Daily fresh-food ordering with an AI second opinion

The problem

Store managers eyeball ~200 SKUs daily from what's on the shelf. On flyer-promo days 40% of orders miss demand. When markdowns hit 50% the margin collapses to zero. A regional manager has 12 stores and no time to coach any of them individually. Weather and local events — Formula-1 weekend traffic, school holidays — never enter the ordering decision.

What we'd ship

An assisted ordering copilot blending POS data from the existing Azure Synapse DWH with Helios stock, promo calendars, local weather and event data. It proposes the day's order per SKU, shows the top three reasoning factors visibly, and lets the manager override in one tap. Built as an EU AI Act limited-risk recommender with a supervised fallback to current behaviour.

Stack

• Azure Machine Learning for per-SKU demand forecasting
• Azure Synapse (existing DWH) as the feature store
• ERP Helios integration via the existing pipeline
• Azure Maps + weather API for local context signals
• Blazor mobile PWA for the store-manager app (house stack)

Guardrails & compliance

• EU AI Act limited-risk recommender — every suggestion shows top-3 factors
• One-tap override always available; no forced acceptance
• A/B test 2× against control behaviour in the pilot stores
• ESG reporting pipeline aligned with the 2027 waste target
• Loyalty-program data GDPR-scoped — no personal data in forecasts

Typical timeline

1. Week 0–4 Baseline measurement + data prep

   Instrument waste and out-of-stock metrics per store, extract the feature set from Synapse, define the A/B methodology.

2. Week 4–10 Four-store pilot

   Train per-store models, ship the manager PWA, run control vs. assisted side-by-side for six weeks.

3. Week 10–22 Rollout to all 62 stores

   Phased enablement by region, training materials for store managers, hand the ESG reporting pipeline to the merchant director.