Intelligent & Smart Store Orders for a Global Beverage Giant

A global beverage manufacturer serving thousands of neighborhood stores in Latin America needed to reduce stockouts caused by missed or late orders. Store staff relied on manual shelf checks and manual entry in an ordering app, leading to empty shelves, lost sales, and lower customer satisfaction.

We integrated smart ordering agents into the existing store-order mobile app. Store keepers snap a photo of low or empty shelves; the system detects missing SKUs, pack sizes, and required quantities using in-house vision models and multimodal language models, then proposes an order that factors inventory availability, current prices, and active promotions. Users can adjust and submit immediately.

Store teams faced recurring stockouts due to manual workflows and SKU complexity:

Manual shelf checks: Staff frequently forgot to re-order or under-ordered during busy periods, creating gaps in high-velocity SKUs and popular pack sizes.

SKU, size, and promo complexity: Similar packaging across sizes and flavors, plus frequent promotions, made accurate ordering error-prone.

Disconnected signals: The ordering app lacked real-time context on DC/store inventory, price changes, and promos, delaying optimal recommendations.

Operational constraints: Variable connectivity and diverse device capabilities across LATAM demanded lightweight, resilient mobile experiences.

A vision-enabled, agentic ordering flow embedded in the existing mobile app:

Shelf Capture & Understanding: Users take shelf photos; in-house vision models and MLMs identify products, pack sizes, facings, and out-of-stock cues. Multimodal LLMs reconcile ambiguous labels and packaging variants.

Auto-Quantity Recommendations: The agent estimates recommended quantities based on shelf gaps, velocity tiers, historical orders, and planogram targets (where available).

Availability, Price & Promo Checks: Real-time lookups against distributor/DC inventory, current price lists, and active promotions refine the suggested cart to maximize fill rates and promo compliance.

Editable Smart Cart: Store keepers review, tweak quantities, substitute sizes/flavors if needed, and submit orders instantly to the manufacturer/distributor.

Learning Loop & Guardrails: Post-order outcomes and user edits retrain heuristics; policy guardrails prevent over-ordering and flag anomalies for review.

Empty-shelf incidents reduced by 90% at pilot stores through proactive, camera-driven detection and guided ordering.

Inventory optimization improved by 30%, with right-sized replenishment decreasing carrying costs while maintaining service levels.

Always-on context (inventory, price, promo) enabled 24/7 recommendation freshness, improving fill rates and promo adherence.

Store staff reported faster ordering cycles and fewer missed SKUs, with high adoption due to seamless integration into the existing app.

Key factors behind success:

In-App Simplicity: A photo-first workflow embedded in the existing app minimized behavior change and training needs.

Robust Vision for Variants: Custom models handled look-alike packaging, regional labels, and glare/angle/occlusion.

Multimodal Reasoning: Language+vision models improved disambiguation of pack sizes and flavors, reducing false matches.

Tight Systems Integration: Real-time inventory, pricing, and promotions ensured recommendations aligned with operational constraints and sales objectives.

Resilience by Design: Offline capture, lightweight models, and progressive sync supported diverse LATAM device and network conditions.

Privacy & Governance: On-device preprocessing where feasible, secure APIs, and audit trails protected store and customer data.