Using Edge AI to Stop Checkout Loss at the Point of Sale

Retailers are under constant pressure to make checkout faster and easier while reducing shrink, improving inventory precision, and controlling costs.

Self-checkout has made that balancing act harder. The same lanes that give shoppers speed and convenience also give them more opportunities, whether accidental or deliberate, for items to leave the store without being paid for. Non-scans and mis-scans, ticket switching, and items left in the cart or basket that never reach the scanner all add up to a significant and growing source of loss.

Retailers have traditionally responded with more staff intervention, more friction, or both. Neither is a great answer. But computer vision and edge AI can now verify what is happening at the checkout, catch errors in the moment, and give shoppers the chance to correct them before an associate ever needs to step in.

To explore how it works and how grocers and retailers are using these technologies to address checkout shrink, I recently sat down with Mitch Goldman of Edgify, an AI and computer vision company focused on loss prevention and product recognition for retail.

Catching Missed Scans, Ticket Switching, and Cart-Based Loss

Edgify's AI and computer vision technology runs directly on in-store devices, including Zebra solutions, integrating seamlessly with point-of-sale systems and scanner/scales. Using cameras already at the lane, it watches the transaction as it happens and checks that what is presented matches what is scanned.

That means the system can flag an item that passed through the scan zone without registering, an item whose barcode does not match the product it is attached to, or an item that stayed in the cart and was never presented at all. Because the analysis happens on the device in real time, the shopper can be prompted to resolve the issue themselves, which in most cases means simply scanning the item again or removing the wrong label. Interventions can be reserved for cases that genuinely warrant one.

As Mitch explained, this reframes loss prevention from something that happens after the fact, via receipts, audits, and exception reports, to something that happens during the transaction, without slowing honest shoppers down.

Making Non-Barcoded Items Easier to Identify

Verifying a transaction starts with knowing what is on the belt, and some of the simplest products create surprisingly complex challenges. Fresh produce, bakery items, and even hardware often have no traditional barcode, so customers or associates have to look them up in the point of sale, remember price look-up (PLU) codes, or manually identify what is being purchased. Those extra steps create friction, slow transactions, and introduce their own risk of errors and shrink.

The same camera and Edgify software that verify scans also automatically and accurately identify non-barcoded items at a self-checkout, staffed lane, or self-service scale, almost instantaneously. Customers spend less time working through self-checkout screens, associates no longer need to memorize large numbers of product codes, and because the system recognizes what is actually being presented, it reduces the incorrect selections and missed labels that affect both inventory precision and shrink.

Training AI in the Store Environment

One of the most interesting aspects of Edgify's approach is how its AI models are trained. Rather than relying on models developed externally by a third party, Edgify learns directly from activity in retailers’ own environments.

During an initial silent mode, the system observes transactions without changing the existing customer or associate workflow. When a shopper identifies an avocado, banana, or another non-barcoded item, Edgify's AI uses that real-world interaction as training data, pairing the transaction with the associated visuals to learn that item.

According to Mitch, Edgify can self-train a product recognition model over approximately three to five weeks and achieve greater than 95% accuracy across relevant non-barcoded items. This helps the technology adapt quickly to the realities of retail, where products, packaging, lighting, seasons, and regional assortments are constantly changing.

Mitch shared an example involving a major grocer whose Edgify system started encountering pumpkins when Halloween arrived. Once enough customers and associates had identified the new item, the model learned what a pumpkin looked like, and that knowledge was distributed across approximately 1,300 devices in about an hour and a half. Instead of becoming less effective as products change, the system keeps learning from the environment around it.

Why Running AI on Edge Devices Matters

Edgify's AI is built to run directly on devices at the edge, where retail transactions yhappen. This includes point-of-sale systems and self-checkouts as well as integration with barcode scanners, scales, or other devices that already have cameras and sufficient processing power.

Rather than continuously sending images and video to a cloud environment or investing in additional in-store servers, retailers can simply use the computing resources they already have in their technology environments. This offers several crucial advantages.

Processing data locally reduces the need to transfer large amounts of information across store networks, helping to control infrastructure and processing costs. And it allows retailers to use more of the available data without having to compress or restrict it for transmission.

This matters even more for loss prevention than for product recognition. Recognizing a product is a single-image problem. Verifying a transaction is a video problem: the system has to follow items continuously from cart to scanner to bagging area, which demands far more bandwidth and compute than a snapshot. Doing that in the cloud for every lane in every store is expensive and introduces latency at exactly the moment a shopper needs a prompt. Doing it on the device, on hardware already at the lane, is what makes real-time loss prevention practical at scale.

This approach also simplifies deployment because Edgify operates as a software layer across multiple operating systems and hardware environments, allowing retailers and their technology partners to take advantage of existing infrastructure rather than redesigning their entire store technology stack.

Ultimately, AI and computer vision make everyday retail processes faster, easier, and more accurate while improving the efficiency and economics of deploying the technology.

In our conversation, Mitch and I went deeper on how retailers are using these capabilities to address self-checkout shrink, how shopper self-correction changes the role of the attendant, and how to balance loss prevention with customer experience rather than trading one for the other.

To hear the full discussion and learn more about how Edgify and Zebra are applying edge AI to real-world retail challenges, watch or listen to the full podcast below.