Operations
AI for demand and stock forecasting
Start with two or more years of sales data by product and week, and use a forecasting tool to project demand and flag stock risks. Add what the data cannot see: promotions, holidays, tourist season and supplier delays. A buyer reviews the forecast and decides the orders. Begin with your best-selling products, where errors cost the most.
Effort:
High
Time saved:
Often several hours saved per ordering cycle (estimate)
The problem today
Many SMEs order stock based on last year's numbers and gut feeling. The result is stockouts in peak season and cash tied up in slow-moving products.
How it works, in plain words
It is like an experienced buyer who looks at past seasons and tells you what you will probably need, while you add what only you know.
What goes in: two or more years of sales by product and week, current stock, supplier lead times, and a calendar of promotions, holidays and season dates.
What the AI does: it finds repeating patterns such as seasons and trends, projects demand for the coming weeks and flags stock risks.
What a person checks: the buyer compares the forecast with what they know about promotions, suppliers and the market, and decides the orders.
What comes out: a suggested order list and a short risk report for your top products, reviewed every week or month.
It cannot foresee one-off events such as a heatwave, a strike or a competitor's sale, and it is only as good as your sales history.
How it could work for businesses like yours
Illustrative examples, not client case studies.
A beachwear retailer with five stores in Halkidiki
Each spring they guess quantities by size and color and end the season with too much of some lines. They would forecast the top 20% of products from three seasons of weekly sales, adjusted for Easter and school holidays. They would measure the stock left at the end of September and the sales lost to stockouts.
A food wholesaler on Santorini that supplies hotels and restaurants
Demand swings with tourist arrivals and ferry schedules. They would combine sales history with the season calendar and supplier lead times, and the purchasing manager would review the forecast every week. A realistic first target could be fewer emergency orders in the peak months.
An 18-person craft brewery near Trikala
They brew by gut feeling and some summers run out of their best seller. They would forecast demand per beer and month and plan brewing slots from it. They would track stockouts and how much beer passes its best-before date.
How to do it, step by step
Check your data. You need clean sales history per product, ideally two years or more by week or month, plus current stock and supplier lead times. Fix gaps and changed product codes first.
Start with your top products. Pick the 20% of products that bring most revenue. Forecasting everything at once hides whether it works.
Build a simple baseline. Use a basic forecast, such as Excel's Forecast Sheet, and compare it with what actually sold last season.
Add known events. Mark promotions, Easter, the summer season and other local peaks so the model or the buyer can adjust.
Treat it as a signal, not an order. Let the tool flag likely stockouts and overstock, while the buyer decides quantities.
Pilot over one full season. Track forecast error, stockouts and excess stock against the previous year.
Tools to consider
Microsoft Excel (Forecast Sheet) with Copilot: built-in forecasting from historical data, and Copilot can help explain trends. Caveat: it handles one series at a time well, so many products need a more structured setup.
ChatGPT (data analysis) or Claude: can analyze a sales export, find seasonality and explain a draft forecast. Caveat: results depend heavily on your prompt, so validate against actual sales before trusting them.
Inventory planning add-ons, for example Inventory Planner by Sage: connect to e-shop platforms such as Shopify and suggest reorder points. Caveat: they are only as good as your product data and lead times, and add another subscription.
Data protection and GDPR
Demand forecasting works on aggregated sales and stock data, so it usually involves little personal data. Keep it that way: forecast by product, store and week, not by named customer. If you add loyalty or customer-level data to forecast by segment, remove direct identifiers and limit who can access it. The main risk here is commercial rather than legal, because a confident but wrong forecast can tie up cash. Keep a person responsible for ordering decisions and document how forecasts are made. Staff using these tools also need basic AI literacy under the EU AI Act.
Mistakes to avoid
Feeding the tool messy data with renamed product codes and missing months.
Ignoring promotions and seasonality, then blaming the model for missing the Easter peak.
Letting the forecast place orders automatically before you have measured how often it is wrong.