Written by Tsujita — Paddy & Water Inc. / Supply Chain

An Approach to Supply Chain Optimization Strategy

The Mechanics of PSI (Production–Sales–Inventory) Planning and Practical Methods for Inventory Reduction

Demand Forecasting ・ Safety Stock Optimization ・ Multi-Echelon Inventory Optimization (MEIO) ・ S&OP/IBP

July 2026

This report systematically explains a strategic approach for transforming supply chain management (SCM) — which has traditionally relied on intuition and experience — into a “data-driven” discipline. In particular, it delves into the mechanics of PSI (Production, Sales, Inventory) planning, which integrates supply and demand, and into concrete inventory-reduction methods that simultaneously curb both excess inventory and stockouts (safety stock theory, MEIO, ABC/XYZ analysis, and more). The primary intended readers are practitioners in SCM, production management, and corporate planning at manufacturing and wholesale companies.

Table of Contents

・Chapter 1　Why “Data-Driven SCM” Now

・Chapter 2　The Overall Picture and Framework of SCM Optimization

・Chapter 3　The Mechanics of PSI (Production, Sales, Inventory) Planning

・Chapter 4　Advancing Demand Forecasting ― The Starting Point of PSI

・Chapter 5　Concrete Methods for Inventory Optimization and Reduction

・Chapter 6　SCM KPIs and Monitoring

・Chapter 7　A Roadmap for Implementing Data-Driven SCM

・Chapter 8　Industry-Specific Requirements for Global PSI

・Chapter 9　Methods for Optimizing Purchasing Costs

・Chapter 10　The Technology Stack for SCM Optimization and a Comparison by Technology

・Chapter 11　Implementation Case Studies and Investment Evaluation for Data-Driven SCM

・Chapter 12　Recommended Architecture, Governance, and an Implementation Checklist

・Chapter 13　Conclusion ― Inventory Should Be “Optimized,” Not Simply “Reduced”

Chapter 1　Why “Data-Driven SCM” Now

1-1　Structural Challenges Facing SCM

Supply chains are exposed to unprecedented uncertainty (VUCA) due to widening demand volatility, geopolitical risk, shorter lead times, and increasing SKU (item) diversity. In this environment, companies constantly face the trade-off between “lost opportunity from stockouts” and “cash and disposal losses from excess inventory.” Relying solely on the intuition and experience of individual staff has reached its limits in controlling both simultaneously; data-based visualization and optimization of supply and demand have become indispensable.

1-2　The Two Faces of Inventory ― Asset or Evil

Inventory is indispensable as a “buffer that absorbs demand fluctuations,” but when excessive it becomes a “hidden liability” that strains working capital and generates storage costs, obsolescence, and disposal costs. What matters is not uniformly reducing inventory, but achieving optimal placement — holding “only the necessary amount, in the necessary place.”

・Benefits of inventory: Absorbs timing gaps and fluctuations between demand and supply, prevents stockouts, and levels production

・Drawbacks of inventory: Ties up working capital, incurs storage and insurance costs, causes obsolescence and disposal, leads to quality degradation, and conceals problems (inventory masks underlying issues)

1-3　The Bullwhip Effect ― How Demand Fluctuations Are Amplified

The phenomenon in which small fluctuations in actual retail demand are amplified as they travel upstream through the supply chain — from wholesaler to manufacturer to parts to materials — is known as the “bullwhip effect.” Because each stage forecasts based on “order information” rather than actual demand, and orders extra as a safety measure, the swings in demand grow larger the further upstream one goes, alternately triggering excess inventory and stockouts.

・Main causes: Fragmentation of actual demand information, independent demand forecasting at each stage, batch ordering (lot-sizing), and speculative buying driven by price changes and promotions

・Countermeasures: Sharing actual demand (POS/shipment) data upstream, collaborative replenishment through VMI/CPFR, revisiting order lot sizes, and shortening lead times

✱ The essence of the bullwhip effect lies in the “fragmentation of information.” The first step of data-driven SCM is to create a “single source of truth” in which every stage looks at the same actual-demand data.

Chapter 2　The Overall Picture and Framework of SCM Optimization

2-1　The SCOR Model ― Capturing SCM as Five Processes

The SCOR (Supply Chain Operations Reference) model is widely used as a common language for SCM reform. It breaks the supply chain down into five basic processes and evaluates them with standard KPIs. Data-driven SCM focuses on collecting and integrating data from each of these processes to improve the accuracy of the Plan process.

Process	Content	Key Data	Representative KPI
Plan	Demand forecasting, supply-demand planning, inventory planning	Demand history, forecasts, capacity	Forecast accuracy, plan adherence rate
Source (Procurement)	Procurement of materials, supplier management	Orders, receipts, actual lead times	Procurement lead time, on-time receipt rate
Make (Manufacturing)	Production, assembly	Production orders, results, capacity	Production adherence rate, utilization rate
Deliver	Shipping, transportation, delivery	Orders received, shipments, transport	On-time delivery rate, fill rate
Return	Returns, recalls, retrieval	Returns, defects	Return rate, retrieval lead time

2-2　S&OP／IBP ― Connecting Planning to Management

The process that connects PSI and inventory optimization to management decision-making — rather than leaving them as mere “operational plans” — is S&OP (Sales and Operations Planning), and its evolved form, IBP (Integrated Business Planning). These integrate demand, supply, and finance into a single plan, with top management participating monthly to agree on supply-demand gaps and inventory policy. Data-driven SCM sophisticates the forecasting and inventory planning that underpin this S&OP.

2-3　The Maturity Model of Data-Driven SCM

Stage	State	Planning Method	How Inventory Is Determined
Lv1 Individual-dependent	Excel, intuition and experience	Individual staff judge independently	Fixed number of days based on rules of thumb
Lv2 Visualization	Data aggregation, BI	Visualization of actual results	Historical average + buffer
Lv3 Integrated planning	PSI/S&OP operation	Integrated supply-demand coordination	Statistical safety stock
Lv4 Optimization	MEIO, optimization engines	Optimal solutions under constraints	Optimal network placement
Lv5 Autonomous	AI, digital twin	Demand sensing, automatic replenishment	Real-time dynamic optimization

✱ Many Japanese companies remain at Lv1–2. A realistic path to optimization is to first establish integrated supply-demand management through PSI (Lv3), and then progress in stages toward MEIO and AI (Lv4–5).

Chapter 3　The Mechanics of PSI (Production, Sales, Inventory) Planning

3-1　What Is PSI ― Connecting Production, Sales, and Inventory

PSI is a supply-demand coordination concept that manages the three elements of Production (manufacturing/receiving), Sales (selling/shipping), and Inventory as an integrated whole along a time axis. It links “how much goes out through sales,” “how much comes in through production,” and “what happens to inventory as a result” in a single planning table, preventing excess and stockouts before they occur. In Japanese manufacturing, this is also called “Seihanzai (Production-Sales-Inventory) planning.”

・P (Production/Receiving): Movements that increase inventory, such as production, purchasing, and receiving

・S (Sales/Shipping): Movements that decrease inventory, such as sales, shipments, and consumption

・I (Inventory): The cumulative difference between P and S — the adjustment lever that absorbs timing gaps between supply and demand

3-2　The Structure of the PSI Table ― Inventory Is Determined by “Subtraction”

At the core of PSI is a simple identity that accumulates the inventory balance for each period (week/month). Building a plan so that this equation holds true across all periods, all SKUs, and all locations is, in itself, the practice of PSI planning.

■ The Basic PSI Identity

Ending Inventory ＝ Beginning Inventory ＋ Receipts (P) － Shipments (S)

Example) Beginning 100 ＋ Production receipts 80 － Sales shipments 120 ＝ Ending inventory 60

　→ This ending balance of 60 is carried forward as the beginning balance of the following period (rolling)

Expanding this table horizontally (along the time axis) makes it possible to see at a glance in which future week inventory will become excessive and in which week a stockout (negative inventory) will occur. The value of PSI lies in being able to take early action — such as “advancing production” or “sourcing alternative supply” — as soon as a stockout becomes visible.

Item	Week 1	Week 2	Week 3	Week 4
Beginning inventory	100	60	40	70
Receipts P (Production)	80	100	130	90
Shipments S (Sales)	120	120	100	110
Ending inventory I	60	40	70	50
Safety stock line	50	50	50	50

In the table above, inventory approaches the safety stock line (50) in Week 2. PSI visualizes such “danger zones ahead” in advance, prompting adjustment decisions such as bringing production forward to Weeks 1–2.

3-3　Sales-Driven “Backward-Calculation” Planning

PSI is basically built top-down, starting from the sales plan (demand forecast) and working backward to derive the production/replenishment volume needed to maintain the required inventory. The necessary receipt volume is derived from “expected sales volume + inventory that should be held − current inventory.”

■ Backward Calculation of Required Production/Replenishment

Required Receipts (P) ＝ Period Sales (S) ＋ Target Ending Inventory － Beginning Inventory

　　Target Ending Inventory ＝ Safety Stock ＋ Cycle Stock (provision for the following period’s sales)

In practice, however, there are constraints such as production capacity, minimum order quantities (MOQ), and procurement lead times. The “plan as it should be,” derived by backward calculation, is therefore reconciled bottom-up against the supply side’s “plan as it can actually be produced/transported,” adjusting the supply-demand gap. This reconciliation of supply and demand is the crux of PSI operations.

3-4　The Multidimensional Nature and Operating Cycle of PSI

・Multidimensional: PSI is managed across three dimensions — “SKU × location (plant, warehouse, channel) × period.” Inter-site inventory transfers are also expressed as receipts and shipments

・Buckets: Short cycles (weekly/daily) are used alongside medium-to-long term (monthly) cycles. The near term is planned with precision, the far term in broad strokes

・Rolling: Ending inventory is carried forward to the beginning of the following period, and the plan is continually updated on a weekly/monthly basis through “rolling planning”

・Role division: Demand planning (sales/marketing), supply planning (production/procurement), and inventory policy (SCM/management) are all agreed upon on the same PSI

[Case Study] A consumer goods manufacturer with large seasonal swings: Introduced “backward-calculation management,” using the monthly sales plan as the starting point and working backward through PSI to determine production and inventory. By leveling production ahead of demand peaks and dynamically reviewing safety stock, it reduced the risk of excess inventory by more than 30%.

Chapter 4　Advancing Demand Forecasting ― The Starting Point of PSI

The accuracy of PSI is determined by the accuracy of the demand forecast (sales plan) that serves as its starting point. If the forecast is off, that error translates directly into either excess inventory or stockouts. Data-driven SCM combines statistics and AI to improve forecast accuracy, while also operating it together with a safety-stock design that assumes forecasts will, to some degree, be wrong.

4-1　Forecasting Methods ― From Statistics to AI and Demand Sensing

Method	Overview	Suitable Cases
Moving average / exponential smoothing	Smoothing of past results	Stable, staple items
Seasonal/trend decomposition	Separates seasonality and trend	Seasonal products
Machine learning (regression, gradient boosting)	Learns multiple variables (price, weather, promotions)	Items with many volatility factors
Demand sensing	Detects short-term demand in near real time using POS, weather, SNS, etc.	Items with highly volatile demand

In recent years, “demand sensing,” which incorporates a variety of external data — weather, temperature, promotional calendars, social media sentiment, and more — to capture localized, short-term demand fluctuations with high precision, has become practical AI.

4-2　How to Measure Forecast Accuracy

To improve a forecast, its accuracy must first be measured quantitatively. Representative metrics are as follows.

・MAPE (Mean Absolute Percentage Error): The average of |actual − forecast| / actual. The lower, the more accurate. Target values should be set according to item characteristics

・Bias: Whether the forecast is consistently too high or too low. A positive bias breeds excess inventory

・Forecast Value Added (FVA): Verifies whether manual intervention actually improves accuracy over a naive forecast, eliminating wasteful manual adjustments

✱ “100% forecast accuracy” is impossible. What matters is running both efforts in parallel: improving accuracy, and designing safety stock to absorb the error that remains. Forecast improvement and inventory optimization cannot be separated.

Chapter 5　Concrete Methods for Inventory Optimization and Reduction

This chapter, alongside PSI, is the core of this report. It breaks down “why inventory is held” and explains, from both theoretical (formula-based) and practical perspectives, concrete levers for reducing inventory without worsening stockouts.

5-1　Breaking Down Inventory ― Which Inventory Can Be Reduced

The first step in reducing inventory is to break it down by purpose, separating “inventory that can be reduced” from “inventory that is necessary.”

Type of Inventory	Role	Reduction Approach
Cycle stock	Regular inventory held in order-lot units	Optimize order lot size/frequency (EOQ)
Safety stock	Buffer for demand and lead-time variability	Improve forecast accuracy, shorten lead times, right-size service levels
Work-in-process (WIP) inventory	Intermediate inventory between processes	Shorten lead time, level processes, eliminate bottlenecks
In-transit inventory	Inventory in transit	Shorten transport lead time, review site locations
Strategic / anticipation stock	Provision for demand peaks and risk	Improve supply-demand planning accuracy, level production
Stagnant / dead stock	Unsold, obsolete inventory	Early visibility, disposal, halt ordering

5-2　The Theory of Safety Stock ― The Key Lever for Reduction

The single most effective factor in reducing inventory is optimizing safety stock. Safety stock is the buffer inventory held to meet a target service level (the probability of not stocking out) against “variability in demand and lead time,” and is expressed by the following formula.

■ The Basic Safety Stock Formula

SS ＝ Z × σ_d × √LT (when lead time is stable)

More precisely (when both demand and lead time vary):

SS ＝ Z × √( LT × σ_d² ＋ d² × σ_LT² )

　Z ＝ Service-level coefficient (e.g., 95%→1.65, 99%→2.33)

　σ_d ＝ Variability of demand (standard deviation)

　σ_LT ＝ Variability of lead time (standard deviation)

　d ＝ Average demand　／　LT ＝ Average lead time

This formula clearly indicates which levers are effective for reducing inventory. Because safety stock is proportional to each of these factors, improving any of the following directly reduces safety stock.

・Lower Z: Rather than uniformly imposing a high service level on every item, use ABC analysis to apply a high level only to important items (differentiation)

・Lower σ_d: Improve demand forecast accuracy and level out demand volatility itself (leveling promotions, sharing information)

・Shorten LT: Safety stock is proportional to the square root of lead time. Shortening lead time is extremely effective

・Lower σ_LT: Reduce variability in lead time. Stable procurement and reliable suppliers reduce inventory

✱ The most important point is that “safety stock is proportional to the square root of lead time.” For example, cutting lead time to a quarter of its original length halves safety stock (√1/4＝1/2). Shortening lead time is the single strongest lever for inventory reduction.

5-3　12 Levers for Inventory Reduction (Overview)

#	Lever	Mechanism / Effect
1	Improve demand forecast accuracy	Lowers σ_d, reducing safety stock. Leverages AI/demand sensing
2	Shorten lead time	The strongest lever, working through √LT. Local sourcing, in-house production, shorter processes
3	Reduce lead-time variability	Compresses σ_LT through stable procurement, dual sourcing, and supplier evaluation
4	Right-size service levels	Differentiates Z through ABC analysis. Stops over-guaranteeing uniformly across all items
5	ABC/XYZ analysis	Differentiates inventory policy by importance × variability
6	MEIO (multi-echelon optimization)	Pools risk across the entire network, compressing total inventory
7	Optimize order lot size (EOQ)	Minimizes cycle stock by balancing ordering cost and holding cost
8	Risk pooling (site consolidation)	Consolidates inventory to offset variability, reducing total safety stock
9	Postponement	Holds common semi-finished goods, finalizing them after demand is confirmed, compressing the variety of inventory
10	Parts commonization / modularization	Reduces the number of items, aggregating demand and making it more predictable
11	VMI/CPFR	Suppresses the bullwhip effect and inventory through actual-demand-linked collaborative replenishment
12	Visualize and dispose of stagnant inventory	Identifies dead stock early and halts ordering, disposes of, or repurposes it

5-4　ABC/XYZ Analysis ― A Differentiated Inventory Policy

Applying the same inventory policy to every item is inefficient. By combining ABC analysis (importance by value/shipment volume) with XYZ analysis (magnitude of demand variability), items are divided into nine quadrants, each with a differentiated policy.

Category	Characteristics	Recommended Inventory Policy
AX (Important, stable)	Best-sellers, easy to forecast	Lean inventory, high turnover. Automatic replenishment / JIT
AZ (Important, highly variable)	Best-sellers but hard to read	Thicker safety stock. Focus on demand sensing
CX (Low importance, stable)	Low value, stable	Simplify ordering, batch orders
CZ (Low importance, highly variable)	Low value, hard to read	Make-to-order / arrange as needed. Minimize standard stock

5-5　MEIO ― Optimizing Inventory Across the Entire Network

Traditional inventory management calculates safety stock site by site, at a single echelon, which tends to cause inventory to be duplicated and excessive across the supply chain as a whole. MEIO (Multi-Echelon Inventory Optimization) captures the entire multi-tier network — plant → central warehouse → regional warehouse → store — as a single system, and uses mathematical optimization to solve “at which tier, and how much, inventory should be placed to satisfy overall service with the minimum total inventory.”

・Risk pooling: Consolidating inventory upstream offsets individual fluctuations, reducing the total safety stock required

・Optimal placement: Places responsiveness-oriented inventory downstream, close to demand, and general-purpose buffer inventory upstream, to achieve overall optimization

・Required data: Item master, BOM, site hierarchy, lead times and their variability, current/in-transit inventory, actual demand history, ordering constraints (MOQ, etc.)

[Illustrative Effect] A manufacturer with a multi-echelon network: By optimizing, via MEIO, safety stock that had previously been optimized independently at each site, the company reduced total inventory by 10–30% while maintaining the same service level. This was driven by risk pooling that consolidated upstream inventory.

5-6　Order Methods and EOQ ― Minimizing Cycle Stock

Cycle stock arises from “batch ordering.” Because the cost per order (setup, transport) and the holding cost are in a trade-off relationship, the order quantity that minimizes the sum of both is the economic order quantity (EOQ).

■ Economic Order Quantity (EOQ)

EOQ ＝ √( 2 × D × S ／ H )

　D ＝ Annual demand　S ＝ Cost per order　H ＝ Annual holding cost per unit

Order methods: choose between fixed-quantity ordering (reorder-point method) and fixed-interval ordering (periodic ordering) depending on item characteristics

5-7　Countermeasures for Stagnant and Dead Stock

・Visualization: Automatically extract dead stock by days stagnant and last shipment date, visualizing inventory value and space occupied

・Order-hold rules: Automatically place items that have not moved for a certain period on order hold and issue alerts

・Exit design: Decide in advance, as “cut-loss rules,” on discounted sales, transfer to other sites, repurposing, or planned disposal

⚠ Leaving stagnant inventory unaddressed causes a double loss. Storage costs continue to be paid while the inventory is eventually written off as a disposal loss. Rather than holding onto it out of reluctance to “waste” it, an early cut-loss decision ultimately protects both cash and space.

Chapter 6　SCM KPIs and Monitoring

6-1　Key KPIs for Measuring Inventory and Supply-Demand

KPI	Definition	Meaning
Inventory turnover	Cost of goods sold ÷ average inventory	How many times inventory turns over per year. Higher is more efficient
Days of inventory (DOS/DII)	Average inventory ÷ daily shipments	How many days’ worth of inventory is held. Shorter is leaner
Fill rate	Quantity delivered immediately ÷ quantity requested	The proportion met without stocking out
Stockout rate	Occurrences of stockout ÷ demand	The inverse of service level
Forecast accuracy (MAPE)	Forecast error rate	The quality of the foundation of PSI/inventory planning
CCC (Cash Conversion Cycle)	Days of inventory ＋ days sales outstanding − days payable outstanding	How quickly inventory converts back to cash. Working-capital efficiency
GMROI	Gross margin ÷ average inventory cost	The gross-margin efficiency generated by inventory investment

6-2　Visualization Through a Control Tower

Rather than leaving KPIs as monthly after-the-fact reports, they are operated through a “control tower” that integrates and monitors supply-demand, inventory, and transportation in real time. When an anomaly (stockout risk, excess, delay) is detected, it triggers an immediate alert and feeds back into PSI replanning. This evolves SCM from “reactive response” to “anticipatory response.”

✱ KPIs should be viewed as a “trade-off set.” Reducing inventory alone increases stockouts. Monitoring turnover rate, fill rate, and forecast accuracy simultaneously on a dashboard, and searching for the point of balance, is the essence of optimization.

Chapter 7　A Roadmap for Implementing Data-Driven SCM

7-1　A Phased Approach

Phase	Initiative	Prerequisite Data	Outcome Achieved
Phase 1: Visualization	Centralize inventory/supply-demand data and BI	Item/site/lead-time/actual-demand data readiness	Quantitative grasp of the current state
Phase 2: PSI integration	Integrate production-sales-inventory via PSI and reconcile supply-demand	Sales plan, production capacity	Early detection of excess/stockouts
Phase 3: Inventory optimization	Statistical safety stock, ABC/XYZ	Demand and lead-time variability	Realization of appropriate inventory levels
Phase 4: Advanced optimization	MEIO, S&OP/IBP operation	Network-wide data	Structural reduction of total inventory
Phase 5: Autonomy	AI forecasting, demand sensing, automatic replenishment	Real-time data	Dynamic, autonomous supply-demand adjustment

7-2　Success Factors and Pitfalls

・Data quality is everything: Both forecasting and optimization are only as good as the quality of the item master, lead-time history, and actual demand history (garbage in, garbage out)

・Process and organization: The real challenge is building the “operating model” through which sales, production, and SCM agree on a single PSI — more than tool adoption itself

・Phased introduction: Rather than aiming straight for MEIO/AI, first solidify supply-demand integration through PSI, then advance in stages

・Simultaneous KPI monitoring: Targeting inventory reduction alone invites stockouts. Manage fill rate and turnover rate together as a set

⚠ Tools are not magic. Even an advanced optimization engine will not function without a demand-planning consensus process and proper data readiness. Building the foundation in the order “mechanism (PSI/S&OP)” → “data” → “tools” may look like a detour, but is in fact the fastest path.

Chapter 8　Industry-Specific Requirements for Global PSI

Operating the PSI described so far across a global supply chain spanning multiple countries, sites, and currencies is what constitutes global PSI. Compared with domestic PSI, transport lead times are longer, tariffs, trade regulations, and geopolitical risk come into play, and demand locations and supply locations are geographically separated. Furthermore, because product characteristics, regulations, and demand patterns differ greatly by industry, the functionality required of PSI must also be tailored industry by industry.

8-1　How Global PSI Differs From Domestic PSI

・Multi-site, multi-country, multi-currency: Plants, regional warehouses, and sales sites span national borders, requiring conversion of currencies, units, and calendars

・Long transport lead times: Ocean freight and other lead times spanning several weeks. In-transit inventory is substantial, increasing the importance of forward-looking planning

・Trade, tariffs, and regulation: Country of origin, tariffs, and import/export regulations (the GTS domain) affect procurement and allocation decisions

・Inter-site allocation (DRP): Distribution Requirements Planning — deciding which site produces and which site to allocate to — is added

・Risk diversification: The risk of relying on a single supply source or single site requires a design of dual/multi-sourcing and risk pooling

8-2　PSI Requirements by Industry

Even with the same PSI, the “functions that matter most” differ completely between consumer goods, pharmaceuticals, automotive, and capital equipment. The demand characteristics and required PSI functionality for four representative industries are organized below.

Industry	Product/Demand Characteristics	PSI Functions Emphasized	Inventory Considerations
Consumer goods (CPG)	Many SKUs, multiple channels, sharp swings from promotions and seasonality, short life cycles	Promotion-aware demand sensing / channel-specific PSI / expiration (FEFO) management	High service levels. Keep best-sellers lean and fast-moving, minimize slow-movers
Healthcare / Pharmaceuticals	Strict regulation, traceability, temperature control, stockouts affect patients	Lot/expiration management, FEFO, serialization, emphasis on supply responsibility	Thicker safety stock. Prioritize regulation and quality over cost efficiency, strictly avoid stockouts
Automotive	JIT/JIS, two-tier forecast and firm orders, multi-tier supplier network	Two-tier planning of forecast (advance notice) and firm orders / sequential kanban replenishment / leveling	Minimize in-process inventory. Absorb variability through leveled production
Capital equipment / industrial machinery	Engineer-to-order / make-to-order (ETO/MTO), long-lead-time parts, project-based	Multi-year demand forecasting / advance procurement of long-lead-time parts / project-specific P&S plus service-parts planning	Minimize finished-goods inventory; treat long-lead-time parts and maintenance parts as strategic stock

8-3　Consumer Goods (CPG) ― A Battle of Speed and Freshness

Consumer goods handle numerous SKUs across many channels, and demand shifts sharply with promotions, seasons, and trends. PSI must incorporate demand sensing that accounts for promotional effects and weather, channel-specific supply-demand planning, and — for food products — inventory turnover based on expiration dates (FEFO). Because service level (avoiding shelf stockouts) directly affects brand value, the key is “differentiation” — keeping best-sellers lean and fast-turning while quickly narrowing down slow-movers.

8-4　Healthcare / Pharmaceuticals ― Regulation and Supply Responsibility

Pharmaceuticals and medical devices presuppose strict regulation, complete traceability, and a cold chain (temperature control). Rigorous management of lots and expiration dates, FEFO (First-Expired, First-Out), and serialization (unit-level identification) are essential. Because stockouts can directly threaten patients’ lives, safety stock is generally set thicker than in other industries, and the basic design prioritizes service level and compliance over cost efficiency.

8-5　Automotive ― Two-Tier Planning of Forecast and Firm Orders

Automotive supply is fundamentally JIT/JIS, synchronized with the OEM’s production schedule, requiring a PSI that handles both medium-to-long-term “forecasts (advance notice)” and short-term “firm orders” in two tiers. Demand is leveled across the entire multi-tier supplier network, and kanban-based sequential replenishment minimizes in-process inventory. Batch traceability and EDI integration, along with close information sharing upstream, are key to suppressing the bullwhip effect.

8-6　Capital Equipment / Industrial Machinery ― Long Lead Times and Project-Based Work

Capital equipment and industrial machinery are centered on make-to-order production (ETO/MTO), converting the project pipeline into a multi-year demand forecast and procuring long-lead-time parts in advance to mitigate risk. Both project-specific PSI (material planning by project) and long-term after-delivery maintenance-parts (service parts) planning are required in tandem. Evaluating the supply risk of long-lead-time parts in advance through what-if scenarios determines competitiveness.

✱ There is no single “correct answer” for global PSI. The starting point of global inventory optimization is designing which functions to emphasize, tailored to the product characteristics, regulations, and demand volatility of each industry.

8-7　Functions Commonly Required Across a Global PSI Platform

・Visualization of multi-site inventory: Consolidates inventory and in-transit inventory worldwide into a single source of truth

・Inter-site allocation (DRP): Plans optimal allocation and replenishment to demand locations

・Constraint handling: Incorporates production capacity, transportation, tariffs, and currency into scenarios

・Inventory policy by service level: Differentiation based on ABC/XYZ and industry characteristics

・Collaboration: Suppresses the bullwhip effect by sharing plans across sites and with suppliers

Chapter 9　Methods for Optimizing Purchasing Costs

Purchasing costs (procurement spend) for materials and outsourcing account for the majority of SCM costs. Optimizing purchasing costs therefore directly affects the overall profitability of SCM. Rather than mere “price negotiation,” it is important to systematically combine multiple layers of methods, from spend visibility to strategic sourcing to specification review. Strategic purchasing reform is generally said to bring about a 10–20% cost reduction.

9-1　Spend Analysis ― Visibility as the Starting Point

Purchasing reform begins with accurately understanding “what is being bought, at what price, and from whom.” Spend analysis classifies and visualizes company-wide spend by category, supplier, and department, identifying opportunities for reduction.

・Grasping tail spend: Small-value, numerous, dispersed purchases (the long tail) are a breeding ground for unmanaged cost

・Maverick spend: Detects off-contract, rogue purchases, and leverages contract consolidation to improve pricing

・Opportunities for supplier consolidation: Visualizes whether the same item is being purchased separately from multiple suppliers

9-2　Category Management and the Kraljic Matrix

Category management divides spend into categories (item groups) and formulates a strategy for each category based on market conditions. The classic framework for this strategy formulation is the Kraljic Matrix, which divides items into four quadrants along two axes: supply risk and profit impact.

Quadrant	Characteristics (Supply Risk × Profit Impact)	Recommended Purchasing Strategy
Strategic items	High risk × high impact	Long-term partnership, joint development, dual sourcing
Bottleneck items	High risk × low impact	Prioritize securing supply, safety stock, develop alternatives
Leverage items	Low risk × high impact	Competitive bidding, consolidated purchasing to maximize price reduction
Non-critical items	Low risk × low impact	Automate and simplify the purchasing process

9-3　Key Methods for Cost Optimization

#	Method	Mechanism / Effect
1	Consolidated purchasing (volume aggregation)	Bundles purchase volume across sites and departments, lowering unit price through scale
2	Competitive bidding / dual sourcing (e-sourcing)	Creates a competitive environment through RFQs/reverse auctions to optimize price
3	TCO (Total Cost of Ownership) analysis	Judges based on total cost — including maintenance, inventory, defects, and disposal — rather than purchase price alone
4	Should-cost analysis	Negotiates based on the “cost that should be” by building up raw-material cost, processing cost, and margin
5	VA/VE (Value Analysis / Value Engineering)	Reviews design and specifications to lower cost without sacrificing function
6	Specification standardization / commonization	Corrects over-specification and commonizes parts to reduce both item count and unit price
7	Demand management	Reviews the volume and frequency purchased in the first place, curbing spend from the demand side
8	Optimizing payment terms	Optimizes cash and cost through early-payment discounts and adjusted payment terms
9	Global / nearshore sourcing	Reviews sourcing locations to balance unit price, tariffs, and lead time
10	Contract/SLA optimization	Designs price-revision clauses, volume commitments, and SLAs favorably within contracts

9-4　The Strategic Sourcing Process

These methods deliver lasting effect when run not as one-off actions but as a continuous cycle known as “strategic sourcing.”

1. Spend analysis: Visualize spend by category and supplier and select priority areas

2. Market analysis: Investigate the structure of the supply market, supplier capabilities, and cost structure

3. Strategy formulation: Based on the Kraljic matrix, decide on policies such as consolidation, dual sourcing, or partnering

4. Sourcing execution: Evaluate and select candidates through RFI/RFQ/reverse auctions

5. Negotiation and contracting: Negotiate based on TCO/should-cost evidence, and conclude contracts and SLAs

6. Execution and evaluation: Measure delivery quality and cost-reduction impact, feeding into the next cycle (continuous improvement)

9-5　The Use of Digital and AI

・Spend analysis AI: Automatically classifies and deduplicates spend data, surfacing reduction opportunities

・E-sourcing: Digitizes RFQs, reverse auctions, and bid evaluation (AI-assisted proposal comparison)

・Should-cost AI: Automatically estimates the “cost that should be” from drawings and specifications, supporting negotiation

・Contract intelligence: AI analyzes contract clauses, detecting risk clauses and renewal deadlines

[Illustrative Effect] Thorough strategic sourcing and TCO application: Combining spend visibility, supplier consolidation, competitive bidding, and specification optimization is generally said to reduce procurement costs by 10–20%, with some examples of TCO-based procurement achieving up to 30% cost reduction over three years.

✱ The key to purchasing cost optimization is not “hammering on price” but thinking in terms of total cost (TCO) and value (VA/VE). Short-term discounting can damage relationships and lower quality. Running the cycle of visibility → strategy → continuous improvement leads to sustainable reduction.

Chapter 10　The Technology Stack for SCM Optimization and a Comparison by Technology

The results of data-driven SCM rarely emerge from “AI/ML alone.” What recent review research shows is that the configuration with the highest reproducibility layers machine learning for demand forecasting, mathematical optimization for constrained decision-making, simulation for uncertainty assessment, and a digital twin for overall visualization, all on top of a common data foundation, forming a “closed loop” that feeds results back into execution through ERP, WMS, TMS, and S&OP. This chapter surveys the applicable conditions, strengths, and limitations of each technology.

10-1　A Comparison by Technology

Technology Area	Representative Methods	Suitable Conditions	Main Strengths	Main Limitations
Statistical forecasting	ETS, ARIMA, seasonal adjustment	Stable history, few explanatory variables	Highly interpretable, lightweight to implement	Weak for new products, intermittent demand, external shocks
Gradient-boosting ML	XGBoost, LightGBM	External features available (weather, promotions, price)	Handles external factors well; good accuracy/speed	Long-term dependencies require additional design
Deep-learning forecasting	LSTM, TCN, Transformer	High granularity, many series, large data volume	Learns nonlinear patterns and long-term dependencies well	Training cost, explainability, MLOps burden
Constrained optimization	LP, MILP, MINLP	Clear constraints, auditability required	Constraint compliance, cost minimization, accountability	Modeling effort, design for dynamic adaptation
Optimization under uncertainty	Stochastic programming, robust optimization	High demand/supply/risk uncertainty	Prepares for variability and worst-case scenarios	Scenario design and computational load
Reinforcement learning	DQN, A2C, MARL	Frequent sequential decision-making, dynamic control	Sequentially improves policy; strong for dynamic control	Large action spaces, explainability, safety constraints
Simulation	Discrete-event, Monte Carlo, agent-based modeling	Situations hard to formulate mathematically, what-if analysis	Strong reproducibility of real-world conditions and sensitivity analysis	Not itself optimization; depends on assumptions
Digital twin	Synchronization + simulation + optimization	End-to-end visibility and scenario comparison	E2E visibility, predictive/prescriptive analysis	Often stalls at pilot stage; heavy integration/investment burden
Data pipeline	CDC, API/EDI, event streaming, Lake	Integration across multiple systems/companies/layers	Single source of truth, real-time, auditable	Requires lead time before value is realized

10-2　The Most Reproducible “Winning Approach”

Synthesizing the body of review research, the winning approach is not to deploy demand-forecasting ML in isolation, but to connect “how forecast error affects safety stock, replenishment, and transportation” all the way through to optimization and simulation. Results come from this chain of mathematical connections: improving demand-forecast accuracy directly reduces safety stock (lowering σ in the safety-stock formula of Chapter 5), and sensitivity-analysis simulation then verifies robustness.

10-3　Data Pipelines ― Creating a Single Source of Truth

・Common data cloud: Integrates demand, inventory, transportation, and supplier information, preventing a patchwork of locally optimized silos

・Integration methods: CDC, API/EDI, event streaming, and lakehouse architectures ensure real-time reflection and auditability

・Event standards (GS1 EPCIS): Facilitates cross-company sharing of traceability, freshness, and certificate information (IBM Food Trust is a reference example)

✱ SCM optimization failures often stem not from “AI being weak,” but from the operational domain not yet being sufficiently streamlined for AI to handle. It is a cardinal rule to put data integration and business-process integration in place before advancing model sophistication.

Chapter 11　Implementation Case Studies and Investment Evaluation for Data-Driven SCM

11-1　Publicly Disclosed Real-World Cases (Named, Quantified)

Below are relatively concrete publicly available case studies, organized by success and by difficulty. Without exception, the successful cases put data integration and business-process integration in place before the forecasting model itself.

Case	Industry / Scale	Key Effects	Implication
SLB	Energy, ultra-large scale	~90% forecast accuracy in some businesses; over $1B in inventory reduction over 4 years; DIO down 37% in 3 years	E2E process, change management, and data integration were the success factors (SAP IBP)
Super Retail Group	Retail, revenue over $2B	Rolled out integrated forecasting/replenishment in 10 months; on-hand and safety inventory down approximately 20%	Ease of piloting and scaling in the cloud was key (Blue Yonder)
ODP Corporation	Retail, B2B, 3PL	Forecast accuracy up 16% in retail / up 5% in B2B and e-commerce; $30M in inventory reduction	Moving away from hand-built logic (Blue Yonder)
Amazon Pharmacy	Healthcare retail	Daily MAPE of 5% (better than the industry target of under 10%); approximately 5 hours of manual work saved per week	Lowering the granularity to “daily” was key to success (AWS Supply Chain)
Starbucks	Foodservice, global	(A difficult case) AI inventory-count misreads, rollback of automated ordering	Caused by fragmented suppliers, legacy core systems, and overly hasty field automation

✱ What the difficult cases have in common is a lack of supplier standardization, inconsistent field inventory data, legacy core systems, and overly hasty automation. The failed cases show that “rushing automation on top of fragmented data spreads distrust before it improves accuracy.”

11-2　A Risk-Adjusted ROI Model

The effect of a digital twin and similar initiatives does not appear as a single, maximized KPI, but as the sum of many small-to-medium effects. Investment evaluation is therefore best served by a risk-adjusted ROI model that treats cost, probability of realization, and business impact separately.

■ Risk-Adjusted ROI Model

Total Benefit ＝ Inventory reduction ＋ Holding-cost reduction ＋ Stockout/lost-sales improvement ＋ Expedited-shipping reduction

　　　　＋ Disposal/obsolescence reduction ＋ Planning-productivity improvement

Risk-Adjusted Benefit ＝ Total Benefit × DQ × AD × PX × CX

　DQ ＝ Data-quality coefficient　AD ＝ Field-adoption coefficient

　PX ＝ Partner-connectivity coefficient　CX ＝ Compliance/security coefficient

5-Year ROI ＝ (5-year cumulative risk-adjusted benefit − 5-year cumulative cost) ／ 5-year cumulative cost

11-3　Base-Case Estimate and Sensitivity Analysis

In an estimate modeled on a mid-to-large company with $1B in revenue, $0.7B in COGS, and 60 days of inventory, the base scenario yields a 5-year ROI of approximately 73%, with payback in about 20 months. Importantly, the “realization coefficients (DQ/AD/PX)” affect ROI far more than the technology itself.

Scenario	Annual Benefit	5-Year ROI	Payback Period	Representative Assumptions
Pessimistic	$2.6M	~8%	~36–40 months	Delayed master-data readiness, low adoption rate, limited external connectivity
Base	$3.8M	~73%	~20 months	Phased rollout limited to demand, inventory, and replenishment
Optimistic	$5.2M	~130%	~12–14 months	Integrated through forecasting, inventory, and transportation, with high field utilization

11-4　A Quantification Framework for Key Risks

Risk	Measurement Indicator	Example Threshold	Reflection in ROI
Data quality	Missing-data rate, master-data consistency rate, lead-time-history reliability	Caution below 95% consistency rate for key fields	Adjust DQ 0.6–0.9
Organizational adoption	Planner override rate, exception-handling time	Persistently high override rate erodes benefit	Adjust AD 0.5–0.95
External connectivity	Tier 1/3PL connection rate, ASN/EDI/API coverage	Limited effect if Tier 1 coverage is under 60%	Adjust PX 0.4–0.95
Regulation / security	DPIA, cross-border transfer controls, access rights	Operational restrictions and added cost if not in place	Adjust CX 0.6–1.0

⚠ The realization coefficients are decisive: If any of data quality, field adoption, or partner connectivity falls below 0.7, realized ROI is significantly eroded even when the modeled benefit is large. The success or failure of an AI implementation is explained by “data and system requirements,” “the implementation process,” and “cross-organizational integration.”

Chapter 12　Recommended Architecture, Governance, and an Implementation Checklist

12-1　Recommended Architecture ― Connecting the Analytics Layer and Execution Systems in a Closed Loop

The recommended architecture places an analytics layer (forecasting, optimization, simulation, digital twin) on top of a common data cloud, continuously feeding back with the execution systems — ERP, WMS, TMS, and IBP — through KPIs and events.

・Common data cloud: Unifies demand, inventory, transportation, and supplier information, preventing a patchwork of locally optimized silos

・Digital twin as the innovation layer: Rather than replacing existing systems wholesale, it is layered on top as a higher-level layer that optimizes the input policies of TMS/WMS/IBP

・Closed loop: Feeding results back into execution systems and KPI feedback are placed in the same loop, making it possible to learn not just model accuracy but operational quality

・Event-driven integration: In addition to API/EDI, standard event formats such as GS1 EPCIS connect traceability, freshness, and certificate information

12-2　Governance and Regulatory Response ― As a Constraint on Realizing Returns

Over the next five years, requirements for data sharing, cross-border transfer, AI governance, and cyber controls will rise under the EU AI Act, EU Data Act, GDPR, NIS2, and similar regulations. The recommended strategy is not one that competes purely on forecast accuracy, but one advanced together with governance design.

・Data contracts and access control: Data-sharing rules (Data Act) together with permissions, encryption, and audit logs

・Model auditing: Risk classification, model review, and drift monitoring in preparation for AI Act applicability

・Human approval for exceptions: Expand the scope of automated AI decisions gradually, while retaining human approval and preserving logs

・Privacy: GDPR technical and organizational measures, DPIA, cross-border transfer (SCC/BCR)

12-3　Principles of the Execution Roadmap ― Avoiding “Big Bang”

Execution follows METI’s three stages (internal visibility → inter-company visibility → predictive analysis and plan reflection through a digital twin), progressing from internal visibility and foundational readiness in the short term, to intra-company optimization in the medium term, and to inter-company collaboration and digital-twin adoption in the long term.

・Avoiding big bang: Broadly rolling out field automation on top of fragmented data spreads distrust before it improves accuracy

・Pilot KPI criteria: In the short term, base decisions not on “company-wide adoption rate” but on “how much the pilot changed KPIs”

12-4　Implementation Checklist

Area	Minimum Requirements	Desired State
Data governance	Common master for SKU, site, business partner, lead time, and unit of measure	Data contracts, lineage, quality SLAs
Security	Access management, encryption, audit logs, backups	DPIA, cross-border transfer assessment, zero trust
Model operations	Baseline, accuracy monitoring, override records	Drift detection, champion/challenger
Business process design	Division of responsibility for S&OP/S&OE, manual-intervention rules	Exception-centric operations, reason-code analysis
Organization / skills	Joint team of SCM and IT	A composite team spanning OR/ML/MLOps/business design
Inter-company collaboration	Priority connections with Tier 1 and 3PL, EDI/API	Multi-tier visibility, event standards
Regulatory response	Mapping applicability to GDPR/NIS2/AI Act/Data Act	Regional data residency, AI risk classification

12-5　Four Best Practices

・Don’t stop at forecasting: Connect it through to inventory, replenishment, and transportation as a single mathematical chain

・Keep pilots to a single domain: Collect field override reasons and feed them back into model improvement

・Be selective about partner connections: Rather than all directions at once, start with the Tier 1/Tier 2 partners that contribute most to OTIF and inventory

・Expand automated decisions in stages: Gradually widen the scope of AI’s automated decisions while retaining human approval and preserving logs

✱ The essence of the investment decision is not “whether to introduce AI,” but to what level of the hierarchy data is connected, which decisions are quantified, and which KPIs close the loop with field operations. Advancing this together with a governance design that includes data contracts, audits, and exception approvals is the condition for sustained results.

Chapter 13　Conclusion ― Inventory Should Be “Optimized,” Not Simply “Reduced”

The purpose of data-driven SCM is not to cut inventory indiscriminately. Under the premise of demand uncertainty, it is to place inventory “only where necessary, only in the amount necessary,” suppressing both stockouts and excess simultaneously.

At the core of achieving this is PSI, which links production, sales, and inventory together on a single sheet to anticipate supply-demand gaps, supported by advanced demand forecasting and inventory-optimization methods such as safety stock, MEIO, and ABC/XYZ analysis. As the safety-stock formula shows, the most effective levers for inventory reduction are “improving forecast accuracy” and “shortening and stabilizing lead time.” These, in turn, only function once data readiness and cross-departmental consensus processes are in place — even before any tool is introduced.

✱ Final message: Inventory is neither friend nor foe — it is a “mirror reflecting the uncertainty of supply and demand.” Using data to shrink that uncertainty, and absorbing what variability remains through optimal placement — this is the essence of the supply chain optimization that data-driven SCM aims for.

(This report was prepared based on general methodologies in production management and SCM — PSI/production-sales-inventory planning, safety stock theory, MEIO, S&OP/IBP, SCOR, ABC/XYZ analysis, and others — as well as practitioner literature from 2025–2026. The figures presented are representative estimates; actual results will vary by industry and product characteristics.)