Vinasoy cut out-of-stock rates by 20% within two months of deploying GenAI retail monitoring across its Vietnamese distribution network.

Inside Vinasoy’s GenAI Retail Monitoring

Vinasoy is a Vietnamese soy milk manufacturer selling through a retail network that spans 34 provinces. It now runs GenAI retail monitoring over the photographs its staff take during store visits. First, image recognition running on Amazon SageMaker identifies the products on each shelf. Then a scoring engine built on Amazon Bedrock judges the display against Vinasoy’s approved standards and returns a compliance score to the sales team. The company built the system with Renova Cloud, an AWS Premier Tier Services Partner. The recognition stage handles 22 images per second, which Amazon Web Services describes as 1,300 times faster than manual inspection.

Twenty Days to Learn a Shelf Was Empty

Previously, staff assessed the same photographs by hand inside the company’s distribution management system. Each scoring cycle consumed nearly 2,000 work hours and took more than 20 days to finish. As a result, it reached only about 17% of outlets in a month, and any single store was checked once. Meanwhile, the shelves kept moving. “Our sales teams used to wait weeks for display compliance data, and by the time they got it, the shelf problem had already cost us,” said Le Ba Be, national sales director at Vinasoy.

What Changed on the Shelf

Since GenAI retail monitoring went live, out-of-stock rates have fallen 20% within two months. Coverage rose more than four times, to over 70% of distribution outlets per month, and each store is now checked weekly instead of monthly. Compliance scores reach the sales team within one to two days, up to 20 times faster than before, because the full scoring cycle now runs in under two days rather than more than 20. Preparing a sales report takes roughly five minutes, down from about three days. Vinasoy plans to extend coverage to all of its outlets.

“This has freed our people to do what they do best: build relationships with retailers and keep our products within reach of consumers,” Le Ba Be said. Other operators have pushed GenAI across similarly dispersed networks; Adecco, for instance, rolled out GenAI recruiting agents in 40 countries.

Why It Matters

  • The case for GenAI retail monitoring rests on decision latency rather than labour cost. Data that lands after the selling window has closed is worth little at any price.
  • A clear readiness signal is a manual inspection process that samples only a small share of sites because full coverage costs too much.
  • The model is judging, not seeing. Image recognition names the products; the generative layer applies the merchandising rules. Therefore a company without written display standards has nothing for a model to score against.
  • Field teams act on these scores only if they trust them. Expect to spend the first months arbitrating disputed scores rather than celebrating coverage.