Inventory Optimization: How a Knowledge Graph of 3,500 Substitutes Cuts Stockouts 63% and Frees $840K

Key takeaways

• A 520-employee industrial parts distributor running NetSuite manages 12,000 SKUs across 4 warehouses with an 8% stockout rate costing $1.2M/year in lost sales and expedited sourcing — safety stock is set by quarterly spreadsheet review, too slow for seasonal demand shifts.
• 3,500 part substitutes live in 3 senior buyers' heads, not in any system — when a buyer retires, the substitution knowledge leaves the building, and the remaining team cannot fill the gap without months of re-learning.
• A knowledge-graph-augmented agent encoding 3,500 form-fit-function equivalents recommends safety stock per SKU based on demand variability, lead time, and substitute availability, cutting stockouts from 8% to 3% and freeing $840K in working capital — the graph knows which part works, not which part sounds similar.
• Demand-sensing agents improve inventory turnover by 15-25% by reducing safety stock without increasing stockout risk — the substitute-aware agent orders less of part A when it knows part B is a qualified fallback already in stock.
• 94% of procurement executives use generative AI weekly, but only 4% have reached large-scale deployment — inventory optimization is where that gap shows up as measurable balance-sheet impact.

A 520-employee industrial parts distributor with $140M annual revenue runs NetSuite for ERP and a custom demand-forecasting tool across 4 warehouses. It manages 12,000 SKUs, carries $8.2M in safety stock, and posts an 8% stockout rate that costs an estimated $1.2M per year in lost sales and expedited sourcing. Overstock on slow-moving items ties up $2.1M in working capital. The knowledge of which parts substitute for which — 3,500 form-fit-function equivalents — lives in 3 senior buyers' heads. When one of them retires, that substitution knowledge goes with them, and the remaining team spends months re-learning what was never written down. This article maps how a knowledge-graph-augmented agent encoding those 3,500 substitutes, wired to NetSuite through MCP modules and orchestrated with A2A task delegation, drives safety stock per SKU against demand variability, lead time, and substitute availability — cutting stockouts 63% and freeing $840K in working capital, with the human buyer on replenishment decisions above a value threshold.

The problem: quarterly spreadsheet safety stock and tribal substitution knowledge

The distributor's inventory operates on a pattern common to mid-market industrial companies: a quarterly spreadsheet review sets safety stock levels per SKU based on historical demand averages and supplier lead times. The review covers the full 12,000-SKU catalog in batches of 3,000 per week, meaning every SKU is reviewed every 4 weeks at best. Seasonal demand shifts — a 40% volume spike in Q3 for bearings, a 25% drop in Q1 for packaging — hit between review cycles, and safety stock is stale when it matters most.

The 8% stockout rate is the visible cost. The less visible cost is overcorrection: when a stockout occurs, the buying team increases safety stock on that SKU by 30-50% for the next cycle, which produces overstock when demand normalizes. $2.1M sits in slow-moving overstock, much of it the residue of stockout-driven overordering. The quarterly review cycle cannot distinguish between a genuine demand shift and a one-time spike, so it treats both the same way — order more — and the balance sheet absorbs the error.

Flexera's 2026 IT budget report found that 59% of enterprise AI spend produces no measurable ROI — the distributor's quarterly spreadsheet is the inventory equivalent. The team works hard, the spreadsheet is accurate when it is produced, and the output is stale within weeks. The 3,500 substitute mappings compound the problem: when part A stocks out, a senior buyer knows part B is a qualified substitute and can redirect demand. But the spreadsheet does not know this — it treats part A and part B as independent SKUs, each with its own safety stock, when in fact they share a demand pool. The result is duplicate safety stock on substitute pairs and stockouts on parts with no known substitute, simultaneously.

Art of Procurement's 2026 survey reports that 94% of procurement executives use generative AI at least weekly, but only 4% have reached large-scale deployment — this distributor sits squarely in that gap, using AI for market research and supplier discovery but not for the inventory workflow that determines stockout rate and working capital.

The agent-orchestrated solution

The agent layer wraps the existing NetSuite ERP, the demand-forecasting tool, and the supplier database with typed MCP tool calls — the same pattern documented in the NetSuite MCP module pattern and the MCP module code standard. The agent does not replace any system. It connects them as tools and orchestrates the inventory workflow:

MCP modules connect each system. A NetSuite MCP module exposes inventory levels, demand history, and PO creation as typed tools. A supplier database module exposes lead times and reliability scores. The demand-forecasting tool module exposes projected demand per SKU per week. The agent calls these tools with structured requests and receives structured responses — no manual spreadsheet exports, no copy-paste between systems.

The knowledge graph encodes 3,500 form-fit-function equivalents. The graph maps substitute relationships across the 12,000-SKU catalog — if part A stocks out or its price spikes, the agent knows part B is a qualified substitute and adjusts safety stock for both SKUs as a shared demand pool. This is the substitution knowledge that currently lives in 3 senior buyers' heads, captured as structured data that survives retirement. The same GraphRAG retrieval pattern described for customer support — where a knowledge graph walks dependencies and substitutes to find the right answer, not a semantically-similar guess — applies to inventory: the graph knows which part works, not which part sounds similar. LinkedIn Engineering's GraphRAG production deployment improved retrieval accuracy by 77.6% — the same structured-retrieval advantage applies to substitute-aware safety stock.

The agent recommends safety stock per SKU. Instead of a quarterly spreadsheet average, the agent computes safety stock per SKU based on three inputs: demand variability (from the forecasting tool), supplier lead time (from the supplier database), and substitute availability (from the knowledge graph). If part A has a qualified substitute (part B) already in stock, the agent reduces part A's safety stock — the combined pool covers the same demand with less total inventory. If part A has no substitute, the agent increases safety stock to compensate for the higher stockout cost. This is demand-sensing that improves inventory turnover by 15-25% by reducing safety stock without increasing stockout risk — the substitute graph is what makes the reduction safe.

A2A delegation handles demand-forecast subtasks. The orchestrating agent delegates demand-forecasting and anomaly-detection subtasks to a forecasting agent via A2A — the same agent-to-agent delegation pattern described in B2B RFQ Automation with A2A and Hermes Agent. The forecasting agent runs weekly demand projections per SKU, flags anomalous spikes (a 40% week-over-week increase triggers a review), and returns consolidated forecasts. The main agent uses those forecasts to adjust safety stock continuously, not quarterly.

The human stays in the loop on replenishment above a value threshold. The agent auto-recommends PO generation for routine replenishment below $5,000 per line — common MRO and fast-moving items. For replenishment above $5,000, or for any SKU flagged as single-source (no substitute in the graph), the agent prepares the PO but the buyer approves. Every auto-recommendation is logged with a full audit trail in NetSuite, including the substitute analysis that informed the safety stock adjustment.

The outcome

With the agent layer in place, the distributor's inventory shifts from quarterly spreadsheet review to continuous, substitute-aware optimization:

• Stockout rate drops from 8% to 3% — the agent adjusts safety stock weekly against demand forecasts and substitute availability, catching seasonal shifts that the quarterly review missed. The 3% residual covers genuine supplier disruptions, not demand-cycle errors.
• Overstock reduced 40% — $2.1M to $1.26M, freeing $840K in working capital — duplicate safety stock on substitute pairs is eliminated. The graph knows part A and part B share a demand pool, so it carries one pool, not two.
• Inventory turnover improves 15-25% — Ivalua's procurement benchmarking documents this range for demand-sensing agents; the distributor moves from 6 annual turns to a projected 7-7.5, with substitute-aware safety stock as the mechanism.
• Substitution knowledge captured in the graph — 3,500 form-fit-function equivalents encoded as structured data, not in 3 senior buyers' heads. A retirement no longer means lost inventory knowledge.
• Quarterly spreadsheet review replaced by continuous optimization — the 4-week batch cycle is gone. The agent reviews every SKU weekly, and only anomalies (spikes, disruptions, new SKUs) reach the buyer's desk.

The broader context: the AI in procurement market is projected to grow from $4.25B in 2026 to $39.2B by 2035 at 28% CAGR. Mid-market distributors running NetSuite are early in this curve — the agent layer described here is production-ready, not experimental. The balance-sheet impact ($840K freed, $750K in stockout cost recovered) is measurable on the next quarterly close, not a multi-year ROI projection. MarketScale's August 2026 analysis of AI capex scrutiny notes that "vague productivity multipliers and long-horizon payback periods are increasingly insufficient" — inventory optimization is where the multiplier is specific and the payback is one quarter.

Quarterly spreadsheet safety stock and substitute-aware continuous optimization, side by side:

Related reading

• Customer Support at 7-Hour Resolution: How a Knowledge Graph Cuts Ticket Time by 75% — the same GraphRAG substitute-and-dependency retrieval pattern, applied to support tickets instead of inventory pools
• Connecting an AI Agent to NetSuite with MCP — the MCP module pattern that wraps NetSuite's inventory, demand history, and PO creation as typed tools
• MCP Module Code Standard — the structural pattern that makes custom MCP modules production-ready across all connectors

---

A mid-market industrial parts distributor running NetSuite needs a knowledge-graph-augmented agent that encodes 3,500 form-fit-function substitutes, pulls demand variability and lead times through MCP modules, and recommends safety stock per SKU as a shared pool across substitute pairs. The agent cuts stockouts 63%, frees $840K in working capital, and captures substitution knowledge before senior buyers retire. The human stays on replenishment above $5,000.

One-week discovery. You get a system inventory, workflow map, and fixed scope — whether or not you build with us.

Request a scoped build.