Supply chain planning generates an enormous volume of unstructured text that traditional planning systems were never built to handle: planner notes explaining why a forecast was overridden, supplier emails buried in an inbox, exception comments scattered across spreadsheets, meeting notes from S&OP reviews that capture decisions nobody formally documents. That's precisely the kind of problem large language models are good at — and it's a very different use case from the demand-forecasting AI that usually dominates the supply chain AI conversation.
The unstructured data problem in S&OP
- ›Forecast override rationale is written in free text by planners and almost never analysed in aggregate — so patterns in why forecasts get overridden are invisible to the organisation
- ›Supplier communication about delays, quality issues, and capacity constraints lives in email threads that don't feed into planning systems until someone manually re-enters the information
- ›S&OP meeting decisions are captured in minutes or slide notes that are rarely searchable and almost never linked back to the plan they affected
- ›Exception handling notes — why a shipment was expedited, why a safety stock level was overridden — accumulate as institutional knowledge that leaves when the planner does
Where LLMs are genuinely useful today
- ›Summarising planner rationale across hundreds of SKUs into a digestible exception report for the S&OP review, instead of a planner manually compiling notes the night before
- ›Structuring inbound supplier emails into standardised fields — delay reason, revised date, affected order — that feed directly into planning systems rather than requiring manual re-entry
- ›Drafting first-pass what-if scenario narratives from planning system outputs, so planners start from a draft explanation rather than a blank page when presenting a plan change
- ›Answering natural-language questions against planning documentation and historical decisions — 'why did we override the forecast for this SKU last quarter' — through retrieval over the organisation's own planning history
- ›Translating between systems and formats — turning a supplier's non-standard Excel forecast submission into the structured format the planning system expects
The common thread across every use case that's actually working: the LLM is handling language — reading, structuring, summarising, translating — not doing the numerical forecasting itself. That distinction is the difference between a deployment that delivers value and one that quietly erodes trust in the plan.
Where they fall short
Generative AI is a poor substitute for statistical and machine learning forecasting methods, and using it that way is a common and costly mistake. LLMs are not trained to reason reliably about numerical time series, and asking one to 'predict next quarter's demand' from historical data will produce plausible-sounding numbers that are frequently wrong in ways that are hard to detect until the plan fails. The forecasting engine should stay a purpose-built statistical or ML model. The LLM's job is everything around that engine — the inputs it needs structured, and the outputs it needs explained.
A practical pattern: LLM as a translation layer
The most durable pattern I've seen for generative AI in supply chain planning is to treat it as a translation layer sitting between unstructured human communication and structured planning systems — not as a planning engine in its own right. Supplier emails go in unstructured, structured data comes out. Planner rationale goes in as free text, a synthesised exception summary comes out. The plan itself is still produced by the purpose-built forecasting and optimisation tools that have always done that job well; the LLM removes the manual translation work that used to sit between those tools and the humans who feed and interpret them.
The value of generative AI in supply chain planning isn't in replacing the planner or the forecasting model — it's in eliminating the hours planners spend translating between unstructured human communication and structured systems. Deploy it there first, and the ROI is immediate and measurable.