Supply chain AI, like logistics AI, has a mature OR core and an emerging LLM layer. The S&OP (sales and operations planning) cycle — monthly or weekly planning meetings where commercial, operations, finance, and supply meet to agree on the plan — is where most large companies actually make supply chain decisions. This is the workflow where LLMs add genuine value without disrupting the quantitative foundation.
The S&OP cycle as it actually runs
Demand planners produce a statistical forecast. Commercial teams adjust for known events (promotions, launches, lost customers). Supply planners assess feasibility against capacity. Exceptions are debated. A consensus plan emerges. This is a monthly or weekly cycle at most companies.
Where the cycle breaks: (1) statistical forecasts don't catch qualitative signals (news, weather, geopolitics), (2) commercial adjustments are opinion-heavy and hard to defend, (3) exceptions consume the meeting; strategic topics get squeezed, (4) the plan document — what was agreed, why — often doesn't exist in any usable form by week two.
Where LLMs add value in the cycle
Forecast overlay with qualitative signals
LLMs read external signals — trade news, weather forecasts, port congestion, supplier earnings releases — and flag implications for specific SKUs or regions. A planner reviewing the forecast sees not just the numbers but 'heavy-truck demand historically drops 8% in Q1 following tier-1 supplier earnings downgrades; two such downgrades occurred last week.' This is synthesis the planner couldn't feasibly do by reading 50 news items a day.
Supplier risk intelligence
Proactive monitoring of supplier financial health, operational disruptions, geopolitical exposure, compliance issues. LLMs aggregate signals and surface risks before they become incidents. This is the highest-ROI AI use case for companies with deep supply chains — catching a supplier financial risk one quarter early saves enormous response cost.
Meeting preparation
One-page summary before the S&OP meeting: this week's forecast changes, why, notable exceptions, questions to resolve. Takes a planner 2-3 hours to produce manually. LLM draft reviewed and edited in 20 minutes.
Plan documentation
Meeting transcription + action items + rationale capture. The plan as agreed is documented in a form that's actually findable next week, with the reasoning attached.
Disruption response
When something goes wrong — supplier failure, weather event, labor action — the response involves reading many sources, assessing exposure, coordinating with ops, sales, and logistics. LLMs cut the time-to-understand from hours to minutes.
What stays classical
Base statistical forecasting. ARIMA, exponential smoothing, gradient boosting — these work well, are interpretable, and are what CFOs trust for financial planning. LLMs may provide overlays but should not replace the base.
Inventory optimization. (s, S) policies, safety-stock calculations, base-stock models — specialized math. LLMs can explain the results; they do not compute the optimal policies.
Capacity planning. Linear programming and mixed-integer optimization. Solvers exist; use them.
Sales forecasting at the commercial level. The commercial team's judgment is a feature, not a bug — they know things the statistical forecast doesn't know. LLMs can help structure the adjustment (what assumptions, what support) but the judgment is human.
Data integration is where projects stall
ERP (SAP, Oracle) + planning system (SAP IBP, Kinaxis, o9, Blue Yonder) + demand planning + supplier systems. Getting data out of these for AI use is the 6-month engineering project most AI pilots underestimate. Start with use cases where the data is accessible; work outward. Our rule of thumb: no supply chain AI pilot lasts less than 4 months end-to-end; plan for 6-8.