4 patterns for this goal
Inventory-optimization agents miscalibrate safety stock using demand-variance estimates computed at the wrong aggregation level (monthly when replenishment is weekly), treat quality-hold flags recorded in free-text inspection notes as not relevant to available-to-promise calculations, perform unit-conversion arithmetic via free-text reasoning instead of deterministic formulas, and borrow variance profiles from wrong-demand-driver analogs when setting safety stock for new SKUs. These patterns concentrate around three categories: safety-stock miscalibration (variance underestimation, wrong analogs), multi-agent handoff brittleness (quality holds lost, reorder calculations operate on stale ATP), and arithmetic and unit-conversion errors in replenishment logic. Inventory errors manifest as stockouts (when safety stock is too low) or excess inventory (when it is too high), both with full working-capital impact.
All four inventory-optimization patterns share a vulnerability to silent state divergence: the replenishment agent’s calculation operates on a view of inventory state (available-to-promise, variance estimates) that is cached, aggregated, or computed on assumptions that no longer hold. When variance is computed at monthly aggregation but replenishment is weekly, the agent’s safety-stock calculation is silently operating on understated variance. When quality holds exist in inspection notes but not in the ATP schema, the agent’s reorder calculation counts held inventory as available. When arithmetic is performed via free-text reasoning instead of a deterministic formula, the result drifts silently from the correct value. None of these failures produce an obvious signal — the replenishment recommendation looks plausible — until the actual demand or supply pattern reveals the miscalibration. Mitigation requires: variance computed at replenishment-cycle granularity with explicit lead-time variance included; structured quality-hold and change-event fields on inventory records; and deterministic calculation formulas for reorder quantities, never free-text reasoning about arithmetic.
Compare realized in-stock rate against the target service level for each SKU on a rolling basis. If a SKU with a 95% target service level consistently runs out (realized 85% in-stock), the safety-stock inputs (variance, lead time) are likely understated. Recalibrate variance inputs and recompute. Test the new variance estimate on a holdout period to check whether realized service level improves to target.
Monthly aggregation smooths out week-to-week spikes. A SKU that sees 50 units one week and 150 the next will show a monthly aggregate (averaging ~600 over 4 weeks) that misses the true week-to-week volatility. Safety stock based on the monthly aggregate is calibrated to a narrower distribution than what actually occurs, producing stockouts during high-variance weeks. Compute variance at the same time granularity as the replenishment cycle.
Add a structured quality-hold field to inventory schemas separate from usable on-hand. Require receiving agents to populate it directly from inspection determinations. Require replenishment agents to exclude held quantities from available-to-promise calculations by default. Run reconciliation checks comparing hold-status language in inspection notes against the structured field before reorder calculations run.
Use a deterministic calculation function/tool for the final order quantity instead of free-text reasoning. Pass tool outputs (forecast, lead-time, pack-size) as typed, structured values to the calculation function, not embedded in a natural-language prompt the model reasons over in prose. The model’s reasoning about which values to retrieve is fine; its arithmetic on those values should be delegated to a formula.
Once the new SKU accumulates enough sales history (typically 4-8 weeks), compute its actual coefficient of variation. Compare it to the borrowed-variance proxy’s COV. If they diverge by more than 20-30%, the proxy was mismatched. Replace the borrowed estimate with the directly-computed one and adjust safety stock accordingly. Track which proxy SKUs consistently mismatch for demand-driver classification audit.
| Pattern | Mechanism |
|---|---|
| Safety Stock Miscalibration from Demand Variance Underestimation | Variance computed at monthly aggregation instead of weekly replenishment cycle; lead-time variance held constant instead of estimated |
| Embedding Retrieval Pulls Wrong Substitute SKU as Safety-Stock Variance Proxy | Variance-proxy selection by description similarity instead of demand-driver classification; topically similar SKU has different volatility profile |
| Multi-Agent Handoff Drops Quality-Hold Flag Between Receiving Agent and Replenishment Agent | Quality-hold status exists in inspection notes but not in available-to-promise schema; replenishment counts held lot as usable |
| Unit-Conversion Arithmetic Drift in LLM-Generated Reorder Quantity | Reorder quantity computed via free-text reasoning over forecast/lead-time/pack-size instead of deterministic formula; arithmetic drifts 1-5 units on non-round packs |
Total: 4 patterns
To Calculate Safety Stock for a New SKU Lacking Sufficient Sales History, an Inventory-Optimization Agent Retrieves the Most Similar Existing SKU via Embedding Similarity Over Product Descriptions to Borrow Its Demand-Variance Profile, but Selects a SKU That Is Textually Similar Yet Has a Fundamentally Different Volatility Pattern, Producing a Safety-Stock Level Calibrated to the Wrong Risk Profile
A Receiving Agent That Notes, in Its Own Inspection Reasoning, That a Newly Received Lot Has Been Placed on Quality Hold Pending Inspection Hands Off Inventory Levels to a Replenishment Agent Through a Structured Available-to-Promise Field That Counts the Held Lot as Available, So the Replenishment Agent Treats Held Stock as Usable and Under-Orders Replacement Inventory
Agent Calculates Safety Stock Using a Demand Variance Estimate That Understates True Variability, Producing Stockouts Despite a "Safety" Buffer
A Replenishment Agent Calls a Demand-Forecast Tool and a Lead-Time/Pack-Size Tool, Both of Which Return Correct Values, but Combines Them Into a Final Purchase-Order Quantity via Free-Text Reasoning Rather Than a Deterministic Calculation, Introducing Arithmetic and Unit-Conversion Errors the Individual Tool Calls Never Made