Some jobs require a person to look at everything and miss nothing. No one can do that for a whole shift, and the defects that get through are the expensive ones. Vision models can hold that kind of attention indefinitely - and every correction the inspector makes becomes a training label.
We have put this into production four times, in settings that look nothing alike: a European stainless steel producer watching two kilometres of moving metal, a vehicle fleet assessing damage on cars coming back off lease, a customs authority inspecting goods at the border, and a state police force detecting illegal content in seized material. The asset changes, the economics do not.
Every one of these engagements began the same way. Someone is required to inspect a stream of things - coils, cars, crates, files - and is physically unable to look at all of it properly. So they sample, or they hurry, and the organisation absorbs whatever gets through.
This is not a skill problem, and treating it as one is how these projects usually go wrong. It is an attention problem with an expensive tail: the defect costs cents to catch at the point of inspection and thousands once it has moved downstream. A flaw rolled into a coil has tonnes of steel wound on top of it. A scratch missed at lease return is a dispute with a customer. A crate waved through customs is a compliance failure.
What a vision model changes is not accuracy in the abstract. It changes the job from watching to reviewing - and reviewing is something people are good at.
At a European stainless steel producer, a finished coil runs to around two kilometres in length and a metre and a half across. An operator watches it pass and marks surface defects - for a whole shift.
The decision the system rests on is where the cameras go. They are not at the operator's position; they are five minutes upstream of it, measured in line speed. That gap is what converts a detection into something still actionable, because the steel the model just looked at has not reached the desk yet. Put the cameras at the desk and you have bought an expensive second pair of eyes with no more time to think than the first pair had.
A car comes back off lease and someone has to decide what the damage is worth. Done by hand it is slow, inconsistent between assessors, and the source of most of the arguments that follow - because the customer and the fleet rarely see the same car.
Automated damage assessment from images cut the cost of that process by 80%. The gain is not only speed: a model applies the same standard to the thousandth vehicle as to the first, which is what turns an assessment into something defensible in a dispute.
Customs is the purest version of the problem: far more goods arrive than can be opened, so inspection is a sampling exercise and the question is only ever which crates are worth the officer's time.
Vision models applied to inspection imagery reduced inspection time by 70%. That figure is really a statement about targeting - the same officers, pointed at the consignments that merit opening.
For a German state police force, investigators had to review seized media by hand to find illegal content. The volume is large, the work is distressing, and the backlog is what decides how fast a case moves.
The detection method developed for it is patented. It is also the clearest demonstration of why these systems are built to triage rather than to decide: the model orders the queue, an investigator makes every determination that matters.
Four deployments, four industries, one capability - and in each of them the measurable change is the same one: the people doing the inspecting stopped having to watch everything, and started reviewing what the model had already flagged.
The pattern transfers wherever someone is obliged to look at everything, a miss is cheap now and expensive later, and the people doing the looking can correct the model. We have not delivered the settings below - they are where the same approach applies, and where most of the enquiries we get come from.
Two things decide whether one of these works, and neither is the model.
The first is where the camera goes, and how much time that buys. Detection with no time to act on it is a report, not a system. On the steel line the whole design followed from five minutes; if your process gives you five seconds, we would tell you so before you spent anything.
The second is whether the people doing the inspecting today will correct the model tomorrow. That correction loop is what keeps these systems from decaying in their second year. Where the inspection role is being cut at the same time as the model goes in, the labels stop, and so does the accuracy.
Our founding team designed and delivered all four of these deployments in a senior role at a global technology leader. This is the experience we bring to ProDataAI client engagements.
Let's talk about what's achievable for your specific context.
Start the Conversation →