Ask most operations how accurate their stock is and you will get a number from a cycle count. Ask them how often a picker arrives at a location and finds something that is not what the screen said, and you will usually get a shrug and a story.
Those are two different questions, and only one of them is being measured.
A count checks quantity. It does not check identity
A cycle count answers “are there twelve here?” It does not answer “are these twelve the right thing, and is this the right place for them?”
So the failure mode that survives every count is the one where the numbers are perfect and the goods are wrong:
- Two variants of the same part sit in adjacent bins. Someone puts away into the neighbour. Both locations still count correctly. Both are now wrong.
- A returned item goes back to a location that looks right and isn’t. Quantity restored, identity broken.
- A retail facing gets fronted with whatever was nearest at close. The shelf is full, the planogram says something else, and the count agrees with neither.
None of these show up as a discrepancy. They show up weeks later as a short pick, a wrong delivery, or a customer complaint, by which time the trail is cold and the cost is somebody else’s problem.
Why nobody catches it
Because catching it means looking at every unit, and looking at every unit is exactly the work no operation has spare hands for.
The checks that do exist are sampled: a supervisor spot-checks a lane, a team leader walks the aisle on a Friday. Sampling finds systemic problems. It reliably misses the single mis-slotted pallet, which is the one that actually costs you, because a single mis-slot is invisible right up until the moment it is expensive.
This is the shape of problem a camera is genuinely good at. Not because it is cleverer than the person — it is not — but because it does not get tired, does not skip a lane at the end of a shift, and can be asked the same narrow question ten thousand times a day without its standards drifting.
The question, stated narrowly
“Is the item at this location the item the system says is at this location?”
That is a comparison, not a judgement. The system already knows what should be there. The camera reports what is there. The interesting output is only ever the disagreement.
Stated that narrowly, three things become true at once:
It is tractable. Distinguishing your own catalogue under your own lighting is a far easier problem than open-world recognition, and it gets easier the more your items differ visually — which, for most catalogues, they do more than people assume.
It is cheap to evaluate. You do not need a system to find out whether it would work. You need a few hundred photographs of your own locations and an afternoon.
It produces evidence, not just alerts. Every check keeps its image. When the disagreement is real, you can see it. When the model is wrong, you can see that too — which is the only way anyone ever comes to trust one of these.
Where it actually pays
Four moments, in rough order of return:
- Putaway. Catching a mis-slot at the moment it happens costs a re-scan. Catching it at the pick costs a delayed order. Catching it at the customer costs the customer.
- Goods-in. What arrived, against what the ASN said arrived — including the case where the label is right and the pallet underneath is not.
- Replenishment and fronting. Whether the facing matches the plan, checked continuously rather than on the merchandiser’s next visit.
- Returns. The highest-risk restock in the building, because the item has already been handled by someone with no reason to care where it goes back.
Where it does not
Be equally clear about this, because vendors rarely are.
If your items are visually identical and distinguished only by a code — two cartons of the same size, same print, different revision — a camera reading the code is a barcode scanner with extra steps, and you should buy the scanner.
If the location is a bulk pile rather than a discrete position, “what is at this location” stops having a single answer.
And if your locations are unreliable in your own system, a camera will simply tell you that in higher resolution. That is worth knowing, but it is a data problem and it will not be solved with a lens.
A test you can run this week
No budget, no vendor, no project.
Pick thirty locations at random. Not the ones you suspect — random, or the exercise measures your hunch instead of your operation. Go and look at each one. Record two things: whether the item matches what the system says, and whether the quantity matches.
Then compare the two error rates.
If quantity is accurate and identity is not, you have the problem this article is about, and you now have a number for it. If both are accurate, you have better discipline than most operations and your money belongs somewhere else — which is a genuinely useful thing to have proven for the cost of an hour’s walk.
Either way you will know something you did not know on Monday, and you will not have bought anything to find it out.
If the walk turns up more than you expected, image recognition is the capability that automates exactly that comparison — and the Operations Diagnostic puts a cost against the manual checking you are doing instead.