The Retail Loss Prevention Math - Where AI Cameras Actually Change the Equation

Retail shrink is the gap between the inventory a store should have and what is on the shelves. It cost US retailers roughly $112 billion in 2022, the last year the National Retail Federation published an industry-wide figure. That number is the backdrop for a fast-growing pitch: use an AI agent to keep watching your camera feeds, and watch shrink go down.

Some categories of shrink respond to AI cameras while others are invisible to them. A few get worse when a camera is added. For a small retailer deciding whether the system pays for itself, the question is not whether AI works, but which kind of loss they have to begin with. The rest of this post is that math, by category.

Where the money actually goes

The last reliable NRF breakdown, from the 2022 survey, put US retail shrink at 1.6% of sales - roughly $112 billion. The composition matters more than the headline:

• External theft (shoplifting and organized retail crime): ~36%
• Employee theft: ~29%
• Process and control errors (paperwork, damaged stock, miscounts): ~27%
• Other and unknown: ~8%

A camera, AI or otherwise, watches the sales floor and the stockroom. It does not reconcile a receiving log against a purchase order. It does not catch a returns-desk fraud where the merchandise never enters the store. So before you spend a dollar on cameras, the structural ceiling is roughly two-thirds of your shrink - the part that involves a person physically taking something. The remaining third is an inventory-management and policy problem, and no AI is going to fix it.

Where AI cameras can move the number

Within that two-thirds, AI vision is genuinely useful in a few places.

Self-checkout monitoring. Untrained shoppers and bad actors at self-checkout produce a steady stream of mis-scans; items in the bagging area that were never read, label swaps, basket items that walk past unscanned. An AI model trained to flag "scanned vs. observed" mismatches can prompt a human to glance at the screen. This is one of the genuine wins.

Stockroom and back-of-house cameras. A non-trivial share of employee theft happens away from the sales floor. Cameras here are not new, but the AI-enabled change is searchability - instead of a Loss Prevention (LP) team scrubbing 200 hours of tape after a missing-inventory report, they ask: "Who entered the stockroom between 9 and 11 pm last Tuesday?" That used to take a day. Now it can take a sentence.

Pre-violence escalation. The NRF's 2025 Impact of Theft and Violence report found a 17% year-over-year rise in threats and violence during theft incidents. The most defensible use of AI vision is not catching shoplifters - it is alerting staff before a confrontation happens, so they can step back instead of intervene.

After-the-fact lookup. Most cameras get watched almost never. Industry estimates suggest less than 1% of surveillance footage is monitored live, and only a fraction of recordings are reviewed afterward. Making that footage queryable in plain English changes the economics of investigations more than it changes the rate of theft.

Where they don't help, and where they backfire

The vendor pitch usually skips the parts where AI cameras don't work, or actively make things worse.

Process error and returns fraud are invisible to vision. A pallet that was logged as 48 units but actually contained 46 will show no anomaly to the smartest AI in the world. Returns-counter fraud - bringing back merchandise that was never bought, or swapping receipts - is a paper crime. Cameras catch the act but rarely the pattern, because the pattern lives in the POS data.

False positives at scale create legal exposure. In December 2023, the FTC banned Rite Aid from using facial recognition for security purposes for five years. The reasoning is worth reading directly: the system generated thousands of false positive alerts, including more than 1,000 alerts from a single uploaded photo. Alerts in stores located in plurality-Black and plurality-Asian areas were significantly more likely to be low-confidence matches than alerts in plurality-White stores. Employees followed and accused customers based on those alerts. The result was a federal enforcement action, a five-year ban, and a stack of civil suits.

Alert fatigue kills systems within weeks. Walmart deployed Everseen in thousands of stores starting in 2017. By 2020, employees had nicknamed it "NeverSeen" and shot demonstration videos for Wired showing it missing obvious theft while flagging innocent shoppers. If the operator turns the system off - or worse, learns to ignore it - the camera ROI drops to zero regardless of the AI's stated accuracy.

Vendor accuracy claims are usually best-case. A peer-reviewed 2025 deep learning study reported 95% overall accuracy on shoplifting detection. That is in a research dataset, not a Tuesday afternoon at a busy convenience store. Independent benchmarks of in-store performance are scarce, and the systems most often cited for high accuracy are evaluated by the vendor selling them.

The math for a small store

Take a four-camera convenience store doing $1.5M in annual revenue with a typical 1.6% shrink rate - that's $24,000 in losses per year. Suppose the breakdown holds: about $8,600 from shoplifting, $7,000 from employee theft, $6,500 from process error.

A business-grade AI camera deployment is not cheap. A reseller-published 3-camera Verkada small-business setup runs $4,494 in year one, with annual license fees thereafter of $199 to $1,799 per camera depending on tier. Amortized over five years, a four-camera shop lands somewhere between $2,000 and $4,000 a year all-in. Call it $3,000. That is already 12% of total annual shrink, before the system catches a single incident. To break even, the system has to reduce just the addressable two-thirds of shrink by roughly 18%. That is a more reachable bar than at large chains, but it is the bar that has to clear, not the one a vendor's case study claims.

This is also why most "AI camera" pitches lead with major retailers as case studies. A national chain with thousands of stores can absorb high system costs against a much larger absolute shrink number. A four-camera shop runs the numbers differently. For SMBs, the system has to be cheap enough that even modest reductions clear the cost - or the math just doesn't work.

The math has to clear, or the camera comes off the wall.

Shobdo VideoRAG is an AI agent for surveillance cameras. Shobdo can send alerts on events you describe using plain English. Learn more or book a conversation.