Every time an SME runs out of a product right when demand peaks, or accumulates stock of something nobody wants, it is paying the price of poor demand forecasting. AI demand forecasting turns historical data, calendars and external signals into actionable predictions: how much you will sell next week, which SKUs will spike during a campaign and which ones you will need to clear. This article explains, without jargon, how it works, what data you need, which models exist and when it makes sense to deploy it in a company of 10 to 250 employees.
What is demand forecasting and why does AI make a difference?
Demand forecasting is the estimation of how many units of a product or service will be needed over a given time horizon. In its classic form it is done with spreadsheets, moving averages and the judgement of the purchasing manager. The problem is that this approach assumes the past repeats itself linearly, when in practice demand is driven by dozens of simultaneous variables: seasonality, promotions, competitor prices, weather, local events or social-media trends.
Artificial intelligence models — in particular those based on recurrent neural networks (LSTM), gradient boosting (XGBoost, LightGBM) and, more recently, Transformer architectures trained on time series — learn the non-linear interactions among all those variables. The result is a more accurate prediction, with explicit confidence intervals and the ability to retrain automatically when market behaviour changes.
According to Gartner's 2025 Supply Chain Planning Survey, companies that apply AI to forecasting reduce their forecast error (MAPE) by 20 % to 50 % compared with traditional statistical methods, and cut capital tied up in inventory by 10 % to 30 %. These figures are not magic: they depend on data quality and model design, but they indicate the order of magnitude of the opportunity.
Types of AI forecasting models
There is no single AI model for demand prediction. The choice depends on the number of SKUs, the time horizon and the nature of the available data.
| Model family | Use cases | Main advantages | Limitations |
|---|---|---|---|
| Classical statistical models (ARIMA, ETS) | Few SKUs, long and stable series | Explainable, fast to deploy | Do not capture external variables; sensitive to trend changes |
| Gradient Boosting (XGBoost, LightGBM) | Medium-sized catalogues, multiple covariates | High accuracy, tolerant of missing data | Require manual feature engineering |
| LSTM / GRU networks | Long series with complex patterns | Capture long temporal dependencies | Need lots of data; slow to train |
| Time-series Transformers (TFT, PatchTST, Chronos) | Large catalogues, heterogeneous signals | State of the art in multivariate accuracy | High computational cost; require MLOps expertise |
| Foundation models (Amazon Chronos, TimeGPT) | Fast start with limited historical data | Zero-shot or few-shot without retraining | Less customisable; cloud inference cost |
For most industrial or distribution SMEs, the most practical starting point in 2025-2026 is a combination of LightGBM with enriched external variables (working calendar, public holidays, promotional history) plus a statistical baseline model for low-turnover SKUs. The foundation-model layer is added when there are new SKUs with insufficient historical data.
Data you need for reliable forecasting
Forecast quality depends directly on input data quality. Before talking about models, you need to audit what information exists and in what state.
Essential internal data
- Sales history by SKU: at least 24 months at daily or weekly granularity. The more granular, the better.
- Promotions and discounts calendar: campaigns distort demand and the model needs to know this to avoid confusing them with real trends.
- Stock-out incidents: if a product was out of stock for three weeks, actual sales were lower than actual demand. Those periods must be corrected.
- Differentiated sales channels: e-commerce, physical store, distributors and B2B typically have completely different patterns.
External data that improves the model
- Macroeconomic indices: CPI, Consumer Confidence Index (CCI), exchange rate if imports or exports are involved.
- Weather data: relevant for food, beverages, fashion, DIY or any seasonal sector.
- Search trends: Google Trends can anticipate demand spikes weeks before they show up at the till.
- Competitor prices: in e-commerce, competitor price web scraping allows the model to adjust the forecast when a price war breaks out.
In our demand forecasting service, the first step is always a data diagnosis: what is available, what needs to be extracted from the ERP, which external sources are worth connecting and what cleaning the series require before training the first model.
The implementation process step by step
Phase 1 — Diagnosis and data extraction
Sales histories are extracted from the ERP or management system, periods with missing data or anomalies (stock-outs, mass returns, liquidations) are identified, and a clean dataset is built. This phase typically takes two to four weeks depending on how scattered the data is.
Phase 2 — Feature engineering and baseline
Derived variables the model needs are created: week of the year, day of the week, days until the next public holiday, 7-, 14- and 28-day sales lags, weighted moving average. A reference statistical model (usually SARIMA or Holt-Winters per product family) is trained to serve as an accuracy benchmark.
Phase 3 — AI model training and validation
The main model is trained with temporal validation (walk-forward validation): future data are never used for training, only for evaluation. Common metrics are MAPE (Mean Absolute Percentage Error) and wMAE (volume-weighted Mean Absolute Error). The goal is to improve the statistical baseline by at least 15-20 % on the validation set.
Phase 4 — Integration and production deployment
The trained model is connected to the ERP or warehouse management system (WMS) via API or structured interchange files. Purchasing managers receive a dashboard with the forecast for the next 4-8 weeks, confidence intervals and alerts for potential stock-outs or overstock. The model retrains automatically on a weekly or monthly schedule.
Concrete use cases by sector
AI demand forecasting is not an abstract tool: it has very specific applications depending on the sector.
Distribution and logistics
A distribution company with 3,000 SKUs can reduce its Days of Inventory Outstanding (DIO) by accurately planning when to replenish each SKU. AI identifies which ones have seasonal demand, which react to retailer promotions and which move erratically. With that segmentation, the right replenishment policy is applied to each group.
Agri-food industry
Production has long lead times and perishable ingredients cannot be overstocked. A model that combines sales history, weather data and public holidays reduces finished-product waste and optimises raw-material purchase orders.
Retail and e-commerce
Discount campaigns (Black Friday, seasonal sales, Prime Day if the channel is Amazon) generate sharp demand spikes that are difficult to capture with historical averages. AI models learn the lift effect of each promotion type and incorporate it into future forecasts.
Services with limited capacity
Clinics, workshops, training studios or any business with finite human or technical resources can use forecasting to anticipate demand peaks and adjust workforce planning with sufficient lead time.
Integration with ERP and purchasing systems
A forecasting model that generates predictions in an isolated Excel file has limited value. Real impact comes when the forecast is integrated directly into the purchasing and production workflow.
The most common ERPs in Spanish SMEs — Sage, Odoo, Microsoft Business Central, Holded — have APIs or import modules that allow them to receive AI-model forecasts and automatically convert them into purchase proposals or manufacturing orders. This integration is managed by our applied AI team, coordinating with the Systems division when the ERP integration requires custom development.
A critical point is model governance: who validates predictions before they become real orders, how manual corrections are recorded and how the model is fed back with actual sales data. Without a human-review process, the model can perpetuate errors or react late to structural market changes.
What results you can expect (and what you cannot)
AI forecasting does not eliminate uncertainty: no model predicts the future with absolute certainty. What it does is reduce forecast error consistently and, above all, quantify uncertainty explicitly through confidence intervals.
Typical results documented in case studies of mid-sized distribution and manufacturing companies in Europe include:
- MAPE reduction of 20 % to 45 % compared with manual or statistical forecasting.
- Safety stock level reduction of 10 % to 25 % without an increase in stock-outs.
- Stock-out reduction of 15 % to 30 % on high-turnover SKUs.
- Savings on urgent transport costs for emergency replenishment.
What the model cannot do without human intervention: anticipate a sudden supplier change, a global supply crisis with no historical precedent or a commercial decision taken the day before. That is why the process always includes a manual review and adjustment layer for extraordinary events.
Frequently asked questions
How much historical data do I need to get started?
The reasonable minimum is 18-24 months of sales history at daily or weekly granularity per SKU. With less data a model can be trained, but accuracy will be lower, especially for capturing annual seasonality. Foundation models (such as Amazon Chronos or TimeGPT) can deliver acceptable results with shorter histories thanks to pre-training on millions of external time series, although having sufficient proprietary data is always preferable.
Does AI demand forecasting require an in-house technical team?
Not necessarily. The model can be deployed and maintained by a specialist external team that handles periodic retraining, accuracy monitoring and data pipeline updates. What the company does need is an internal owner who reviews forecasts before converting them into orders and who communicates to the AI team the extraordinary events (launches, customer losses, price changes) that the model cannot infer on its own.
How quickly does the investment pay back?
It depends on inventory value, stock-out costs and business margin, but in distribution companies with average inventory above 500,000 euros and more than 500 SKUs, return on investment typically occurs within the first year. The reduction in capital tied up in stock is the most direct and measurable saving; savings on urgent orders and the revenue increase from improved product availability are additional benefits.
Can it be combined with the existing S&OP planning process?
Yes, and it is the most recommended integration. The AI model generates the statistical baseline forecast; sales, purchasing and production managers review it in the monthly S&OP (Sales & Operations Planning) cycle and apply the qualitative adjustments that correspond. The system records those corrections and incorporates them as an additional signal in the next retraining cycle. The result is a forecast that combines the mathematical rigour of the model with the business knowledge of the team.