AI how-to guides Operations and processes
How to forecast demand and inventory with AI
Train forecasting models on your sales history together with promotions, prices, calendar and external signals, producing coherent forecasts by product, location and channel with a confidence range. Planners review the forecast, override it with a reason when they know something the model does not, and the result feeds replenishment. Start with one category, compare against your current forecast and widen.
What AI does
- Forecasts demand at the level you plan, such as SKU, store, region and channel, and keeps the levels consistent with each other.
- Separates baseline demand from the effect of promotions, prices, events and seasonality.
- Gives a confidence range for each forecast, not just a single number, so safety stock can follow the uncertainty.
- Explains which factors drove each forecast so planners can judge it.
- Simulates what-if scenarios, such as a planned promotion or a price change, before committing stock.
What stays with people
- Overriding the forecast when they know something the data does not show, with a reason code.
- Setting service levels and inventory policy: how much stockout and overstock risk the business accepts.
- Deciding on new products, assortment changes and supplier issues the history cannot anticipate.
How to put it in place
-
Map the current process and its accuracy gaps
Document how forecasts are made today, at what level and how often, and where they fail: peaks, promotions, new products or slow movers. Measure the current forecast accuracy so you have a baseline to compare against.
-
Gather and reconcile the data
Bring together sales history, promotions, prices, stock-outs, calendar and events, and the product, location and channel hierarchies. Flag periods where sales were limited by lack of stock, because they understate real demand.
-
Build and back-test the models
Train models that handle regular weeks and peaks, decompose promotion lift and reconcile forecasts across the hierarchy. Test them on past periods they did not see and compare with your current forecast, category by category.
-
Give planners a workbench to review and override
Show forecasts with their range and main drivers, and let planners override with a reason code. Track whether overrides improve or worsen accuracy so both the model and the team learn.
-
Connect to replenishment and pilot one category
Post the final forecast, model plus overrides, to the ERP or replenishment engine with full lineage. Run one category live, measure stockouts, overstock and accuracy against the baseline, and adjust.
-
Roll out by category and govern the loop
Extend category by category, adding promotion modelling and channels as you go. Monitor accuracy and bias continuously and retrain when patterns change.
What you need
- Several seasons of sales history at the level you plan, with stock-out information if available.
- The promotion calendar, prices and relevant events, past and planned.
- Clean product, location and channel hierarchies and master data.
- Access to the ERP and replenishment systems to read data and post forecasts.
- Planners involved from the start, with an owner for the forecast process.
When it is not worth it
- Very few products with stable demand, where a simple statistical method or a spreadsheet already works.
- When sales history is missing, unreliable or does not reflect the current business.
- If inventory decisions are fixed by contracts or suppliers and a better forecast would not change them.
Risks and how to control them
- Biased history: stock-outs and past promotions distort demand. Correct for them before training and check results by segment.
- Black-box forecasts planners do not trust: show drivers and ranges, and keep the override in their hands.
- Breaks in patterns, such as a crisis, a new competitor or a channel shift: monitor accuracy and bias and trigger retraining or manual review when they drift.
- Overrides that make things worse: measure the value each override adds and review recurring ones with the team.
- Silent integration errors: reconcile what the ERP received against what was forecast and alert on mismatches.
How Thinkia does it
We start by mapping categories, the promotion calendar, your current forecast process and the biggest accuracy gaps, then define the hierarchy, the override workflow, the replenishment integration and the accuracy KPIs. Sales, promotions, weather, events and prices are normalised and the hierarchies are built once and reused, with a single model family for peaks and regular weeks and promotion lift decomposed from baseline.
Planners see which factors drove each forecast, with confidence intervals, and override with reason codes; replenishment uses both model and override with full lineage, and the model learns from overrides over time. We pilot on one category, integrate with the ERP and roll out category by category.
The Thinkia solution behind this guide
- Demand forecasting & inventoryForecasts that survive peaks and Tuesdays
- Supply-chain risk monitoringSee disruption before it lands
Thinkia products involved
Frequently asked questions
Is AI demand forecasting better than traditional methods?
Not always. Machine learning tends to help when many factors interact, such as promotions, prices, events and many products and locations. For a few stable products, classical statistical methods can be just as good, so always compare against your current forecast.
What data do I need to start?
Sales history at the level you plan, the promotion calendar and prices, product and location hierarchies and, ideally, information on stock-outs. External data like weather or events helps in some categories but is not where to start.
How do you forecast new products with no history?
By using similar products, attributes and launch patterns as reference, and by giving planners more weight in the early weeks. Treat it as a separate case with its own review.
Will it replace our planners?
No. It removes manual number crunching and gives them a better starting point, but planners still own overrides, exceptions, new products and the inventory policy.
What is hierarchical forecast reconciliation?
It makes forecasts at different levels add up: SKU forecasts sum to the category, stores to regions and channels to the total. Without it, each level tells a different story and plans do not match.
Does it work with SAP or Oracle?
Yes. The forecasts can be read from and posted to ERP and replenishment systems through their standard interfaces, so planners keep working in the tools they already use.
Keep exploring
Related guides
- How to use AI in procurement
- How to do financial forecasting with AI
- How to improve data quality with AI