A demand-planning process that explains its own uncertainty.

01

Better predictions were not enough

The fictional retailer Willow Distribution planned replenishment across 36 product groups. Forecasts existed, but buyers often overrode them without a recorded reason. Promotions, supplier constraints, and product introductions were discussed in spreadsheets that never reached the evaluation pipeline.

The design objective was not to eliminate overrides. A buyer may know something the data does not. The objective was to make that information visible and compare decisions with the same baseline.

02

Use the right model for the right task

A forecasting pipeline generated demand estimates and prediction intervals from time-series inputs. Business rules applied minimum order quantities, lead times, and available storage capacity. A language model explained the resulting proposal using the structured inputs; it did not invent new demand numbers.

The buyer interface showed the forecast, the operational constraint, and the reason for any override. That created a shared decision record. Forecast evaluation stayed separate from the quality of the generated explanation.

03

Evaluate at the decision horizon

A forecast made one day before a shipment is not comparable to one made six weeks before a purchasing decision. We organized evaluation by lead-time horizon, product family, and promotion status. Stockout and waste measures were reviewed alongside forecast error.

The team also monitored when an explanation cited a factor that was not in the input. Those responses were rejected even if they sounded plausible. An explanation is useful only when it accurately represents the calculation and the available business context.

04

Illustrative business outcomes

The following invented results cover the same planning horizon and product groups. They illustrate why decision quality and traceability belong beside forecast accuracy.

05

A useful operating model for AI-assisted planning

The scenario shows a layered approach: predictive models for numerical estimates, deterministic rules for business constraints, and generative models for explanation. Each layer has a different test and a different owner.

That separation makes disagreements easier to resolve. A buyer can challenge an assumption without rejecting the entire system, and an engineer can investigate a model issue without rewriting the purchasing policy. The result is a planning process that can improve through evidence rather than persuasion.