AI Order Recommendations for Food Retail Stores

Store restocking used to rely on sales reps’ experience. We built a data platform on GCP with external variables such as weather, holidays and trade area, used Vertex AI to forecast the next 14 days of sales, and turned the forecast into recommended quantities for every product in every store—with 87.77% forecast accuracy.

Background & challenges

• The chain has stores across Taiwan, and restocking quantities relied largely on sales reps’ experience.
• Sales swing with weather, holidays, promotions and nearby competitors—static reports couldn’t keep up.
• Sales and store data sat in two separate on-premises databases, with no integrated dataset a model could use.

Our approach

Following Datas’ four-step method: get the data right first, then deliver the application, then keep it running.

1. Data audit: Organized sales, store tier, store format, shelf count and promotion data from 2021 onward, and defined external variables: Central Weather Administration weather, lunar-calendar and holiday dates, and distances to nearby hypermarkets and same-chain stores.
2. Build the platform: Connected on-premises MS SQL (9 tables) with daily Cloud Data Fusion pipelines, pulled external data with Cloud Run, and built an ODS → DW → ADS architecture in BigQuery.
3. AI applications: Vertex AI AutoML forecasts daily sales for each product in each store over the next 14 days; once reps pick a replenishment window, forecasts are compared with inventory and full-shelf levels to produce recommended quantities and low-stock alerts, synced to the sales app and Power BI.
4. Operate & optimize: Cloud Scheduler refreshes data and re-runs predictions; Power BI forecast-vs-actual and feature-attribution reports track accuracy; a technical manual and training completed the knowledge transfer.

Results

• 87.77% forecast accuracy (forecast within 1 case of actual sales), with an R² of 0.89 and a mean absolute error (MAE) of 0.18 cases.
• Daily sales for every product are forecast 14 days ahead; reps choose a replenishment window and see recommended quantities and stock alerts for every store.
• Forecast-vs-actual and feature-attribution reports make the model explainable: recent inventory and the previous day’s sales are the strongest drivers.
• The BigQuery data platform is a foundation for data governance and can extend to more products and channels.

Related services & next steps

• Predictive & Segmentation Models: the core of this project—Vertex AI store-sales forecasts turned into recommended order quantities.
• Data Platform & Data Engineering: a three-layer BigQuery platform combining internal data with weather, holiday and trade-area data.
• AI Agent Process Automation: can extend to automatic low-stock alerts for reps and a daily restocking list.