Retail loses an estimated $1.75 trillion annually to out-of-stock events. But the number that makes that figure so damaging isn't the scale — it's the timing. Most out-of-stock losses accumulate silently, in the hours or days between when a shelf gap develops and when someone finally notices it.
That detection lag is the core problem. And it's one that store walks, however diligent, are structurally unable to close.
The out-of-stock problem in retail
Shelf gaps don't announce themselves. A product sells through on a Tuesday afternoon. By Wednesday morning, that section of shelf is bare — but no alert fires, no dashboard updates, and no ticket routes to a replenishment team. The gap just sits there, costing revenue with every customer who walks past and doesn't find what they came for.
Most out-of-stock events develop between store walks. In practice, that means many gaps go undetected for 6–8 hours or more — sometimes for multiple days if the walk schedule doesn't cover that aisle. By the time an associate or manager finds the empty shelf, the loss has already happened.
What makes this worse: customers who can't find a product don't usually ask for help. Research consistently shows that when shoppers encounter an out-of-stock, the most common responses are substitution (switching to a competitor product in the same store) or abandonment (leaving without purchasing at all). Only a small fraction of customers flag the issue to staff. The rest vote with their feet — quietly, and invisibly to the retailer.
For an individual store, out-of-stocks typically account for 4–8% of lost sales. Multiplied across a multi-location network and across a full year, that figure becomes a meaningful drag on revenue that most retailers have no systematic way to track.
How retailers currently detect out-of-stocks
Three approaches dominate the market today. Each has a different cost-accuracy-frequency tradeoff:
1. Manual store walks
The baseline approach: store associates or managers walk aisles on a set schedule — once a week at many chains, more frequently at high-volume locations. Store walks are low-cost and require no additional infrastructure, but they are fundamentally limited by frequency. A walk that happens on Thursday afternoon cannot catch the gap that opened on Tuesday morning. Most out-of-stock events in a typical store will never be seen by a scheduled walk before they self-resolve (through replenishment) or persist long enough to show up in inventory reconciliation.
2. IoT shelf sensors
Weight-based or RFID shelf sensors provide accurate, continuous stock level data. When a shelf section drops below a configured weight threshold, a replenishment alert fires automatically — no walk required, no visual inspection needed. The tradeoff is cost and setup overhead. Sensor hardware represents a significant per-shelf capital investment, and each SKU requires individual calibration. Deploying sensors across a full store estate — let alone a multi-location network — is a substantial undertaking. IoT sensors also cannot detect planogram deviations or promotional display execution failures; they measure weight or RFID signal, not visual presence or position.
3. Computer vision shelf monitoring
Computer vision uses existing store cameras to detect shelf conditions visually, continuously, and without additional per-shelf hardware. AI models analyze camera feeds in store aisles, identifying when sections fall below stock thresholds, when products are placed outside their planogram position, and when promotional displays are missing or incorrectly installed. Alerts route to staff in real time via mobile app.
The critical advantage over store walks is detection frequency — gaps are flagged as they develop, not days later. The advantage over IoT sensors is coverage — visual detection works with existing camera infrastructure and captures planogram compliance and display execution that sensors cannot.
How AI shelf monitoring works
In a computer vision shelf monitoring deployment, an AI Gateway device connects to existing CCTV cameras in store aisles and processes video locally. No footage is transmitted to the cloud; only structured intelligence — shelf status, alert events, compliance scores — flows to the operations dashboard and staff app.
The models look for specific visual signals of shelf conditions:
- Visible backing: When shelf backing or peg hooks become exposed, the model registers a stock gap in that section. The threshold is configurable by zone and product category.
- Missing facings: Products that should be frontfaced but are absent or pushed to the back register as potential out-of-stocks, even when some inventory technically remains.
- Planogram deviations: Products placed outside their designated position — wrong shelf, wrong section, wrong facing — are flagged for correction.
- Missing promotional displays: During campaign windows, the system verifies that displays are installed, correctly positioned, and maintained throughout the campaign period.
When a gap or deviation is detected, an alert routes to the responsible staff member via mobile app. The average detection-to-alert latency is measured in minutes, not hours. That changes the operational calculus entirely.
"A shelf gap found on Thursday's store walk was a gap that lost revenue since Tuesday. Real-time detection changes the calculus from mitigation to prevention."
Replenishment staff receive the alert, pull stock from the backroom, and close the gap — often within a single customer-facing hour rather than after a multi-day revenue loss.
What computer vision detects (and what it misses)
Computer vision shelf monitoring is not a complete substitute for all shelf intelligence. Understanding its scope — including its limitations — is important when evaluating platforms.
What it reliably catches:
- Visible shelf gaps: sections where backing or hooks are exposed
- Planogram deviations: products placed outside their designated position
- Missing promotional displays: absent or incorrectly installed campaign fixtures
- Promotional execution gaps: displays present but improperly assembled or incomplete
What it typically misses or requires additional integration to address:
- Wrong variant, right position: If a product is stocked but with the wrong color, size, or flavor variant, computer vision may not flag it — particularly for SKU categories where packaging looks similar across variants. Planogram integration with product image databases can improve detection here.
- Stock in backroom, not on shelf: Computer vision sees what's on the floor. Whether replenishment inventory exists in the backroom requires WMS or inventory system integration — the camera alone cannot know.
- Very small or narrow products at adverse angles: Accuracy degrades for certain product categories depending on camera angle, lighting conditions, and product profile. Pilot deployments typically include a calibration period to tune model confidence thresholds for specific SKU categories in each store format.
Evaluating shelf monitoring platforms
When assessing out-of-stock detection systems, the evaluation criteria that separate capable platforms from underwhelming ones are more specific than most buyers realize upfront.
| Criterion | What to verify |
|---|---|
| Detection latency | How quickly does a shelf gap trigger an alert? Minutes matter. Ask for documented latency benchmarks from live deployments — not lab conditions. |
| Camera compatibility | Does the platform work with your existing CCTV hardware and VMS? Confirm specific camera model compatibility before signing — don't assume. |
| Planogram integration | Can the system detect position deviations, not just stock presence? Does it integrate with your planogram data or require manual zone configuration? |
| Alert routing | Do alerts go to staff via a mobile app, or only to a dashboard that managers check intermittently? Staff app routing is what closes the gap fast. |
| Accuracy per product category | Ask for accuracy data segmented by product type — packaged goods, apparel, home, etc. A platform that's 95% accurate on grocery may perform differently in a fashion aisle. |
| Pilot support | Does the vendor include a calibration period in the pilot? Models need to be tuned to your specific store format, camera angles, and SKU categories. Vendors that skip this step tend to underperform. |
Before committing to a full rollout, run a structured pilot in 2–3 stores with different format profiles. Measure actual detection rates, alert response times, and replenishment cycle times over 4–8 weeks. Verify that the platform surfaces shelf intelligence your team acts on — not just data that goes unreviewed.
→ See EdgeRetail Shelf — Out-of-Stock Detection & Shelf Monitoring