Be clear about what forecasting needs
- History: at least two full cycles, ideally three years
- Stability: the patterns that held must still hold
- Clean data: consistent categories, no unexplained gaps
- Enough volume: forecasting five orders a month is not forecasting
If two of those are missing, the honest answer is that no technique will produce a forecast you should act on.
Language models are the wrong tool here
For numerical forecasting, established statistical methods outperform language models substantially and cost a fraction. The right AI answer is often a well-fitted seasonal model, not a chat interface.
Where language models do help is explaining a forecast, surfacing the drivers in plain English, and incorporating unstructured signals like sales notes.
Where it works well in practice
| Application | Fit | Requirement |
|---|---|---|
| Seasonal demand for established lines | Strong | Three years of history |
| Cash flow from invoice and payment history | Strong | Consistent payment behaviour |
| Staffing against booked work | Strong | Reliable booking data |
| New product demand | Weak | No history to learn from |
| Anything after a business model change | Weak | Past patterns no longer apply |
Present ranges, not numbers
A single forecast figure invites false confidence. A range with a stated confidence level invites planning, which is what a forecast is for.
We also always show the same model's accuracy on the last six months, so the reader can calibrate their trust rather than assume it.
The realistic value
Better than intuition for stable, seasonal, high-volume patterns. Not better for novel situations, which is precisely when people most want a forecast.
Used for reordering, staffing and cash planning, it saves real money. Used for strategy, it produces spurious precision.