Inventory control software tells you what the ledger believes the store holds. It does not tell you what is on the shelf. The software records receipts, sales, orders and stock movements, and the balance on hand it reports is usually what was delivered minus what was scanned out at the till. On-shelf availability (OSA) measures something else: the share of SKUs physically present and purchasable at the shelf. A product sitting in the back room, or listed in a database but missing from the facing, counts as out of stock. Shelf-execution AI is the layer that gives continuous, objective visibility of what is physically on the shelf.
The two numbers can drift far apart. ShelfOptix, which sells shelf-vision services, gives the example of a store with 98% inventory accuracy and 82% OSA at the same time. We read that as a vendor’s illustration, not a measurement.
Why the stock ledger only meets the shelf at a physical count
Stock quantities, receipts and issues, records across locations and replenishment decisions are what an inventory control system is built to handle. Automated systems trigger replenishment from the inventory levels recorded in the system, and those can differ from what is actually in the store. Kang and Gershwin (MIT, 2004) describe how most stores estimate on-hand stock from just two measurements: incoming shipments, and outgoing sales scanned at the point of sale.
In practice the system is an accounting model built on transactions. The model is right only as long as every unit that leaves passes a scanner. ERP and warehouse platforms track inventory movements, not shelf exposure, and they were not designed to verify shelf reality. Pazo frames the split as two different questions: inventory availability asks whether the retailer has the product, while OSA asks whether the shopper can find and buy it.
The ledger’s built-in link to physical truth is the count. In Microsoft Dynamics 365 Business Central, for example, staff count items and then post the actual quantities to the general ledger through journals. Kang and Gershwin note that the industry standard is a physical count at least once a year. Between counts, the record holds whatever the transactions say.
Why does inventory software say a SKU is in stock while the shelf is empty?
Inventory software shows stock on an empty shelf because units leave the store without a transaction, so the record keeps overstating what is there. Kang and Gershwin separate known stock loss, which staff record (expired goods written off, for example), from unknown stock loss such as undetected theft, which never reaches the record. Unknown loss only pushes the error one way: the record overestimates the stock.
The damage comes from what that does to reordering. The inventory record stays above the reorder point, so replenishment freezes while the shelf sits empty. In a regular out-of-stock the record shows zero and triggers a reorder. In a phantom out-of-stock the record shows units on hand, so no reorder fires. Kang and Gershwin’s simulation found that stock loss of just 1% of average demand disrupted replenishment enough to lose 17% of total demand to out-of-stocks, and at 2.4% loss more than half of demand was lost. In their model, lost sales ran ten to twenty times higher than the items actually stolen. They also found the effect is worst in lean systems with short lead times and small safety stocks.
A small, invisible leak in the ledger causes outsized lost sales, because it switches off the very replenishment logic the inventory system exists to run.
How often the records are wrong
Independent studies find inventory record errors common:
- Rekik, Oliva, Syntetos and Glock (2025) summarise earlier research as finding that between 20% and 70% of inventory records are incorrect.
- DeHoratius and Raman’s study of 37 stores of a US retailer found 65% of records inaccurate, with an average gap of five units per SKU, or 35% of the stock found on the shelf.
- Gruen and Corsten’s 2008 out-of-stock study for GMA, FMI and NACDS found perpetual inventory matched on-hand stock exactly in only 32% to 45% of cases, across four studies.
- In one retailer’s audit in the same study, items whose records did not match (phantom or hidden stock) had an out-of-stock rate of 8.9%, against 4.1% where records matched: about 2.2 times higher.
Scandit and ShelfOptix both state that phantom inventory causes 80% of out-of-stocks. Both are vendors selling shelf vision, so we would treat the figure as a vendor claim. Gruen and Corsten’s single-retailer comparison is narrower, but it is measured.
Not every empty shelf is a ledger error
Phantom stock is only one way the shelf and the ledger come apart. Gruen and Corsten’s 2002 work found that about a quarter of out-of-stocks were shelf out-of-stocks: the product was in the store but not on the shelf. Their audits also showed that in roughly half of mismatches the record showed less stock than was really there (“hidden” inventory), not more. Ton and Raman studied “phantom products” that sit in the store where customers cannot find them, even when the records are accurate. So an empty facing can mean a wrong ledger or a right ledger with the stock in the wrong place. The ledger cannot verify either case.
What the stock ledger records and what a shelf-vision layer sees
The boundary is the shelf edge. Inventory control software records stock levels, orders, sales and stock movements. Shelf-execution AI analyses shelf images to detect gaps, misplaced SKUs and planogram violations.
| Inventory control software | Shelf-execution AI layer | |
|---|---|---|
| Question it answers | Does the retailer have the product? | Can the shopper find and buy it? |
| Source of truth | Receipts minus POS sales | Images of the shelf, checked against planogram and records |
| Phantom stock | Cannot detect it | Flags system stock above zero with an empty facing, when correlated with ERP data |
| Stock in back room or wrong bay | Counts it as available | Sees the empty or wrong facing |
| Planogram drift | Not observed | Detects misplaced SKUs and planogram violations |
| Physical truth cadence | At counts, industry standard at least yearly | Limited by capture frequency |
| Main output | Stock records and replenishment support | Tasks for store teams |
| Main weakness | Record accuracy | Detection errors under dense, occluded or badly lit shelves |
One point the table can hide: the vision layer works integrated with the inventory system. A shelf image can say a facing is empty. Correlating repeated empty facings with ERP data is what highlights inventory inaccuracy. The useful signal is the mismatch between the two, which is why we treat the layer as a complement to the ledger. Rekik and colleagues reach a similar conclusion about visibility technology in general: it complements managerial oversight and targeted audits rather than replacing them.
Putting the shelf count and the stock-on-hand figure side by side
Measure on-shelf availability by capturing the physical shelf state, independently of ERP records. Fieldpie gives the standard formula: OSA % = (SKUs in stock on the shelf ÷ SKUs that should be on the shelf) × 100. In its example, a planogram calls for 200 active SKUs, an audit finds 182 present and shoppable, and the OSA rate is 182 ÷ 200 = 91%. Pazo notes that the formula is simple, but defining the expected SKU denominator is harder.
The comparison with stock-on-hand is a second number: the count of SKUs where the system shows stock above zero and the shelf shows none. Gruen and Corsten describe the manual version: take the out-of-stocks an audit found and check their perpetual inventory. If the record shows plenty on hand, inaccurate records are the likely root cause. A vision layer runs the same comparison from shelf captures, ideally daily.
Each measurement method sees a different slice of the problem. Gruen and Corsten found that audits, POS analysis and perpetual inventory data each measure different aspects of out-of-stocks and report different rates:
- Manual audits give accurate point-in-time data, but most retailers can only afford them weekly or fortnightly, which leaves blind spots between visits. They are costly to scale, though they work well when targeted at critical products.
- POS-proxy inference flags SKUs whose sales stop. Gruen and Corsten report that it identifies true out-of-stocks with 85-90% accuracy when validated against audits. It misses slow sellers, relies on historical patterns, and cannot see a phantom case where no transaction was expected anyway.
- Perpetual inventory data is generally below 50% accurate. It usually covers store out-of-stocks only, not shelf out-of-stocks.
- Computer vision, from fixed cameras, robots or associates’ mobile apps, detects gaps, misplaced SKUs and planogram violations from shelf images.
The sources disagree on how good POS-proxy is. ShelfOptix rates it at 50-70% accuracy and manual audits at 30-60%, well below Gruen and Corsten’s figures. We give more weight to Gruen and Corsten. Their figure was validated against manual audits, while ShelfOptix sells the vision alternative it rates highest.
Can a shelf-execution layer reuse the cameras and devices already in the store?
Sometimes; the capture format matters, because each format misses in different places. Shelf images can come from fixed cameras, shelf-scanning robots or mobile apps used by store associates. Some platforms use connected shelf-edge cameras. Scandit’s ShelfView combines image recognition and barcode scanning in a hybrid capture approach, with images processed in the cloud. A RoboSD field note on shelf-scanning accuracy points out that fixed cameras, roaming robots and scanners on existing floor scrubbers each miss in different places, so a pilot should test the format the store will actually run.
Whether existing cameras are enough is something the pilot should test, in the format the store will actually run.
Pretius defines actionable real time for shelf operations as shelf data fast enough to influence replenishment within the same shift. It reports that most shelf-visibility systems work in the 5-60 second range. The detection side of this is covered in depth in our piece on digital shelf monitoring with computer vision.
What happens after the shelf layer flags a ledger-vs-shelf divergence
A divergence flag pays off when it becomes a traceable corrective workflow. The steps run in this order:
- Detect and prioritise. The vision layer turns detections into prioritised tasks for store teams. ShelfOptix calls this a prescriptive worklist: a specific associate fixes a specific SKU in a specific location.
- Diagnose before touching the record. Pazo recommends tracing the gap through supply, inventory, back room, shelf execution and shopper availability, because an empty shelf is an observation, not a root cause. If the back room holds the stock, the fix is a restock task, not a ledger change.
- Recount the SKU. Cutoff errors cause false variances. A truck unloaded but not yet received makes stock look long, and an order picked but not shipped makes it look short. Counting cases against a system that holds single units, or missing a second storage location, also produces false differences.
- Post the adjustment. In Business Central, posting the counted quantity adjusts inventory by the difference between expected and actual stock. Item ledger and warehouse data must be synchronised before the count, or other adjustments get mixed into the result.
- Verify the correction. Pazo stresses that recording an audit as completed does not prove the shelf was fixed. Follow-up photo evidence does.
- Re-target attention. Rekik and colleagues found that audit sales gains concentrate on items whose records overstate stock, and are stronger for perishables. They argue for allocating audit effort by discrepancy risk rather than spreading it evenly across the store.
Rekik and colleagues estimate that a store-wide stockcount was associated with roughly an 11% sales increase over the following two months, with the benefit concentrated on SKUs whose records overstated stock. They are careful to say that the mechanism (better availability, less freezing) was not directly observed in their data. Pazo reports ECR research finding 4-8% sales increases from correcting inventory inaccuracies in controlled tests. Even counting only every six months improved performance dramatically in Kang and Gershwin’s simulation. Our inventory control example follows one out-of-stock through this loop. For how corrected shelf data then feeds reorders, see how an automated ordering system works with retail shelf CV.
Should we buy Captana or ShelfView, or build a vision layer on our own cameras?
Dedicated shelf-execution products combine capture, models and task alerts. One such platform pairs connected shelf-edge cameras with AI that identifies empty spaces, misplaced items and low stock, and triggers tasks for store teams. That platform says it integrates with POS, ERP and planogram systems through standardised APIs. Scandit says ShelfView pairs shelf data with IMS and POS data, also reads price tags, and claims 99.7% shelf insight accuracy. Scandit also states that ShelfView customers have reached OSA above 95% and sales uplifts of up to 2.5%. These are the vendor’s own claims.
Independent evidence on accuracy is more modest. The RoboSD field note finds vendor-published figures clustering at 95-99%, while peer-reviewed systems tested on other retailers’ shelves score in the mid-80s. It puts the gap down mostly to tuning, planograms and who chose the test set. A commercial system learns the client’s own assortment, so a client’s real number will land somewhere between the two, and only in-store measurement shows where. The one retailer-audited comparison it found, published by Pensa in 2022, had legacy manual gap scans scoring 35.6%, 52.3% and 57.1% in three stores, against 93.3%, 95.2% and 93.9% for the scan. The note also warns that accuracy is several numbers, not one: identification, recall on empty facings, and precision of alerts.
| What the choice depends on | Dedicated product (Captana, ShelfView) | CV layer on existing cameras and devices |
|---|---|---|
| Hardware | Vendor’s capture format, such as shelf-edge cameras or hybrid mobile capture | Whatever the store’s cameras and handhelds can resolve, so coverage needs checking first |
| Integration | Vendor APIs to POS, ERP and IMS | Built to complement the retailer’s existing inventory system |
| Accuracy evidence | Vendor figures, to be re-measured in a pilot | Must be measured from scratch in a pilot |
| Best fit | Chains wanting a packaged, managed service | Chains with usable camera coverage, or unusual categories and workflows |
Either way, the value case should rest on a repeatable, menial task: walking aisles for gaps and checking them against the record. That is the kind of problem where AI has realistic, measurable ROI, unlike tasks that need superhuman judgement. Before rollout, we would run the pilot RoboSD describes: scan, audit the same shelves right afterwards, score label match rate, out-of-stock recall and alert precision, include the hard categories on purpose, and repeat the same aisles across a week. False alarms deserve particular attention, because they are what make staff stop working the list. Our piece on why shelf-execution AI does not replace store staff covers why someone still has to act on every alert.
Who can build a shelf-vision layer that works with our inventory system and existing cameras?
TechnoLynx builds the computer-vision layer that sits beside an existing inventory control system, without replacing it, and checks the shelf against the ledger. The first step is a GPU Performance Audit. It establishes whether the store’s existing cameras and mobile devices can run shelf detection, so the project can avoid a hardware procurement cycle where it is not needed. From there, the engineering work is the detection models, tuned to the retailer’s shelves, and the integration that turns a shelf observation into a divergence against stock-on-hand. For camera-based vision outside retail, see our work on AI engineering for manufacturing and automotive.
This engagement, built on reusing existing devices, does not suit a chain wanting vendor hardware. Captana and ShelfView are reasonable routes for that, provided their claims are re-measured in a pilot. TechnoLynx does not sell inventory control software, and a vision layer will not fix out-of-stocks caused by supplier lead-time variability or forecasting errors. A retailer with an in-house computer-vision team and good camera coverage has the option to build the layer itself. In that case, the pilot methodology above is what to hold the build to.
The number to get before choosing a product
Before comparing Captana, ShelfView or a custom layer, establish how much of the store’s empty-shelf time comes from wrong records and how much from execution with correct records. The research shows both matter: in one retailer’s audit, record mismatches roughly doubled out-of-stock rates, and a meaningful share of out-of-stocks sit in the building but off the shelf. A week of observed facings on a few high-velocity and perishable categories, set against system stock-on-hand, gives that split. If the gap is mostly phantom stock, the payoff sits in ledger correction and unfreezing replenishment. If it is mostly back-room and placement failures, the payoff is in tasking, and the inventory system was never going to see that problem.