quality signal and established what reliable automated visual inspection would cost.
The situation
The company manages cleaning operations for car-sharing fleets across several European countries, using a large network of freelance workers. After cleaning each car, workers were required to photograph the finished vehicle. Those photos were stored but never systematically reviewed: manual review would have required supervisors to inspect hundreds or thousands of images each day.
When quality problems surfaced, they usually appeared through customer complaints or occasional checks of individual workers. The question was whether an AI system could review the photos and flag poor work before a worker finished the job.
The constraint
The hard part was not processing an image. It was whether a model could reliably recognise the specific dirt that mattered under real operating conditions: different vehicle angles, day and night, winter and summer, and every lighting condition in between. The relevant cases were rarely obviously filthy cars; they were small areas a cleaner had missed.
No public dataset came close. Automotive datasets cover vehicles, damage, dents, and scratches — not subtle cleaning quality — so the dataset had to be built and labelled from the company’s own images.
We tested three approaches. General-purpose generative-vision models were ruled out within about a week because the same image could receive different assessments across runs. That is unsuitable for quality control. A managed computer-vision service on Microsoft Azure was more deterministic, but its rectangular bounding boxes made it hard to represent irregular patches of dirt. The model could associate a vehicle part, such as a rear bumper, with dirt rather than learning the shape of the dirty area itself.
A custom YOLO-based model in Roboflow allowed dirt to be labelled with polygons. It offered the control and domain-specific accuracy the problem required, in exchange for substantially more work on dataset preparation, labelling, review, and iteration.
That trade-off was the finding.
What we built
We built a vision pipeline over the photos workers were already taking. Each image was first classified by vehicle angle, so the system knew which part of the car it was reviewing and managers could see whether workers had captured every requested angle. It then entered a second AI-vision stage for visible dirt and missed areas.
The dataset used historical company images, manually reviewed and labelled with auto-labelling support, and balanced across angles, seasons, lighting conditions, and cleaning scenarios. During the pilot, inference ran on dedicated cloud infrastructure rather than Roboflow’s hosted service, making batch throughput and cost controllable.
We also built a management dashboard for cleaning activity by location and worker. Managers could review the previous day, week, or a custom period, compare the share of vehicles flagged, and see how results varied between workers. A mobile form replaced the spreadsheet managers had used while inspecting new workers.
Process diagram
Where it stands
The pilot ran for several months. It established that the company’s existing photos could become a usable quality signal, and later model iterations produced results managers could act on. It also established the volume of domain-specific data, labelling, and continued refinement needed for reliable production use.
Angle classification proved reliable in the pilot. Dirt detection required more labelled examples, so the client paused further development while weighing that investment against the operational benefit.
Next
Research that updates itself. Judgment that does not.
Ready to boost business results with AI automation?
Book a 30-minute operations review. We will look at where manual work is limiting your operation and whether there is a practical opportunity to create capacity.
Book a 30-minute operations reviewSwiss precision. Valley speed.