Companies must decide where to produce, store and ship (strategic network design) and every month reconcile demand, supply and finance (S&OP / integrated business planning). Both are optimisation problems under uncertainty. Machine-learning forecasts feed mixed-integer models that compare scenarios, and LLM copilots now explain trade-offs and draft the planning narrative. Typical benefit: lower total landed cost, better service level and faster, more fact-based consensus decisions.
Problem & value
- Today: network design is run as a one-off consulting project every few years with a specialist tool; S&OP runs on spreadsheets, ERP exports and slide decks. Demand, supply and finance numbers disagree, scenarios take days, and the plan is stale when it is agreed.
- AI value: (1) ML demand and lead-time forecasts with uncertainty ranges; (2) solver-based optimisation of network flows, sourcing and capacity; (3) fast what-if scenarios for tariffs, site closures, supplier loss or CO2 targets; (4) LLM assistants that summarise plan deltas, answer questions on the model and draft meeting material.
- Why now: tariff shifts, geopolitical disruption and EU supply chain due diligence and reporting duties (CSDDD, CSRD; scope and timelines were narrowed by the 2025-26 Omnibus simplification, so check current status) make network resilience and emissions explicit planning criteria [5].
- Vendors publish case studies but few independent, comparable benefit figures; measure your own baseline before trusting any promised percentage.
Variants
- Strategic network design: facility location, make-vs-buy, sourcing and capacity decisions over a 3-10 year horizon.
- Tactical S&OP / IBP: monthly demand review, supply review, financial reconciliation, executive decision meeting; rolling 12-24 months.
- Resilience and carbon variant: multi-objective design (cost, service, emissions, dual sourcing).
- SMB / Mittelstand: a single plant with a few warehouses; solved in a spreadsheet-plus-open-solver model or an ERP/planning add-on rather than a full IBP suite.
- Retail / CPG: distribution-centre network and store replenishment flows, linked to OPT-005.
How it works
- data (ERP, demand history, cost/rate tables, BOM, capacity)
- forecast (ML time series, probabilistic) + lead-time/risk estimates
- network/supply model (digital twin: nodes, arcs, constraints, costs)
- optimisation (MIP/LP, multi-objective) + simulation of uncertainty
- scenario library (baseline, tariff, outage, demand shock, CO2 cap)
- S&OP meeting pack (LLM-drafted explanations, KPI deltas)
- human decision
- plan released to execution (APS, WMS, TMS)
- Forecasting: gradient boosting or foundation time-series models produce quantile forecasts; see PRED-001.
- Optimisation core: facility-location and multi-commodity flow models solved by MIP solvers. The solver, not the LLM, produces the numbers.
- LLM role: translate a business question into a scenario definition (see OPT-018), explain binding constraints and plan differences in plain language, draft minutes. Outputs are checked against the solver result.
- Human-in-the-loop: demand planners adjust forecasts, the supply review accepts or changes capacity, executives approve network changes. Plan release is always a human decision.
Data preparation
Needed: demand history, forecasts, site and warehouse capacities, transport lanes and tariffs, production capacities, bills of materials, costs and service targets. These sit in ERP, planning tools, carrier invoices, and many Excel files maintained by individual planners. Customer and supplier master data is spread across systems.
Typical gaps: different definitions of the same figure across departments, outdated lane costs, missing capacity data, customer and site IDs that do not match across systems, and S&OP meeting documents as scattered slide decks.
Preparation steps: - Map required data against sources and gaps (PREP-002). - Extract ERP and transport data via exports or connectors (PREP-005). - Reconcile master data and build one agreed KPI glossary (PREP-012). - Convert planner Excel files into tidy tables and track them as data sources (PREP-003, PREP-011). - Name data owners per source (PREP-004).
Rough share of effort: 50-60 percent, mainly alignment of definitions rather than technology. Check readiness with GUIDE-003.
Implementation paths
Which path for whom
- Enterprise: Path B. Network design and S&OP are strategic and company-specific, so scenario models and solvers on own ERP and planning data via APIs beat generic tools.
- SMB: Path A. Few SMBs need a bespoke model; established planning and S&OP SaaS with EU hosting covers the standard process at affordable prices.
- Developer: Path C. Open-source optimisation (e.g. OR-Tools, HiGHS) and LLM-based narrative generation run on HW1 for embedding in own tools.
Path A - Buy (recommended)
- Network design tools: Coupa Supply Chain Design (former LLamasoft), AIMMS SC Navigator (EU, Netherlands), anyLogistix (AnyLogic, simulation-based), Optilogic Cosmic Frog, SAP IBP with network design, Kinaxis Maestro, o9 Solutions, Blue Yonder, OMP (Belgium, EU), Anaplan for finance-linked S&OP, and Lokad for probabilistic planning (France, EU).
- Choose A when: multiple sites and SKUs, an existing ERP (SAP, Microsoft, Oracle) and a team that needs vendor-supported planning workflows. Check EU data hosting, ERP connectors, scenario versioning and how the vendor's LLM assistant handles your data.
Path B - Build on cloud APIs
- Solvers as a service: Gurobi (cloud/Gurobi Optimizer; Gurobi AI Modeling assistant), FICO Xpress, Google OR-Tools, Hexaly (France, EU), Timefold (Belgium, EU) for planning problems.
- Forecasting: Azure/AWS/Google forecasting services or Nixtla TimeGPT; plus a frontier LLM (Claude, Gemini, GPT, Mistral in EU regions) for narrative and scenario translation.
- Orchestration: data in a warehouse (Snowflake, BigQuery, Databricks), model in Python (Pyomo or similar), results back to BI and the planning tool.
- Choose B when: you need a custom model that off-the-shelf tools cannot represent.
Path C - Self-host (open source)
- Solvers: HiGHS (MIT licence), Google OR-Tools (CP-SAT), SCIP, CBC; modelling in Pyomo, PuLP or JuMP. Simulation: SimPy.
- Forecasting: Nixtla StatsForecast / NeuralForecast, Darts, Chronos or other open time-series foundation models.
- LLM layer: an open-weight model (e.g. Qwen, Llama, Mistral) served locally for narrative and model-explanation tasks; the LLM only writes text and model code that is reviewed.
- Hardware: HW1 suffices for solvers and forecasting (CPU, plenty of RAM); HW2/HW3 only for a local LLM. Large MIPs may need commercial solver licences for performance.
Path D - Custom training
Fine-tuning is rarely needed. Worth it for: proprietary lead-time or disruption-risk models, learned surrogate models that speed up large simulations, or reinforcement-learning policies for multi-echelon decisions. Needs multi-year history of orders, shipments and delays.
State of the art (2026)
- Predict-then-optimise is standard; research increasingly pursues decision-focused learning, where forecasts are trained for decision quality rather than accuracy alone.
- Probabilistic and stochastic planning (quantile forecasts, scenario trees, robust optimisation) is replacing single-number plans in leading suites.
- LLM copilots in planning suites (Kinaxis, o9, SAP, Blue Yonder, Coupa) answer questions on plans, build scenarios from natural language and summarise exceptions; agentic features that propose and execute planning actions are in early production.
- LLM-to-solver modelling (OptiMUS, Gurobi AI Modeling, OR-LLM research) helps analysts formulate models but benchmark accuracy on complex industrial models is still limited [4].
- Digital twins of the end-to-end network allow continuous re-optimisation rather than one-off studies.
- Limits and failure modes: garbage-in data (outdated cost and lead-time tables); models that ignore real-world constraints (labour, customs, contracts); over-trust in a single optimum; hallucinated model code or constraints from LLMs; forecast bias propagated into capacity decisions; S&OP failing for organisational, not technical, reasons (no shared targets).
Shared resources & combinations
- Same stack (optimization-solver + time-series): OPT-001 routing, OPT-002 production scheduling, OPT-005 inventory, OPT-012 packing, OPT-037 port operations.
- Upstream: PRED-001 demand forecasting, PRED-008 commodity price forecasting, ANOM-011 supplier and supply chain risk monitoring.
- LLM layer: OPT-018 natural-language optimisation modelling and decision copilot is the natural front end for scenario work.
- Data pipeline: one ERP-to-warehouse pipeline serves planning, inventory and finance reporting.
Effort, cost & KPIs
- Path A: 3-9 months for a suite rollout; licence typically per user/module; data cleanup is the main cost. Network design studies can be run in weeks with a specialist.
- Path B: 3-6 months for a first production model; solver licences and cloud compute.
- Path C: 3-6 months; mostly engineering and modelling time; low licence cost.
- KPIs: total landed cost, service level (OTIF), forecast accuracy and bias (WAPE), inventory turns, plan stability, S&OP cycle time, scenario turnaround time, CO2e per unit shipped.
People & change
Involve management as sponsor, since network decisions change sites and contracts, and department leads from sales, production, purchasing and logistics, because S&OP is only useful if all functions accept one set of numbers. IT and the data owners deliver data; key suppliers and logistics partners may need to confirm capacity assumptions. Legal and compliance review contract and customs impacts.
Works council: usually no, as the data is aggregated and not person-related. §90(1) no. 3 BetrVG still calls for timely information when planning processes change through AI, and site closures or relocations from a network study can trigger separate co-determination rights, so involve it early if scenarios affect jobs. Public-sector equivalents involve the staff council.
Personal data is normally absent; inform the DPO anyway. AI literacy: train planners to question scenario assumptions and read model outputs critically (GUIDE-006). Pilot with one product family or region and compare with the established plan (GUIDE-009). Typical concern: "the model decides over sites". Make clear that scenarios are decision support, assumptions are documented, and management decides. See GUIDE-008 and GUIDE-001.
Risks, regulation & compliance
- EU AI Act: minimal risk; no Annex III category for commercial supply chain planning. Exception: if the planning system manages safety components of critical infrastructure (energy, water, traffic), classification must be reviewed. AI literacy (Art. 4) applies [3].
- GDPR: mostly non-personal data; customer addresses for delivery and employee/driver data in capacity models may be personal data; aggregate where possible.
- Sector and trade rules: EU supply chain due diligence (CSDDD), CSRD emissions reporting, German LkSG, sanctions and export-control screening of new network nodes, customs and tariff changes [5].
- Competition law: sharing planning data between competitors (shared S&OP platforms) needs review.
- Risks and mitigation: wrong cost data (version and audit inputs), single-scenario bias (always run stress tests), LLM-generated model errors (verify against solver output), vendor lock-in (export models and data), cyber exposure of planning data (access control).
Getting started
- Define the decision (new site, sourcing shift, S&OP redesign) and the 3-5 KPIs that judge it.
- Collect a clean baseline dataset: flows, costs, capacities, lead times; reconcile it with finance.
- Build and calibrate a baseline model that reproduces last year's actual costs within a few percent.
- Run 5-10 scenarios (tariff, disruption, demand, CO2 cap) and review them with operations and finance.
- Institutionalise: monthly S&OP cycle with forecast-accuracy tracking; add an LLM assistant for reporting.
Sources
- HiGHS open-source solver — https://highs.dev/
- Google OR-Tools documentation — https://developers.google.com/optimization
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- OptiMUS: scalable optimization modeling with LLMs (Ahmaditeshnizi et al.) — https://arxiv.org/abs/2402.10172
- European Commission — Corporate sustainability due diligence — https://commission.europa.eu/business-economy-euro/doing-business-eu/sustainability-due-diligence-responsible-business/corporate-sustainability-due-diligence_en
- Pyomo optimization modeling — https://www.pyomo.org/
- Nixtla StatsForecast — https://github.com/Nixtla/statsforecast
- Timefold planning solver — https://timefold.ai/
- Gurobi AI Modeling — https://www.gurobi.com/