Shelf-state detection can reliably flag a missing product; predicting theft from behavior cannot. These are distinct capability tiers with different methods, error rates, and deployment needs, and vendor claims should not blur the difference.
The Problem Vision Actually Solves
Retail shrink includes external theft, internal theft or fraud, administrative mistakes, and vendor losses. Computer vision mainly addresses the first two by making activity less anonymous, flagging events resembling theft patterns, and preserving footage that can support later investigation.
Vision detection covers behavior around shelves, not internal theft, vendor fraud, or stockroom shrink when cameras are missing or shelf views hide the pattern. Administrative losses, such as receiving errors, price overrides, void abuse, and inter-store transfer discrepancies, need POS and inventory links plus exception reports. Camera intelligence should therefore complement administrative audits.
Its usable scope remains significant. In busy stores, external theft may represent a substantial portion of shrink and is the category most responsive to visible detection. The business case therefore depends on how much of each store's shrink comes from external theft.
Reliable Uses Today
Shelf-state change detection is the most established commercial application. Showing that an item was present at one timestamp but absent later, while linking the change to no nearby scan, is a well-understood inference problem. With appropriate camera placement and stable lighting, current models reach commercially viable accuracy on actual store footage.
"Commercially viable accuracy" still has conditions. Good placement, steady light, and clear shelf views help keep false positives manageable. Changing light, aisle traffic, and cameras designed for general surveillance can materially reduce performance. Local calibration and zone tuning matter more than headline accuracy shown in vendor demonstrations.
Zone occupancy and dwell-time analysis is also dependable in aggregate. Finding unusual dwell patterns in a store area during selected time windows is practical. Using those signals to prioritize staff checks or locations for closer observation is a legitimate, low-risk use.
Still Aspirational
Individual behavior prediction receives heavy marketing attention but is the least dependable at store scale. Calling someone likely to steal before a theft, based on behavior signals, creates unacceptable false-positive risk. Legal and ethical rules for customer-facing use are also still developing across most jurisdictions.
We do not build or use behavioral prediction systems. Chains evaluating vendors should request the false-positive rate in their own stores, not a controlled test, and the response procedure for an innocent person who is flagged. A vendor unable to answer both precisely is not production-ready for this use.
Real-time facial recognition tied to outside databases presents a separate, more problematic aspiration. Accuracy varies with camera resolution, lighting, and viewing angle, while deployment rules differ by jurisdiction. For most regional chains, the technology is not yet mature for retail and is difficult to deploy compliantly.
Store Deployment Reality
Most regional chains have CCTV intended for passive recording, not analytics. These cameras enable review after an incident and usually cover busy areas broadly rather than specific shelves. Applying vision requires an honest assessment of what each camera can usefully show.
In our pilot-store work, roughly 60 to 70 percent of existing camera positions in a typical regional chain provide a usable shelf-state angle in at least one product zone. The remainder face angle, distance, or occlusion issues that make shelf inference unreliable. Storewide coverage may exist, but deployment still requires zone-by-zone review.
The clearest shrink reductions appear where high-shrink locations are known and concentrated: an end-cap with repeated theft, a display bay near a poorly monitored exit, or a self-checkout area with known scan avoidance. Suitable coverage and alerts there can be compared with historical shrink for the same areas.
Measuring Results
Loss-prevention investment is hard to measure because prevented theft is inherently invisible. The counterfactual cannot be seen. Stronger methods compare similar stores with and without active monitoring over a period that captures seasonal and trend effects. Many regional chains lack SKU-location baseline data.
A more practical approach for most chains is tracking alert volume by location, response times, and staff coverage, then comparing those measures with category-level shrink over time. Faster alert response alongside declining category shrink offers useful correlation without a perfect controlled experiment. It signals operations, but does not establish causation.
For a regional chain, computer vision is a real capability with specific applications that work reliably today. It does not solve every shrink problem. Results are strongest when teams target known high-shrink locations, define response procedures, and add vision to the administrative and process controls needed for other shrink categories.
The Road Ahead
Shelf-state detection and item-removal inference should improve as deployed systems create larger training datasets. Edge inference, with models running on cameras instead of making cloud round-trips, is already offered by some hardware vendors and will reduce latency. The capability curve is moving upward.
The more important development is linking loss-prevention signals with broader operational data. Combining shrink alerts with restock and compliance signals gives staff one operational view instead of separate tools. For chains operating 10 to 30 stores, that integration means alerts can be handled within the same store workflow.