Stage 4 · Deploy · into real workflows
Most forecasting in a New Zealand business is a spreadsheet with last year’s numbers and an experienced person’s feeling. The feeling is often right, and the spreadsheet cannot say when it is wrong. Predictive AI uses your own history to forecast demand, flag risk and optimise schedules, with the model’s confidence shown, so people know when to trust it.
What it does
Every business predicts: how much to order before Christmas, which machine will fail next, which customer will pay late, how many people to roster on Saturday. Most of those predictions are made by an experienced person looking at a spreadsheet, and most of the time they are close. Nobody knows which time, the person is on leave in January, and the spreadsheet cannot say how sure it is.
Predictive AI is a model trained on your own history. It learns the patterns that person has learned and some they have not: the sales pattern after a long weekend, the vibration reading before a bearing failure, the order size and payment history that mean a late invoice. It produces a number, a range around it and a confidence, so the reader knows when to trust it and when to look harder, and it is judged the way you would judge a person: against what actually happened, month after month.
Where it fits
How we build it
From history to a forecast people trust
Compare
| The spreadsheet you have | Forecasting built into your ERP | Be AI Predictive AI | |
|---|---|---|---|
| Uses your full history rather than last year | Rarely | Partly | Yes |
| Shows how confident it is | No | Rarely | Every prediction |
| Explains a prediction in words | The person can | No | Yes |
| Measured against what happened | Sometimes | Rarely | Monthly |
| Sensible when | Volumes are small and one person knows the business | Your patterns are simple and the ERP does it well | The decision is repeated, costly and made on a feeling |
Questions
Less than you fear, more than a quarter. For demand, two or three years of sales history at the level you want to forecast is a good start. For maintenance and risk, it is the number of past failures that matters, because a model learns from what went wrong.
Yes, and we do not build ones that cannot. Each prediction comes with the factors that drove it, in words a buyer, a workshop manager or a credit officer can read and challenge. For decisions about people, such as credit or claims, that is designed in from the start.
The model gets worse, and the monitoring says so. Every prediction is stored beside what actually happened, so a drop in accuracy shows within weeks. The model is retrained on the new pattern, and until then the confidence shown tells people to trust it less.
Then a person ordered stock on a forecast they chose to trust, which is what happens now, with less information. The design keeps the person in the decision, shows the confidence, and runs the model beside the current method for a season before anyone relies on it.
Related
Tell us the decision you make on a feeling every month and we will tell you whether your history can do better, and what it would take to prove it.
