The Background
A new tier-1 city, 32 community fresh-food convenience stores, average ticket 25 to 40 yuan, targeting residents grabbing groceries on the way home.
Fresh food has one core pain: rot.
Leafy greens last two or three days. Unsold means tossed. This chain's shrinkage rate sat at 8% long-term — every 100 yuan of stock, 8 yuan went straight to the dumpster. The owner, Ms. Lin, did the math: every 1 percentage point of shrinkage cut across 32 stores adds nearly 500,000 yuan a year.
In early 2026 she spent 120,000 yuan on an AI selection plus freshness management system. The stack runs on Tongyi Qianwen and Tencent Cloud, customized by a local AI vendor.
Three Things
1. AI Product Selection: What Each Store Stocks, Data Decides
Selection used to be the regional manager walking stores and judging by gut. Now AI analyzes neighborhood profiles — age, household structure, delivery-app preferences — combined with sales history, generating a differentiated product list per store.
Same city: store A near an older neighborhood gets leafy greens and tofu. Store B near offices gets semi-prepared meals and quick-cook items. Store C near a school gets fruit and yogurt.
2. Freshness Management: Shelf Life Made Visible
Each batch gets scanned in with an auto-tagged shelf-life countdown. Near-expiry items auto-trigger discounts — one day left hits 50% off, half a day hits 70%, dynamically repriced.
Simple in concept, but previously staff eyeballed it manually. Busy hours meant nobody checked, and a batch of spinach would get discovered too late at night to discount.
3. Shrinkage Forecasting: How Much Tomorrow
AI predicts each store's daily sales by category, with procurement suggestions down to the kilogram. Store managers review and place orders; suppliers deliver next day.
Two Pits
Fresh food is not dry goods. The first two months of this system lost nearly 60,000 yuan.
Pit One: Dry-Goods Inventory Logic Applied to Fresh
The vendor's original model was built for dry-goods convenience stores, ported directly. The logic: first in first out, hold inventory if it doesn't sell.
Leafy greens don't hold. A 20kg batch of spinach sold 12kg in three days; the remaining 8kg rotted on day four. The system kept calculating safety stock on dry-goods logic, recommending restocks — more in, more rot.
The fix: a rewritten shelf-life model for fresh — different expiry weights per category. Leafy greens trigger discount the same day if unsold, not when they near expiry.
Pit Two: AI Selection Ignored Local Taste
Week two, the system pushed a load of northern vegetables to a southern-community store — napa cabbage, radish, large scallions. Southern customers don't buy those. A week of dead stock, all tossed.
Root cause: AI trained on city-wide sales data without splitting north-south taste. Northern veggies ranked high city-wide, but this store's customer base didn't want them.
The fix: regional taste tags per category, per-store preference calibration. Northern vegetables only pushed to stores with high northern-customer ratios.
The Results
Four months later:
- Shrinkage rate: 8% → 2.1%
- Stockout rate: 15% → 6%
- Gross margin: 18% → 26%
- Near-expiry discount losses: down 70%
- 120K yuan invested, payback in 4 months
The Real Lesson
Lin's takeaway: the core of fresh-food AI isn't prediction accuracy, it's throwing away less.
Being 10% more accurate and tossing 10% less — the latter is worth far more. Shrinkage in fresh food is real cost. A slightly-off forecast just means a few unsold bunches. A tossed crate of spinach is pure loss.
So the system's focus should sit on freshness management and dynamic discounting, not on obsessing over forecast precision. Stop throwing away first, then talk about selling more.
And: don't port dry-goods retail models onto fresh food. Two different businesses. Shelf life is a life-or-death line; dry goods don't have that problem.