Demand forecasting with AI: what it delivers and what it costs
AI demand forecasting earns its place where Excel and your ERP's forecast module only extend your sales history. A model that also weighs promotions, weather and holidays forecasts each item more sharply and shows how certain each forecast is.
For the owner who wants to know what it returns and what it costs, and for the operations or IT manager who has to run it afterwards: what AI adds to Excel and your ERP, the data you need, how to measure accuracy, and who does what alongside your ICT partner.
Last updated: 29 October 2026
A better demand forecast pays back on two sides: you miss fewer sales and you waste less stock, capacity and staff time. For a footwear manufacturer with €235M in turnover, Crux Digits built a forecasting and planning layer on Prophet and XGBoost; within a year revenue grew 8% and production capacity 7% (read the case). A first test on your own data costs €2,500, a working forecast for one product group €20,000 (all prices).
What a better demand forecast delivers
A demand forecast is not a report. It is the number purchasing, production and rostering base their decisions on. When that number runs high, cash sits in slow stock. When it runs low, you turn customers away on exactly the items they want, or pay overtime to catch up.
What you recover depends on where the error lands today: less stock at the same service level, fewer changeovers on the production line, or rosters that match the workload. At the footwear manufacturer in our demand forecasting case, most of the capacity gain came from mould changes that now lined up with real demand instead of reacting to surprises. Those are the figures of one engagement, not a promise for yours. That is why we first measure how well you forecast today, so the gain becomes a number on your own data.
What AI adds to Excel and your ERP's forecast module
| Approach | What it does | Where it stops |
|---|---|---|
| Excel | FORECAST.ETS uses the AAA version of exponential smoothing and detects seasonality automatically (Microsoft). | One series at a time, no promotions or weather, and upkeep sits with one planner. |
| Business Central | The Sales and Inventory Forecast extension predicts sales with Azure AI from your sales history and suggests replenishment (Microsoft Learn). | One aggregated forecast for all locations and variants; no prediction when data is too thin or too variable. |
| Odoo | The Master Production Schedule sets forecasted demand against stock and production (Odoo). | The user enters the forecasted demand; the package does not calculate it. |
| SAP IBP | Demand sensing creates daily forecasts for the next 4–8 weeks, starting from the existing demand plan (SAP Learning). | It sharpens the short term and does not replace demand planning itself. |
| Your own AI model | Learns from history plus promotions, prices, weather and holidays, per item and per location, and is tested against your current method. | Needs clean data and someone to review the output. |
The data you need
- Sales history at the level you plan. Per item, per location and per week or day. Prophet, one of the models we use, works best with strong seasonal effects and several seasons of history (CRAN); two years is a reasonable start.
- Stock-out days. An item that was out of stock sold nothing, but the demand was there. Without that flag a model learns the wrong lesson.
- Promotions and prices. When, which item, what discount, through which channel. This is often scattered across the ERP, the webshop and a marketing calendar.
- Weather. In the Netherlands the KNMI publishes daily data per station, including temperature, sunshine and precipitation (KNMI).
- Holidays and school breaks. Prophet models public holiday effects out of the box; school holidays and local events you add yourself.
- One condition. A driver only helps if you know its future value when you forecast. A weather forecast reaches a few days ahead, a promotion calendar months.
MAPE and bias: how to tell whether a forecast is any good
MAPE is the mean absolute percentage error. If you sold 100 units and the model forecast 90, the error is 10%; averaged over all weeks, that is the MAPE. It is easy to explain, with one weakness: when actual sales are zero it is undefined, and for items that rarely sell it swings wildly (Hyndman and Athanasopoulos). For slow movers, measure total absolute error divided by total sales instead.
Bias is the average error with its sign kept: do you forecast too high or too low, consistently? A model with a tidy MAPE and a bias of a few percent upwards quietly builds stock every week. Steer on both.
The number that matters most is the difference against your current method, measured over the same periods. That is why we first backtest a model on held-out history, as in the case, before it touches a decision.
From forecast to production, inventory and staffing
- Production planning. The forecast sets run sizes and sequence, so changeovers line up with real demand. At the footwear manufacturer this produced most of the capacity gain.
- Inventory and purchasing. Reorder points and safety stock follow from the forecast and from its uncertainty. An item with a wide range around its forecast gets more buffer than a stable fast mover.
- Staffing. A forecast per week and per site or shift becomes a roster that matches the workload, with fewer last-minute temps.
- The planner keeps the final say. The system proposes and shows why; the planner adjusts or overrides, and that correction becomes data for the next round. The model supports the judgement of the people who know the floor rather than replacing it.
Forecasting software or a custom model on your own data
| Your situation | Sensible choice | Why |
|---|---|---|
| Few items, short history, one planner | Excel with fixed rules | A model cannot learn what is not there yet. Start with clean data and one way of measuring. |
| You run Business Central online and demand mostly follows history | The built-in extension first | It comes with your subscription. Measure its MAPE and see where it falls short. |
| A large planning team in SAP with an S&OP process | Your ERP vendor's planning software | The process, roles and integration already exist; custom work fills a gap at most. |
| Promotions, weather or locations drive demand, or production has hard constraints | A custom machine learning model on your own data | Packaged tools do not weigh those drivers, or not per location; your own model does, and you own the code and the model. |
Who does what alongside your ICT partner
| Party | Responsible for | Not for |
|---|---|---|
| Crux Digits | Data pipeline, model, backtest against your current method, forecasts as proposals in your ERP or planning tool, and handover of code and test set. | Your network, workplaces, hosting or helpdesk. |
| Your ICT partner | Identity and access, the environment the model runs in, monitoring and backup under your existing agreements. | The content of the forecast. |
| Your IT or operations manager | Ownership, priorities, the change through your change process and the link with the ERP partner. | Building the model. |
| Your ERP partner | Reviews the integration for the APIs and permissions used, and makes any change inside the package itself. | The model and the data pipeline. |
| Planners and buyers | Reviewing, adjusting and releasing the proposals. | Maintaining the pipeline. |
What AI demand forecasting costs and how long it takes
| Step | What you get | Duration | Price excl. VAT |
|---|---|---|---|
| AI Audit & Strategy | Your sales, stock and promotion data reviewed, the MAPE and bias of your current method measured, and one scoped first build at a fixed price. | 1–2 weeks | €2,500 |
| Production-ready MVP | A forecast for one product group, site or line on your own data, backtested against your current method and run beside the planner for a few weeks. | 4–6 weeks | €20,000 |
| Production | All relevant groups, integrated with your ERP or WMS, with monitoring, documentation and handover to your IT organisation and ICT partner. | 8–12 weeks | from €50,000 |
| After delivery | Your IT organisation runs the model. If a small team would rather not, managed AI support covers monitoring, model updates and up to 4 hours of improvements a month. | Ongoing | €500 a month |
Four examples by sector
- Wholesale (example). A forecast per item and customer group feeds reorder points in the ERP; buyers see a proposal with its uncertainty. More in what AI delivers for a wholesaler.
- Food production (example). Short shelf life makes every error expensive. Weather and holidays drive daily demand for fresh and chilled lines; the forecast sets the production plan per line, and the planner sees expected sales per product for the coming days.
- Energy (example). A company forecasts its own consumption per quarter hour against its contracted capacity and moves batches to quiet hours. See grid congestion in industry.
- Retail (example). A forecast per store and per week, with promotions and holidays, feeds both store replenishment and the staff roster.
Frequently asked questions
What is demand forecasting?
How does AI demand forecasting work?
How accurate is AI demand forecasting?
What does custom demand forecasting software cost?
How much historical data do you need?
Do we need to replace our ERP or planning system?
Want to know what a better forecast is worth to you?
A free 30-minute call: you tell us how you forecast today and what data you have, we tell you whether an audit is worth it and what the first step costs.
Book a free consultation →