Inventory Optimization

The Context

A multi-channel retailer with a national store network and significant catalogue complexity wanted to strengthen inventory decision-making across the product lifecycle — from initial buy through to in-season response. They had extensive transaction history but no unified framework to convert it into systematic decision logic.

The Opportunity

The retailer saw potential to improve across four connected areas:

• Cold-start accuracy: Better demand estimation for new items with no sales history
• Allocation precision: Stronger alignment between initial stock placement and actual store-level demand
• Availability uplift: Capturing revenue currently lost when high-demand items sell through too quickly
• Proactive operations: Earlier signals to adjust production and redistribute stock while there's still time to act

The Approach

An AI Director coordinated specialist agents across five parallel workstreams aligned to the inventory lifecycle. Agents worked autonomously — querying the data warehouse, testing hypotheses, validating assumptions, and documenting findings — while the Director synthesized outputs and surfaced cross-cutting issues.

What the Agents Delivered

Data Foundation Agent

The agent established canonical demand definitions that eliminated double-counting risk, assessed data completeness, and designed a “dual-window” analytical strategy — using historical data for pattern learning and recent data for stock-aware validation. Importantly, it tested and rejected a velocity-based stockout detection method due to an unacceptable false positive rate, preventing a flawed assumption from propagating into downstream models.

Ordering Agent

The agent discovered that existing master data classifications were ineffective and developed a behavioural alternative based on observed sales continuity. It designed a Hybrid Matching Protocol for new-item demand estimation — combining hard constraints with soft-weighted attributes — and validated this against historical outcomes. It also surfaced a blocker: reorder logic required planning parameters not yet available, which was documented and parked with a clear data checklist.

Allocation Agent

The agent quantified that initial allocation patterns did not reflect actual demand distribution. In the most demand-dense stores, early stockout patterns suggested sales were being capped by availability rather than demand. The recommended reallocation fell within normal trading volatility, meaning it could be implemented without a phased rollout.

Replenishment Agent

The agent established that traditional low-stock triggers fire too late for fast-moving items. It found sufficient inventory typically existed elsewhere in the network, supporting a “redistribution-first” approach. It designed transfer rules with donor-store guardrails and proposed a “Red Light” rule: block demand-seeking transfers for older stock and recommend markdown-in-place instead.

Early Signal Agent

The agent demonstrated that demand behaviour differs materially across product categories. It tested multiple early-signal candidates and found that item-level sales velocity significantly outperformed alternatives. The output was a 4-tier classification system — Chase Aggressively, Chase, Maintain, Markdown Risk — calibrated by category, designed to feed directly into production and buying decisions.

Value Delivered

Opportunities identified: Significant recoverable revenue from reducing stockouts, majority of inter-store transfers found avoidable through better initial allocation, material inventory investment that could be redirected, and a hybrid proxy approach that dramatically outperformed simpler methods.

Foundations established: Behavioural classification rules that outperform existing master data, validated analytical methodology, early rejection of flawed approaches before implementation, and clear documentation of data gaps with a checklist for closure.

Key Takeaway

The agents didn't just analyse data — they tested hypotheses, rejected flawed approaches before implementation, and documented their reasoning, creating an auditable foundation for inventory optimization.