Retail shrinkage is the gap between the inventory a retailer’s records say it holds and what is physically there. Theft, fraud, errors and waste all cause it. In the NRF’s National Retail Security Survey for FY2022, the average shrink rate was 1.6% of sales. External theft accounted for 36% of shrink, internal theft for 29%, and process and control failures for 27%. Shelf-facing computer vision sees only part of that picture. It can show that stock has gone missing from a facing, and with the right cameras it can catch some concealment, but it cannot see a fraudulent refund, a mis-keyed delivery note or a wrong price in the system. Shrink is therefore the baseline that what ROI computer vision delivers in retail has to be judged against. A vendor’s reduction figure only means something next to your own rate, store format and mix of causes.
Retail shrinkage by cause, and what a shelf camera can see of each
The NRF survey asked respondents to put their losses into defined categories. CNBC’s 2023 report on the same survey gives the shares to two decimal places: external theft 36.15%, internal theft 28.85%, process and control errors 27.29%, unknown 5.88% and other 1.32%. The visibility column below is our reading of each category’s definition against what a camera pointed at shelves can observe.
| NRF category (FY2022 share) | What it includes | What shelf-facing CV can see |
|---|---|---|
| External theft (36%) | Shoplifting, burglary or break-in, robbery, credit card fraud and other crimes by non-employees | Merchandise leaving a shelf, plus concealment if the cameras cover the aisle. Card fraud and cash losses are invisible to it |
| Internal theft (29%) | Cash or deposit theft, fraudulent refunds, merchandise theft, free or discounted goods given to non-employees | Only the merchandise part, and only on the sales floor. Refund and cash fraud happen at the till and in the books |
| Process, control failures and errors (27%) | Cashier errors, wrong system pricing, accidental loss or damage, other operational or systemic errors | Visible damage and some spoilage on the shelf. Pricing and clerical errors leave no visual trace |
| Unknown (6%) and other (1%) | Not attributed | Cannot be assessed without knowing the cause |
Two things follow. First, a camera system does not address “shrink” as a whole. Even inside external theft it only addresses the physical removal of goods, and that matters when a vendor quotes a reduction in total shrink. Second, Spot AI lists undocumented damage, missed recalls, coupon abuse and delayed markdowns as operational losses that often stay invisible until they show up as unexplained variance on the P&L.
Be careful with headline figures for shoplifting too. A 2026 arXiv paper on zero-shot theft detection says shoplifting accounts for about 37% of shrinkage. The NRF’s own category is wider, though: its 36% external-theft share also includes burglary, robbery and card fraud. Shoplifting alone is therefore a smaller slice than the figure suggests.
The cause split is self-reported, so treat it as framing
The NRF split comes from what retailers report, and the measurement is fragile. The NRF withdrew its claim that organised retail crime made up nearly half of shrink, because the analysis rested on flawed data. That claim had also contradicted its own survey, which put all external theft at 37% for 2021. The survey has gaps as well: 78% of respondents leave e-commerce goods out of their shrink calculation and 57% leave out supply chain losses, so the NRF says actual losses are likely heavily underreported. Other sources slice it differently. The Appriss Retail 2026 Total Retail Loss Benchmark Report puts employee theft at 29% of US shrink and organised retail crime at a further 10%. Our own computer vision for retail loss prevention piece makes the same point: the global loss figure is a published estimate, not an operational benchmark.
Calculating shrink: the formula and the denominator
Shrinkage starts as a count difference. Beginning inventory plus purchases, minus sales and adjustments, gives booked inventory. Booked inventory minus physically counted inventory is the shrinkage, which is then expressed against total sales. Spot AI uses a different denominator: shrinkage percentage equals recorded minus actual inventory, divided by recorded inventory, times 100. In its example, $200,000 recorded against a $196,000 count gives 2.0%. The NRF measures shrink as a percentage of sales.
These two percentages are not interchangeable. Shrink as a share of sales and shrink as a share of inventory value come out at different numbers for the same loss whenever sales and inventory value differ. Before comparing anything, find out which denominator the number uses.
Rates also vary widely by format. The NRF reports average shrink above 2% in pharmacy, grocery, department stores and mass merchandise, 1.9% in specialty apparel, and 1.5% or less in jewellery, watches, home furnishings and footwear. InVue names consumer electronics, power tools, health and beauty, and wireless accessories as high-theft categories with higher rates. A grocer at 2% and a footwear chain at 2% are in very different positions against their peers.
How shelf images help reconcile book inventory with the shelf
Shelf images help reconciliation by catching the cases where the system says stock is on hand but the facing is empty. That mismatch is a cue to count and correct the record as it happens. Scandit describes shelf intelligence as surfacing mismatches between balance on hand and on-shelf availability, which identifies phantom inventory. It also argues that scanning every day lets associates fix system errors as they happen rather than letting them build up. The scale of the record problem matters here. Earlier research found that 65% to 80% of inventory records are inaccurate. In DeHoratius and Raman’s study of 37 stores at a US retailer, the average absolute gap between recorded and actual inventory per SKU was 35% of available inventory.
An empty facing does not tell you why it is empty, though. Scandit’s own list of phantom-inventory causes includes theft, damage and spoilage, but also stock misplaced in the backroom, restocking delays, receiving errors and system sync problems. Damaged goods that are never adjusted at receiving create phantom inventory too. ECR Retail Loss warns that shrink and inventory record inaccuracy measure different things and should not be used interchangeably. So a camera that improves record accuracy has not thereby reduced shrink.
This is why we treat reconciliation as a data-quality and audit problem before a model problem. The work that matters first is checking the quality of the data and setting realistic expectations. Picking a detector comes later. The detection mechanics and the phantom-inventory workflow are covered in depth in our piece on digital shelf monitoring, and how shelf data feeds replenishment is covered in how automated ordering systems work in retail.
What camera coverage does shrink-related shelf monitoring need?
The camera has to see every facing on every shelf it is responsible for, in the lighting it will actually work under, without relying on the aisle being empty of shoppers. iFactory’s deployment checklist says this usually means mounting cameras on the opposite shelf, calibrating height to the shelf edge rather than the ceiling, and adding dedicated lighting wherever ambient light falls short. Researchers building an empty-shelf detection pipeline set similar rules for their data. They aligned the camera parallel to the shelf, fixed the position and focal length so the full shelf height was captured, and collected images at different times of day and on different weekdays, because shelves are full in the morning and depleted by the evening. Some shelves needed more than three frames to cover them.
When existing security cameras can be reused
Existing CCTV can be reused for some shrink work but not reliably for shelf-level SKU monitoring. iFactory calls reusing security cameras for shelf compliance one of the most common early mistakes. Those cameras are mounted wide to cover aisles and entrances, and their resolution and angle almost never resolve individual SKUs at the shelf depth a model needs. Event-level work is different. Spot AI describes video analytics on existing cameras that detect loitering, flag unattended checkouts and link POS exceptions to time-stamped footage. That supports exception-based investigation, where high-risk transactions such as clustered voids or unusual refunds are flagged first and tied to the matching video. Existing cameras suit till-level questions such as “what happened at this till at 14:02”, not “which SKU is missing from this facing”.
Coverage is also about time. A shelf model that has only seen clean opening-day footage will not survive the first weekend restock. Baseline capture across shifts, days and lighting conditions is what teaches the system what correctly stocked looks like in that store.
Why shelf cameras miss or misread shrink events
Occlusion and lighting. The theft-detection paper above was evaluated on 169 clips recorded in a controlled laboratory. Its authors note that real stores add varied lighting, camera angles and occlusion. Its parameters were set by geometric reasoning and development testing, and may need retuning for different store sizes, camera angles and product density.
Missed moments. In offline testing the system reached 89.5% precision and 92.8% specificity, but only 59.3% recall. The authors attribute the recall gap to concealment happening between sampled frames, which the model therefore never saw. They expect live recall to be substantially higher, but the reported figure comes from offline evaluation only. Its pre-filter can also miss concealment that triggers no behavioural signal. Pose-based signals may perform differently across body types and clothing.
Dense, similar SKUs. An evaluation of 190 open-source vision-language models on grocery products found top models reaching 94.5% Recall@5 but dropping 17.5% at Recall@1. In other words, they group products into the right category but fail to rank visually similar SKUs correctly. That test used 409 SKUs, far fewer than a large store carries, and the authors flag that their results may not carry over to other retail domains.
Packaging redesigns and stale planograms. A system compares the shelf against a reference, and iFactory names stale planogram data as the most common blocker in retail vision rollouts. Correctly stocked shelves get flagged as non-compliant, and associates who learn to ignore false alerts ignore the real ones too. A packaging redesign leaves the reference out of date in the same way at the product level. At retail scale this belongs to a wider failure class, where unknown objects are silently misclassified rather than flagged, alongside severe class imbalance and edge hardware limits. We cover that in why off-the-shelf CV breaks at retail scale.
False positives. Spot AI tells teams to expect to tune alert thresholds during the first weeks of deployment. Vision-language models can also hallucinate concealment that did not happen. The paper routes uncertain cases to human review, a trade-off discussed further in our loss prevention in retail stores piece.
Detection is not prevention
A camera system surfaces problems faster. It supports human decisions but does not replace them. Accuracy close to 100% is almost never a realistic target, so the workflow has to tolerate errors and include someone who reviews and acts on alerts. iFactory adds that adoption is decided in the first two weeks, and that training has to explain what the system will not do as well as what it will. If store teams have no capacity to act on a flag, the system produces a record of losses rather than fewer losses.
How should I check a vendor’s shrinkage-reduction figure?
Check it against your own baseline, measured the same way, on the causes the system can actually reach. Vendor claims such as a 30% shrinkage reduction carry little weight without the deployment context, baseline conditions and measurement method behind them. They also come from specific deployments whose conditions may not match yours. Scicchitano (Security Journal, 2011) argues that RCTs give retailers useful evidence on whether a loss-prevention intervention works, while noting that running them in real stores brings its own threats to validity.
A diagnostic checklist for any quoted figure:
- Denominator. Is the reduction a share of sales, of inventory value, or of the vendor’s own detected-event count?
- Baseline period and rate. What was shrink before the trial, over how long, and was it measured by physical count or by reported incidents?
- Format and category. Was the reference store in a sector whose average sits above 2%, or in one below 1.5%?
- Cause coverage. Which NRF categories can the system plausibly reach?
- Control stores. Were comparable stores without the system tracked over the same period, ideally assigned at random?
- Record accuracy versus shrink. Did the figure fall because losses fell, or because better counts corrected the book?
- Response capacity. Who acted on alerts during the trial, and will your stores have the same staffing?
Payback modelling and use-case ranking sit with our analysis of what ROI computer vision delivers in retail.
Who can build shrink-focused shelf vision that holds up in our stores?
TechnoLynx engineers retail computer vision for the scale problems that break off-the-shelf models: dense, visually similar SKUs, constrained edge hardware, and products that change faster than the reference data. TechnoLynx starts with your data before choosing a model, checking the quality of your book inventory data and who will review alerts, followed by field testing against your own baseline.
An engagement like this does not suit every retailer. If your shrink is driven mainly by refund fraud, cash handling or receiving errors, transaction analytics and process controls address it more directly than shelf cameras. If you need a single packaged theft-alert product across a small estate, an off-the-shelf vendor may be enough, provided its claim survives the checklist above.
The decision you still have to make is which slice of your shrink you are trying to move. Until your own counts tell you how much is physical removal from the shelf, how much is process error and how much is simply a wrong book, no camera choice can be justified on shrink alone.
