AI demand forecasting

The short answer

Demand forecasting is a statistics problem with decades of well tested methods behind it. A language model is not the right instrument, and using one because it is the current technology is how you get a worse forecast at a higher cost.

That does not mean AI has no place here. It means the forecast itself should come from a model built for time series, and the AI belongs around it: in the inputs, the explanation and the decision that follows.

What actually wins

Gradient boosted trees and classical time series methods, with good feature engineering, beat almost everything else on tabular demand data. They are cheap, fast, explainable and they have a well understood failure behaviour.

Say that plainly even when it is not the exciting answer. A vendor who proposes an LLM for a forecast either does not know the field or is selling what they have rather than what you need, which is a specific thing to check for.

Where AI genuinely adds value

Around the model, in three places. Bringing in unstructured signals that never reach the model today: supplier emails about delays, news about a competitor, weather advisories, event calendars.

Explaining the forecast in language the planner can act on, including which drivers changed since last week. And handling the exceptions, where a planner needs to ask why this SKU dropped and get a real answer instead of opening a spreadsheet.

Accuracy is the wrong target

A forecast is not an end product. It is an input to an ordering, staffing or production decision, and those have asymmetric costs: being short is usually far more expensive than being long, or the reverse, and never equally.

Optimise against the cost of the decision, not against a symmetric error metric. A forecast that is slightly less accurate and costs less to be wrong about is the better forecast, which is the measuring AI ROI argument in its purest form.

Data readiness decides feasibility

You need history at the granularity of the decision, and you need to know what was actually available rather than what sold: a stockout looks like zero demand and teaches the model to under order forever.

Promotions, price changes and holidays need to be recorded as events, not inferred from the sales curve. If that history does not exist, the first project is instrumentation, per data readiness.

The cold start problem

New products, new locations and new categories have no history, and they are frequently the ones the business most wants a forecast for.

Handle them by analogy and hierarchy rather than pretending the model can extrapolate: what did comparable launches do, what does the category do, what does the region do. State the uncertainty rather than producing a confident number nobody should trust.

Forecasts drift, loudly and quietly

Demand patterns change, and a model trained on last year's behaviour degrades without erroring. The retraining cadence should be a decision, not a thing that happens when somebody notices.

Monitor error by segment rather than in aggregate, because a stable overall number routinely hides one category falling apart. This is the same discipline as model drift, applied to a different model class.

Adoption is the real constraint

Planners override forecasts they do not understand and do not trust, and an overridden forecast has produced nothing regardless of its accuracy.

Track the override rate as a first class metric. Showing why the number moved, and letting the planner interrogate it in plain language, is usually worth more in practice than another point of accuracy.

How to sequence it

Establish a baseline with a simple model and measure honestly against what the business does today, which is often a spreadsheet that is better than people expect.

Then add the unstructured signals and the explanation layer, and only invest further where the decision cost justifies it. That order is common in AI for logistics work, and it is what makes production AI, end to end rather than a science project.