Retail Cameras Finally Started Earning Their Keep

Shops have had cameras forever. Until recently they did exactly one job: record, so someone could review the footage after something went wrong.

What changed is that software can now interpret what a camera sees as it happens. And the interesting part of that story is not the technology — it is which uses turned out to make money, because it was not the ones the industry spent a decade chasing.

The expensive distraction

For years, "AI in retail" meant the store with no checkout. Walk in, pick things up, walk out, get charged automatically.

It genuinely worked. It also needed so many cameras and so much computing power per store that the cost was enormous, and it solved a problem — queueing at the till — that self-service kiosks had already largely solved for a tiny fraction of the price.

While that got all the coverage, cameras pointed at shelves were quietly paying for themselves in under a year.

What actually makes money

Spotting empty shelves. When a shelf facing is empty, that sale is usually gone — the customer substitutes or leaves. And here is the odd part: the shop's own inventory system often cannot tell. It knows what came in and what went through the tills, but it cannot distinguish twelve units sitting on the shelf from twelve sitting in the stockroom. A camera can.

Catching self-checkout losses. Items passed around the scanner, cheap barcodes scanned for expensive goods. What makes this work is not a brilliant camera — it is that the system compares what it saw against what the till recorded. Two separate signals have to disagree before anything is flagged, so even a moderately accurate system produces reliable alerts.

Predicting queues. Counting people waiting is easy and not very useful, because the manager can already see the queue. Predicting that one will form in eleven minutes based on door traffic and basket sizes is useful, because it gives someone time to open another till before customers get annoyed.

Nothing on that list is exciting. That is rather the point.

The economics people get wrong

This is worth understanding because it explains a lot of cancelled projects.

The cost splits three ways. Hardware in each store, which scales with the number of stores. The software platform, which costs roughly the same whether you run forty stores or eight hundred. And ongoing operations.

That middle one causes the trouble. Run a trial in fifty stores and you pay almost the entire platform cost while capturing a fiftieth of the benefit. The return looks terrible — and it was always going to look terrible, before anyone chose a camera.

Plenty of programmes that would have been strongly profitable across a full estate got cancelled on trial numbers that could never have been good. The fix is to work out the economics at full scale first, decide there, and treat the trial as a way to find out what breaks rather than as proof of value.

Where the processing happens matters enormously

The obvious approach is to send video to a data centre and process it there. It does not survive contact with scale.

One camera streams a couple of megabits per second continuously. Twenty cameras per store is forty. Two hundred stores is around eight gigabits per second, permanently, plus the computing power to analyse four thousand simultaneous video streams that never stop.

You would be paying to move an enormous quantity of pixels in order to extract a small number of useful facts.

So the working design puts a small computer inside each store. It watches the cameras locally and sends out only short text messages: bay 14 has a gap, six people waiting, transaction 88213 looks wrong. A few hundred bytes instead of continuous video.

That keeps costs flat as you add stores, keeps working when the internet connection drops, and — as a genuine side benefit — means customer footage never leaves the building. Most of the privacy question disappears simply because of where the computing happens.

What actually kills these projects

Almost never the technology.

An alert fires, a staff member walks over, and the shelf is full — a shadow confused the system. Four of those in a week and they stop trusting the alerts. Once that happens, improving the software does not help, because nobody is reading it.

The other quiet killer is assuming the work ends at launch. A system that is 94% accurate in the store where it was built typically drops to the low seventies by store forty, because lighting, shelf heights and product ranges all differ. Keeping it accurate means continuously sampling real footage, having people check the uncertain cases, and retraining — permanently. That costs something like a fifth of the build cost every year, and it is the line that always gets cut.

Cut it and you get a system that is excellent on launch day and ignored a year later.

One thing that genuinely improved

The newer AI models removed the need to train a separate custom model for every single task. You can now ask, in ordinary language, whether a photo of a shelf shows a gap.

They are too slow and expensive to run on every frame, so they sit behind a cheap fast detector that flags anything worth a closer look. But the effect is real: adding a new capability went from three months of work to a few days, because the expensive part needs no training data at all.

Full technical version here: Computer Vision in Retail: What Actually Ships in 2026

Frequently Asked Questions

Why can a shop's inventory system not detect an empty shelf?

It calculates stock as deliveries minus sales, which cannot distinguish between units sitting on the shelf and units sitting in the stockroom. Both appear identical in the system while being completely different for a customer standing in front of a gap.

Why is self-checkout loss detection so reliable?

Because it cross-checks camera observations against the till's transaction record rather than relying on the camera alone. Two independent sources must disagree before an alert is raised, which keeps false alarms low even with a moderately accurate model.

Why not process the video in a data centre?

The bandwidth and computing costs grow with cameras multiplied by stores and run continuously, quickly exceeding the value of the information extracted. Processing inside each store and sending only short summaries keeps costs flat as the estate grows.

Do these systems recognise individual people?

The valuable applications do not. They detect objects and activity — gaps, queues, checkout behaviour — without identifying anyone, which avoids the far stricter legal rules that apply to facial recognition and biometric data.

Why do trials often show poor returns?

Because the software platform costs roughly the same regardless of how many stores use it, so a small trial absorbs nearly the whole cost while capturing only a small share of the benefit. The poor result is a mathematical artefact, not evidence about the technology.

Why does accuracy fall when rolling out to more stores?

Lighting, shelf heights, camera angles and product ranges vary between locations, so a model trained in one store generalises poorly to others. Maintaining accuracy requires an ongoing cycle of sampling real footage, human review and retraining rather than a one-time fix.