A merchandising giant makes an AI bet at the shelf
On July 21, Anderson Merchandisers named Seekr as its enterprise AI partner, a deal that puts explainable AI and vision-language models into the hands of more than 4,000 associates who service over 40,000 US retail locations. The stated aim is to attack the "last 100 feet" of retail, the stretch between the stockroom and the shelf where planograms meet reality. Every store visit becomes a source of structured data on shelf conditions, out-of-stocks, and compliance.
This lands at a moment when computer vision in retail has moved past demos and into procurement decisions. Anderson is a merchandising services firm whose product is execution in physical stores, so an AI partnership here is a direct wager on frontline productivity. The question every CIO in retail is now asking is whether vision models are accurate and cheap enough to run at the shelf, and this deployment is one of the larger real-world tests of that thesis to date.
The problem is worth trillions, and it lives in the store
The numbers Seekr and Anderson are chasing are enormous. Out-of-stocks cost retail roughly $1.2 trillion annually, and inventory distortion, the combination of overstocks and understocks, runs to about $1.73 trillion. These are not rounding errors on a P&L, they are structural leakage that has resisted a generation of supply-chain software. The prize is any tool that can reliably close even a fraction of that gap.
The most important statistic in the announcement is where the failure occurs. Seventy-two percent of out-of-stocks trace to in-store execution rather than the supply chain. That reframes the entire problem. The pallet arrived, the product is in the building, and it still is not on the shelf where a shopper can buy it. No amount of upstream forecasting fixes a facing that never got restocked. This is exactly the failure mode a human associate armed with a vision model at the shelf is positioned to catch.
Why vision-language models change the shelf equation
Classic shelf-monitoring computer vision could count facings and flag gaps, but it struggled to reason about what it saw. Vision-language models combine image understanding with language, so an associate can ask a question about a shelf and get an answer grounded in the picture in front of them. "VLM-powered applications are the key to unlocking real AI impact at the shelf level," says Rob Clark, President of Seekr. The claim is that reasoning over shelf imagery, not just detection, is what makes the technology useful to a working associate.
Seekr is emphasizing explainability, which is the detail that separates a pilot from a production rollout. A model that flags a compliance issue has to show why, in terms a field rep and a brand auditor both trust. For a services firm billing clients on execution quality, an opaque score is a liability. Explainable output turns the model into evidence, and evidence is what Anderson can put in front of the consumer brands paying for shelf presence.
The edge deployment is the real architecture decision
Seekr is deploying at the edge so associates get answers on the shelf floor. That is a deliberate architectural choice with real consequences. Retail store connectivity is uneven, aisles are radio dead zones, and a workflow that stalls waiting on a cloud round trip will be abandoned by the associate on their third store of the day. Running inference on the device keeps the interaction fast enough to fit inside an actual store visit rather than interrupting it.
For technology leaders, this is the part worth studying. Edge deployment of vision-language models means managing model updates, on-device performance, and data sync across thousands of endpoints in the field. The payoff is latency low enough to change behavior at the point of work. The cost is an operations problem that looks more like fleet management than a SaaS subscription. Anderson and Seekr are betting the store-execution ROI justifies that added complexity, and the size of the out-of-stock prize is why that math can work.
The people strategy is not an afterthought
Anderson is framing this as augmentation of a skilled workforce, not a replacement for it. "Our associates are the best in the industry, and they deserve technology that matches their expertise," says Jeff King, President of Anderson Merchandisers. The positioning matters, because the fastest way to kill a frontline AI rollout is to make associates feel surveilled or deskilled. Tools that make an expert faster get adopted, tools that second-guess them get worked around.
There is substance behind the framing. The 72% figure means the associate standing at the shelf is the control point for the most expensive failure in retail. Giving that person a VLM that answers questions and captures compliance in seconds raises the value of the visit itself. Done well, the technology turns a store call from a checklist into a structured data event, and it does so without pulling the associate off the floor to file reports later.
What this signals for retail technology roadmaps
This partnership is a marker that vision-language models have crossed into frontline retail operations at scale. A deployment to 4,000 associates across 40,000 locations is a production commitment with training, support, and client billing attached. For retail and CPG leaders, it sets a benchmark: if a services firm can instrument the last 100 feet with explainable edge AI, the pressure to close the in-store execution gap moves from aspiration to expectation.
The decision this puts in front of technology leaders is where to place computer-vision investment. The trillion-dollar out-of-stock problem sits in execution, and 72% of it is addressable at the shelf, so the ROI case for VLMs at the edge is stronger than the case for another layer of upstream forecasting. The hard part is operational: managing models in the field, proving explainable results to brand partners, and keeping associates in the loop. The firms that solve those three problems will convert store execution from a persistent leak into a measurable, defensible advantage.
