Case Study ／ Manufacturing ／ Multi-Agent AI

GenAI Supply-Chain Optimisation for a Global Manufacturer — 20 Sites x 3 Collaborating Agents Triple Decision Speed

For a global manufacturer with sites in 20+ countries, we built a multi-agent platform where three AI agents (demand forecasting, inventory optimisation, production planning) collaborate. By decomposing the Bull Whip effect into four structural drivers, we tripled decision-making speed even under post-COVID market volatility — a 10-month engagement.

Duration

10 months

Globalsites

20+

AI Agent

3unitscollaboration

Decision Speed

3×↑

Inventory turnover

+32%

Challenge

Demand-Forecast Accuracy Decay — Four Structural Shifts Statistical Models Cannot Capture

Post-pandemic, traditional statistical demand forecasting (MAPE 22%) could no longer capture market volatility, and lost opportunities from excess inventory and stock-outs reached tens of billions of yen annually. Decision-making had become tacit and reliant on veteran intuition — a structural fragility that had to be dismantled.

Global Supply-Network Complexity

Across a network spanning 20+ countries, hundreds of suppliers and thousands of components, there was no way to grasp supply/demand balance at each site in real time, and decisions consistently lagged behind events.

Demand-Forecast Accuracy Decay (MAPE 22%)

Post-COVID volatility sharply eroded the accuracy of traditional statistical demand forecasts. MAPE (Mean Absolute Percentage Error) hit 22%, and degraded plan quality fed straight through to bottom-line impact.

Losses from Excess Inventory and Stock-outs

Chronic excess inventory and stock-outs traced back to forecast inaccuracy created lost opportunities and cash-flow drag worth tens of billions of yen a-year — a top-of-agenda issue for both CFO and COO.

The Limits of Tacit Decision-Making

Decision-making had grown dependent on the instincts and experience of veterans. Retirements and rotations made know-how transfer difficult, exposing a structural fragility in organisational capability.

Bull Whip Diagnostic

The Bull Whip Effect Arises from Four Structural Drivers — A Single Agent Cannot Contain It

The Bull Whip effect — demand variance amplifying upstream — cannot be tamed by improving a single forecast model. We decomposed it into four structures (information delay, inventory over-reaction, price volatility, batch ordering) and designed dedicated agent Response patterns for each.

Structural Driver

Amplification Mechanism

Conventional Response

Multi-Agent Response

Containment Effect

Information Delay

Time lag as downstream demand signals propagate upstream

Improve statistical forecast accuracy

Demand-forecast agent integrates external market data

Lag compressed

Inventory Over-reaction

Buffer stock layered on as defence against demand swings

Fixed safety-stock ratio

Inventory-Optimisation Agent applies site-specific dynamic allocation

Over-reaction contained

Price Volatility

Promotions create demand pulses

Manual post-promo adjustment

Demand-Forecast Agent learns price events

Pulses smoothed

Batch Ordering

Order lot size distorts demand

EOQ-based fixed lot

Production-Planning Agent produces dynamic small-lot plans

Distortion minimised

Multi-Agent Architecture

Multi-Agentplatform — 3 Agent × 4 layers, structurally integrated

demand forecasting / Inventory optimisation / production planning Three agents — Data Layer / Agent Layer / Orchestration Layer / Observability-Layer operate on a four-layer shared platform. Each agent retains independence, Orchestration Layer Executive impact via a bespoke design that collaborates on the executive-impact axis.

Tier 1

Data Layer

External market data / Internal operations data / log unification into a single data platform.

20 sites+ operations data (Production / Inventory / Shipment / Order) and External market data (FX / Raw-materials index / Logistics cost) into one platform. Each agent learns from a shared "single source of truth" as a learning platform.

Integration scope20 sites / 12 major external sources

Refresh cadence Daily (operations) ／ Weekly (market)

Tier 2

Agent Layer (3 units)

Agent Layer

demand forecasting / Inventory optimisation / production planning 3 Agent each owns its specialism, and proposes decisions independently.

Each agent has its own machine-learning model and operations logic. Demand-Forecast Agentintegrates an LLM to factor in external drivers, Inventory-Optimisation Agentuses reinforcement learning for dynamic allocation, Production-Planning Agentuses constrained optimisation to generate small-lot plans.

demand forecastingLLM ＋ time-series model

Inventory optimisationreinforcement learning

production planningconstrained optimisation

IndependenceEach agent is loosely coupled

Tier 3

Orchestration Layer

Three-agent proposals proposals integrated on the executive-impact axis, and presented to the executive committee as a decision package.

Each agent proposal , evaluated on four axes — profit / cash flow / service level / reputation, and trade-offs visualised. Final decisions are made by humans, while agents are restricted to structuring the option set.

Integration axisProfit ／ CF ／ SL ／ Reputation

Final decisionHuman (executive／Business Owner)

Tier 4

Observability layer

Three-agent forecasts / proposals / adoption monitored continuously, Contribution to executive KPIs visualised.

Agent Forecast error / Adoption rate / executive KPI logged as time series, Monthlyreview meetings drive structural improvement. Integrated with the second line (risk management), to assure agent decision quality at the organisation level.

Monitored itemsForecast error ／ Adoption rate ／ KPI

Control integrationSecond line (risk management)

Engagement Timeline

3 Phase × 10 months — Diagnosis / Design / Phased rollout end-to-end pipeline

Not a one-off PoC, we designed an end-to-end path from diagnosis to phased rollout. Each phase has an explicit executive decision gate, and Phase 3 confirms effectiveness with five lead sites before drafting the full rollout plan .

Phase 1

2 months

Supply-Chain Diagnosis

20 sites+current processes thoroughly visualised. structural decomposition of the Bull Whip effect, data-quality evaluation, 40+ stakeholder interviews run in parallel. We separated problems AI should solve from those it cannot, up front.

Key Deliverables

SCM end-to-end visualisation map (20 sites)
structural decomposition of the Bull Whip effectreport
Data-pipeline design document
40 interview synthesis (tacit-knowledge codification)

Phase 2

4 months

Multi-Agent Design

3 Agent × 4 layers we designed the reference architecture. Each agent's responsibilities / Orchestration Layer Integration axis / Observability-Layer KPI taxonomy , executive level / frontline / IT were co-designed across these three .

Key Deliverables

3 Agentdetailed design document
Orchestrationspecification (4 axisIntegration logic)
KPI Dashboard design
Tier-by-tier data requirements

Phase 3

4 months

Phased Rollout / Effectiveness Validation

5 siteslead PoC executed, verifying agent-collaboration effectiveness. Monthly KPI reviews drive continuous improvement, establishing the 20-site full rollout plan. Approval received from the executive committee, the following fiscal year secured the full-site rollout budget .

Key Deliverables

PoC validation report (5 sites)
Monthly KPI-review taxonomy
20 sitesFull rollout plan
executiveexecutive-committee-approved edition / investment plan

ROI Projection

Three Domains x 3 Years — Translating Phased Rollout into ROI in Executive Language

5 sites PoC / 10 lead sites / all 20+ sites phased rollout , presented with expected-effect ranges for each of the three domains. Year 1 confirms effectiveness, with leverage applied in Year 2 and Year 3.

Domain

Year 1 (5 sites PoC)

Year 2 (10 lead sites)

Year 3 (all 20+ sites)

Key KPIs

demand forecasting

MAPE improvement

22% → 16% improvement

16% → 12% improvement

12% → 10% target

MAPE ／ Bias

Inventory optimisation

Turnover improvement

+15% improvement (PoC 5 sites)

+25% improvement (10 lead sites)

+32% improvement (all sites)

Turnover ratio / Stock-out rate

production planning

Decision Speed

Monthly → Bi-weekly (compressed)

Bi-weekly → Weekly (compressed)

Weekly / all sitessynchronised

Cycle time / Plan accuracy

Figures This engagement executivecommitteepresented as of expectedrange. Actual values will be re-evaluated from Year 2 onwards in line with phased-rollout results.

Results

Executive impact / operationseffect / Governance / rollout plan — 4 axisValidation

At project completion, Decision speed tripled, demand-forecast accuracy improved, Inventory turnover lifted, the 20-site full rollout plan was simultaneously approved by the executive committee. These are downstream outcomes derived from the multi-agent collaboration design.

Executive impact

3×

Decision Speed 3x compressed

Monthlycycleproduction-planning adjustment was Weeklycycle to . market volatility Response speed to structural improvement, and the resolution of executive decisions rose.

operationseffect

+32% ↑

Inventory turnover structuralimprovement

3 year plan target as of Inventory turnover expected to improve by 32%. tens of billions of yen annuallyscale of inventory excess / stock-outs lost opportunity structural compression is achieved by design.

Governance

Second lineLinkage

Agent decision quality assured at the organisation level

The observability layer is connected to the second line (risk management). Agent forecasts, adoption rates and KPI contribution are kept under continuous oversight, so the firm escapes personality-driven decision-making and establishes organisational capability.

rollout plan

Approved

20-site full rollout plan — the following fiscal yearBudget secured

5 sites PoC Outcome Based on 20-site full rollout plan Formulation, executivecommittee Approval. the following fiscal year secured the full-site rollout budget .

Implementation Principles

Five Implementation Principles — Reproducible Structural Judgements for Global-Manufacturing SCM AI

From the lessons distilled in this engagement we extracted five structures that make global-manufacturing SCM AI implementation reproducible. They are not item-by-item tuning, but a baseline for organisation-level SCM AI design decisions.

Bull Whip cannot be contained by a single agent

Bull Whip Requires Multi-Agent Design

Each of the four structures — information, inventory, price and batch — requires its own dedicated response pattern. Improving demand-forecast accuracy alone delivers only local optima; only a three-agent collaboration design can structurally contain the Bull Whip effect.

"ForecastAccuracy"rather than"Forecast error to Resilience"is what we design

Design for Resilience, Not Accuracy

Rather than driving demand-forecast MAPE to extremes, when forecasts miss Inventory optimisation and production planning Dynamic correction organisationcapability Executive impact is larger. Designing for resilience is the core of SCM AI.

Site specificity "data normalisation" flatten absent

Preserve Site-Specific Logic

20 sites operational difference uniform normalisation and , erase each site's competitive advantage. sitesspecific Parameter Agent Learning Design and , GlobalIntegration and frontlinespecificity Reconcile Realise .

Veteran tacit knowledge AgentinitialParameter to translation

Encode Tacit Knowledge as Priors

40 Veteraninterview gleanedTacit knowledge , Agent initialParameter / Constraints conditions / ExceptionRule to translation. Cold start (cold start) Duration compressed, and knowledge succession is embedded in the technical design at organisation level..

KPI dashboards are 90% 'alert design'

Dashboards Are 90% Alert Design

A dashboard that merely visualises hundreds of metrics is unusable for executive decisions. Decide upfront "who decides what when a threshold is crossed", and rebuild the dashboard as an alert system that drives those decisions.

GenAI Supply-Chain Optimisation for a Global Manufacturer — 20 Sites x 3 Collaborating Agents Triple Decision Speed

Demand-Forecast Accuracy Decay — Four Structural Shifts Statistical Models Cannot Capture

Global Supply-Network Complexity

Demand-Forecast Accuracy Decay (MAPE 22%)

Losses from Excess Inventory and Stock-outs

The Limits of Tacit Decision-Making

The Bull Whip Effect Arises from Four Structural Drivers — A Single Agent Cannot Contain It

Multi-Agentplatform — 3 Agent × 4 layers, structurally integrated

3 Phase × 10 months — Diagnosis / Design / Phased rollout end-to-end pipeline

Supply-Chain Diagnosis

Key Deliverables

Multi-Agent Design

Key Deliverables

Phased Rollout / Effectiveness Validation

Key Deliverables

Three Domains x 3 Years — Translating Phased Rollout into ROI in Executive Language

Executive impact / operationseffect / Governance / rollout plan — 4 axisValidation

Decision Speed 3x compressed

Inventory turnover structuralimprovement

Agent decision quality assured at the organisation level

20-site full rollout plan — the following fiscal yearBudget secured

Five Implementation Principles — Reproducible Structural Judgements for Global-Manufacturing SCM AI

Bull Whip cannot be contained by a single agent

"ForecastAccuracy"rather than"Forecast error to Resilience"is what we design

Site specificity "data normalisation" flatten absent

Veteran tacit knowledge AgentinitialParameter to translation

KPI dashboards are 90% 'alert design'

This engagement Linkage DX Strategy Service

your firm Supply-chain AI , structuringFramework Design / support..

GenAI Supply-Chain Optimisation for a Global Manufacturer — 20 Sites x 3 Collaborating Agents Triple Decision Speed

Demand-Forecast Accuracy Decay — Four Structural Shifts Statistical Models Cannot Capture

Global Supply-Network Complexity

Demand-Forecast Accuracy Decay (MAPE 22%)

Losses from Excess Inventory and Stock-outs

The Limits of Tacit Decision-Making

The Bull Whip Effect Arises from Four Structural Drivers — A Single Agent Cannot Contain It

Multi-Agentplatform — 3 Agent × 4 layers, structurally integrated

3 Phase × 10 months — Diagnosis / Design / Phased rollout end-to-end pipeline

Supply-Chain Diagnosis

Key Deliverables

Multi-Agent Design

Key Deliverables

Phased Rollout / Effectiveness Validation

Key Deliverables

Three Domains x 3 Years — Translating Phased Rollout into ROI in Executive Language

Executive impact / operationseffect / Governance / rollout plan — 4 axisValidation

Decision Speed 3x compressed

Inventory turnover structuralimprovement

Agent decision quality assured at the organisation level

20-site full rollout plan — the following fiscal yearBudget secured

Five Implementation Principles — Reproducible Structural Judgements for Global-Manufacturing SCM AI

Bull Whip cannot be contained by a single agent

"ForecastAccuracy"rather than"Forecast error to Resilience"is what we design

Site specificity "data normalisation" flatten absent

Veteran tacit knowledge AgentinitialParameter to translation

KPI dashboards are 90% 'alert design'

This engagement Linkage DX Strategy Service

Related Case Studies

your firm Supply-chain AI , structuringFramework Design / support..