"We order what we sold last month plus twenty percent, because that's what we always do." We've heard this sentence over a dozen times in the last year. And not from the smallest companies either. In one — a 30-person distributor of spare parts — 800 SKUs were ordered exactly this way. The result was twofold: cash frozen in stock nobody was buying, and at the same time regular shortages of the products that sold best. A classic.
Before somebody shouts "AI will solve it", take a step back. Demand forecasting in an SME doesn't start with a model. It starts with the question: do we even have something worth forecasting? And if so — with the simplest tool that actually works. Excel, a basic ERP module and a dedicated AI model are three very different levels of investment. Picking the wrong level costs about the same as having no forecast at all.
When to forecast, when to guess
Forecasting only makes sense when demand has a repeatable pattern. Sounds trivial? It is — but half the companies we talk to try to forecast things that simply cannot be forecast: one big client, unique projects, sales rhythm driven by individual contracts rather than the market.
In practice, forecasting starts paying off when three conditions are met. First — you have at least 100–200 SKUs (or catalogue items) and it's no longer possible to keep in your head what sells how much. Second — sales history covers a minimum of 18–24 months, so a model can actually see seasonality. Third — lead time (from ordering with a supplier to stock on the shelf) is longer than a week, meaning you have to decide in advance.
If you have 50 SKUs and a one-week lead time — you don't need a forecast. You need discipline in ordering and an ERP alert when stock drops below a threshold. Everything else is over-engineering.
Level 0 — Excel and a weighted average
It's surprising how much forecasting can be squeezed out of a properly built Excel. Exponential smoothing (the FORECAST.ETS function in Excel and Google Sheets) will figure out trend and seasonality on its own if you give it two years of clean data. You don't need statistics — you just need a sales export with dates and quantities.
At the 30-person spare-parts distributor mentioned above, this is exactly where we started. Three days of work: a 24-month sales history export from Comarch Optima, a spreadsheet with a single FORECAST.ETS formula per SKU, and a simple table with a suggested order quantity. After one ordering cycle: 12% less capital frozen in stock and 30% fewer shortages in the top-100 SKUs. Without buying any software at all.
That doesn't mean Excel is forever. It only means that no forecasting system will deliver better results than a solid Excel baseline, until you've proven to yourself that the baseline isn't enough. This is a test that 90% of "AI solution" vendors would rather you didn't run.
Level 1 — What your ERP already does
Most ERPs used by Polish SMEs — Comarch Optima, enova365, Subiekt GT, Symfonia — have some sort of forecasting or minimum-suggested-order module. It usually does the same thing as Excel, but with one huge advantage: the numbers sit in the same place as inventory, orders and suppliers. Nobody has to export, recalculate and paste back.
This level makes sense when the Excel from Level 0 works well but maintaining it starts taking too much time. Typically we move here at 300–500 actively managed SKUs. Deploying a forecasting module inside an existing ERP is about a week of work — you have to configure parameters (season length, aggregation method, safety tolerance) and convince the person doing the ordering that the system's suggestions are worth taking seriously.
Watch out for one trap: the forecasting modules in Polish ERPs generally can't tell a promotion apart from normal sales. If in April you ran a −40% discount on one SKU and it sold ten times more than usual, the module will forecast that next April it will also sell ten times more. That's not the model's fault — it's the data's fault. Before deploying an ERP forecast, tag the "unusual" periods in your sales history.
Level 2 — When you really do need AI
A dedicated machine-learning model (say Prophet, ARIMA with regression, XGBoost with seasonal features) only makes sense once you cross a specific complexity threshold. From what we see, it's one of three scenarios:
- You have several thousand SKUs and their sales patterns differ meaningfully (some seasonal, some flat, some weather-driven).
- Demand depends on external factors Excel can't see — temperature, exchange rates, marketing campaigns, Catholic and Orthodox holidays overlapping.
- Margins are thin enough (typically distribution, e-commerce) that even a 2% accuracy improvement means real money.
Below that threshold the investment won't pay off. An AI model needs 2–3 years of historical data at minimum (five is better), an engineer to build and maintain it, and a process that reacts to its output. The cost isn't a licence — it's the role of a human who does this for the first six months, before the model starts running stably.
There is one sensible shortcut: SaaS forecasting tools (RELEX, Slimstock, ForecastPro in the cloud). They take the model-building off your plate, but cost €500–1200 per month and require clean data on sales, promotions and supplier lead times. Worth considering if warehouse turnover exceeds a few million a year.
What not to do
A few traps SMEs regularly fall into when starting with forecasting.
- Don't buy "AI" without comparing it to Excel. Any vendor who refuses to compare their model against a simple baseline (moving average, exponential smoothing) is selling you a slide deck, not results.
- Don't automate ordering right away. For the first 3–6 months the forecast should be a suggestion the buyer confirms. Automatic ordering off a half-baked model is the shortest route to filling your warehouse with product nobody wants.
- Don't ignore your sales team. A rep who knows "in November the big client stocks up for the whole quarter" carries information the model doesn't. The forecast is the starting point for a conversation, not the final word.
- Don't forecast what you order project-by-project anyway. Products made against a specific customer order get ordered against that customer order, not against history.
Where to start
The practical sequence we take clients through:
- Rank your SKUs by turnover (ABC classification). Forecast only A and B — the rest is noise.
- Export 24 months of sales for those SKUs from your ERP and tag the unusual periods (promotions, outages, one-off orders).
- Build the Excel baseline:
FORECAST.ETSper SKU, suggested order size = forecast × lead time + safety stock. - Test through three ordering cycles. Measure: stockouts, capital frozen in stock, number of items still ordered "by eye".
- If the baseline works well but takes too much manual effort — then reach for the ERP module or a SaaS tool.
Total starting investment: 2–4 weeks of one person's work (yours or an outside consultant's). Zero licence fees. Zero external vendors with a presentation about "digital transformation".
If you want to talk about which forecasting level fits your company and which step to take next — the first call is free and takes half an hour.
No forecasting system will deliver better results than a solid Excel baseline, until you've proven to yourself that the baseline isn't enough.