AI Loss Prevention: What the Cameras Can and Cannot Do

The demonstration is always good. A camera feed, a box drawn around a person putting something into a bag, an alert on a phone. The salesperson says it runs on the cameras you already own and costs less than one shift a month, and the shop owner, who has watched three hundred pounds of stock disappear this quarter, says yes.

Six months later the shop has a phone that buzzes eleven times a day, staff who have stopped looking at it, and roughly the same amount of missing stock. This piece is about why that happens and what the same money buys instead.

How bad is the problem really?

Bad, and measured in ways that make comparison harder than it should be. Two numbers from the same organisation illustrate the trap.

The British Retail Consortium's 2025 survey reported over 20 million incidents of theft, around 55,000 a day, with total retail crime cost including prevention at 4.2 billion pounds, of which 2.2 billion was direct customer theft. Retailers spent a record 1.8 billion pounds on prevention in a single year.

The following year, the 2026 report described 5.5 million detected incidents costing close to 400 million pounds, alongside more than 100 million in parcel theft and 1,600 daily incidents of violence and abuse against staff, down from around 2,000 the year before.

Those figures are not in conflict and they are not comparable. One counts estimated incidents including those nobody caught; the other counts detected ones. A vendor deck will quote whichever is larger and attach it to a product that addresses neither. Knowing which number you are being shown is the first piece of defence.

Across the Atlantic the direction has shifted. The National Retail Federation's study published on 30 July 2026, covering 66 companies and 143 brands representing 1.7 trillion dollars of sales, found shoplifting incidents down 12.4 percent against the previous year and merchandise theft down 8.1 percent. What rose was fraud: phone scams reported by 69 percent of respondents, loyalty fraud by 51 percent and gift card fraud by 42 percent. The threat moved from the shelf to the account.

Where does missing stock actually go?

Not mostly out of the door in a coat. In a small shop the largest share of the gap between what the system says you hold and what is on the shelf is usually clerical, and this is the single most useful thing an owner can internalise before buying anything.

Five sources, in rough order of how often they bite a shop under ten employees. Counting and keying error, where a delivery of twelve was entered as a case of twenty four. Pricing error, where an item sells at the wrong price for three weeks and the value vanishes without the unit doing so. Supplier shortfall, where the invoice says forty and the box held thirty seven. Damage and waste, particularly anything with a date on it. And then theft, internal and external.

A camera addresses the last category only, and within it only the visible part. If your actual mix is sixty percent error, you have bought a solution to forty percent of the problem, and you will conclude that the product does not work when the truth is that it was aimed at the wrong thing.

The diagnostic is cheap. Take your twenty highest value lines, count them physically on a Monday, compare to the system, and repeat weekly for a month. The pattern in the variances tells you which of the five you have. Nobody sells this because there is nothing to sell.

What is the software actually detecting?

This is where the vocabulary matters, because two very different products are marketed with the same word.

Behaviour detection watches for patterns in the video: an item leaving the shelf and not reaching the till, a person lingering by a display, a self checkout where the scanned count does not match the visual count. It does not know who anyone is. The output is an alert about an event.

Identity matching compares faces against a list. The output is an alert about a person. That is biometric processing, it is a special category of personal data under European data protection law, and it is a different legal universe from the first. We wrote the long version of that boundary in our piece on the two regulatory regimes that govern facial recognition in shops, and the summary for present purposes is that the pitch which sounds simplest carries the most risk.

There is a third category worth naming because it is being sold and it is prohibited outright in the European Union. Systems that claim to infer a person's emotional state, sold as detecting suspicious intent or as monitoring staff attentiveness, run into Article 5 of the AI Act, which prohibits using AI to infer emotions of a natural person in the workplace except for medical or safety reasons. If a vendor's deck mentions reading intent or measuring engagement from a camera pointed at your counter, that is not an advanced feature. It is a compliance problem you would be buying.

Why do the alerts stop working?

Because of arithmetic that is identical to the problem in payment fraud screening, and that shop owners meet for the first time here.

Suppose the system is right 95 percent of the time, which would be a good system. A shop with four hundred customers a day and a genuine theft rate of one in two hundred visits has two real events a day. At 95 percent specificity the same system also flags around twenty innocent shoppers. So nineteen or twenty of every twenty two alerts are wrong, and that ratio holds no matter how good the marketing claim sounds, because it is driven by how rare the event is rather than by how clever the model is.

What follows is predictable and human. Staff act on the first few, have two uncomfortable conversations with innocent customers, and quietly stop looking. The system has not failed technically. It has failed operationally, which is the only kind of failure that matters. The identical dynamic, with identical maths, is why card fraud screening turns away good customers, and the fix in both cases is the same: raise the threshold until the alerts are rare enough to act on, and accept that you will miss things.

The bottom row is the point

Reconciling what the supplier invoiced against what physically arrived is unglamorous, nobody demonstrates it at a trade show, and it is where a small shop recovers the most money per hour spent. It is also a task that genuinely suits a model, because it is document comparison rather than judgement: read the delivery note, read the invoice, read the received count, list the discrepancies.

That framing matters. The useful AI in a shop is mostly reading paperwork, not watching people. A tool that photographs a delivery note and reconciles it against the purchase order is doing the same class of work as any document extraction pipeline, and it fails in understandable ways that a person can check. A tool that decides a customer looks like a thief fails in ways nobody can check.

Shelf gap detection deserves a mention for the same reason. A camera that notices an empty facing is not a security product at all, it is a sales product, and in most small shops the revenue lost to something being out on the shelf while sitting in the stockroom exceeds the revenue lost to theft. The technology is identical. The framing decides whether you get value from it.

What about the staff?

Two things, and the first is the one that gets forgotten in a conversation about stock.

The BRC figures put violence and abuse against retail workers at 1,600 incidents a day, including 118 involving physical violence and 36 involving a weapon. Any technology decision that increases the number of confrontations your staff have with customers has a cost that does not appear on the invoice. An alerting system that produces twenty questionable flags a day is, in practice, a machine for generating confrontations, and a sole trader or a part time assistant is the person who has to have them.

The second is monitoring of your own team. Internal loss is real and it is uncomfortable, and the temptation to point the analytics inward is strong. Beyond the emotion recognition prohibition already mentioned, monitoring employees carries its own obligations in most jurisdictions: they have to know, the purpose has to be specific, and the processing has to be proportionate to a problem you can actually evidence. A camera installed because a shop is nervous is different from one installed after a documented pattern, and the difference is the part that gets tested if it is ever challenged.

Where the counter and the website meet

One practical note for shops that sell in both places, which is most of them now. A large slice of apparent shrink in a hybrid business is not shrink at all. It is stock sold online and picked from the shop floor without the till knowing, or a click and collect order reserved twice.

Reconciling that is a systems question rather than a security question, and it is solved by having one stock record rather than two that are synchronised optimistically. If you are rebuilding or setting up the online side, it is worth treating the shop floor as the same inventory rather than a separate one, which is why our ecommerce website builder starts from a single stock model rather than bolting one on.

The diagnostic here is simple too. If your variances cluster on the lines you sell online, the problem is the integration, not the door.

Should you report it, and does that change what to buy?

The NRF figures contain an awkward detail that bears directly on the purchase decision. Sixty three percent of the retailers surveyed report fewer than half of their incidents to police, with 60 percent citing losses too small to meet a felony threshold and 54 percent citing doubts about the response. Those are large companies with loss prevention departments. A single shop is not going to do better.

That matters because the business case for a detection system usually assumes a consequence at the end of it. If the realistic outcome of an alert is that somebody leaves with the item and nothing else happens, the system is not preventing loss, it is documenting it. Documentation has some value, mostly for insurance and for building a pattern over time, but it is a different and much smaller value than the one in the sales pitch.

There is a countervailing signal in the same territory. The BRC's 2026 report notes police response rated good or excellent by 13 percent of retailers, up from 9 percent, and the Pegasus Partnership identifying 395 offenders with 9 million pounds of losses attributed to them since its 2024 launch. Organised repeat offending is the part of the problem where reporting does produce outcomes, because the case is built across many shops rather than on one incident.

The practical reading for a small shop is to separate the two. For a one off opportunist, the realistic goal is deterrence and not detection, and deterrence comes from layout and greeting rather than from software. For a repeat pattern, the goal is a record good enough to contribute to somebody else's case, which means clear footage and a written log with dates, not an alerting system.

A sensible order of spending

If the budget is a few hundred a year, spend it in this order and stop when it runs out.

First, the weekly count of your top twenty lines by value. Free, an hour a week, and it tells you which of the five loss sources you actually have. Second, delivery reconciliation, which pays back immediately in most shops and can genuinely be automated. Third, fixing the till: wrong prices, unlisted lines, staff overrides that nobody reviews.

Fourth, and only now, the physical basics that have decades of evidence behind them: sightlines, where high value stock sits, whether somebody greets people at the door. Fifth, cameras for evidence rather than for alerts, which is a different specification and usually cheaper.

Behaviour detection sits sixth, and it is a reasonable purchase for a shop that has done the first five and still has a measurable theft problem concentrated in a known aisle. Identity matching does not appear on the list at all for a business of this size, because the legal exposure and the cost of being wrong about a regular customer both exceed the recoverable loss.

The camera answers a question about the last five percent. Most shops have not yet asked the question about the first sixty.

What to watch

The NRF data point about fraud overtaking shoplifting is the signal worth tracking, because it says where the next few years of small business loss will come from. Phone scams reported by 69 percent of respondents and gift card fraud by 42 percent are not problems a shop camera addresses, and they arrive by telephone and email at a business with no fraud team.

For a small merchant that reframes the whole question. The loss prevention budget of 2020 went on the shop floor. The loss prevention thinking of the next few years belongs mostly at the till, in the inbox and on the phone, which is a less satisfying conclusion than a camera and a considerably cheaper one.