AI demand forecasting for small and mid-sized manufacturers
Anticipate demand and plan with more confidence. We build forecasts from the sales history you already have, test them against a baseline, and point planners at the products and locations that actually need attention.
Forecasting by instinct has a ceiling
Most small manufacturers forecast with a spreadsheet, last year's numbers and the sales team's judgement. That is a reasonable system and it works for a while. Its limits show up as the catalogue grows: nobody can hold a thousand item-level patterns in their head, so attention goes to the items that shouted loudest most recently.
The cost is quiet. Stock builds on products whose demand has been slowly declining. Seasonal ramps get noticed a month late. A promotion distorts the baseline for a year because nothing recorded that it happened. And when demand does shift, it takes a quarter to be sure it was a shift rather than noise.
A forecast will not remove uncertainty. What it does is quantify it, apply the same rigour to every item, and free your planners to spend their judgement where judgement actually adds something.
What we deliver
A working forecast your team can run, not a one-time projection.
- Item-level forecasts at the time horizon your planning cycle needs, with a prediction interval rather than a single number.
- A documented accuracy comparison against a naive baseline and against your current process, using MAPE, MASE or whichever measure suits your demand.
- A method chosen per demand pattern — seasonal, trending, intermittent or new-product — rather than one model applied to everything.
- An exception list that tells planners which items need a human decision this cycle and which do not.
- A monthly or weekly refresh that reruns the forecast as new sales arrive, with monitoring for drift.
- Documentation of every assumption, so the forecast can be defended in an S&OP meeting.
How it works
-
Assemble and clean the history
Sales history at item level, with the known distortions flagged: promotions, one-off bulk orders, stockout periods where demand was censored by lack of supply. Cleaning this properly matters more than the model choice.
-
Profile every item's demand pattern
Seasonal, trending, stable, intermittent or too new to model. Each pattern gets a method suited to it. This step is why a single-model forecast underperforms.
-
Test candidates against a baseline
Statistical methods and machine learning models compete against a naive forecast on held-out data. The winner is whichever is most accurate on your history, not whichever is most impressive to describe.
-
Put it where planners work
Into your ERP, your BI tool or the spreadsheet your team already uses. The forecast that gets used beats the forecast that is technically better.
-
Monitor and retrain
Accuracy is tracked each cycle. When a model starts to drift, someone is told, and the retraining is already scripted.
What you need to get started
Sales history, an item list, and someone who can tell us what happened in the odd months.
- Item-level sales or shipment history by period — monthly is usually enough, weekly is better.
- An item master, with category and lifecycle status if you track it.
- Anything you know about promotions, price changes, lost customers or supply interruptions. This context is what separates a usable forecast from a confident wrong one.
A CSV export from your ERP is a perfectly good starting point.
Typical outcomes
- Planners reviewing an exception list instead of every item.
- Seasonal ramps anticipated rather than noticed after the fact.
- A forecast the sales and operations meeting can argue with, because the assumptions are written down.
- Better inputs to stock targets, which is where the forecast usually pays for itself.
Common questions
How much sales history do we need?
Two to three years is comfortable because it lets a model separate a seasonal pattern from a trend. Twelve to eighteen months is workable, with wider uncertainty and more conservative methods. Under a year we would normally recommend a simpler approach and revisit once more history accumulates.
Is “AI forecasting” just a rebranded moving average?
Not as we use it, but the honest answer is that a well-chosen simple method beats a complicated one more often than vendors admit. We always test against a naive baseline. If a seasonal naive forecast beats a gradient-boosted model on your data, we will tell you and recommend the simple one.
Can you forecast items that sell irregularly?
Yes, and it needs different methods. Intermittent demand — many periods of zero with occasional spikes — is common for spare parts and industrial items, and standard time-series methods handle it badly. Croston's method and its variants exist for this. The practical output is often a stock policy rather than a point forecast.
What accuracy should we expect?
We will not quote a number before seeing your data, and you should be wary of anyone who does. Achievable accuracy depends on how variable your demand is, which is a property of your business rather than of the model. What we commit to is measuring it honestly against a baseline, and telling you if the improvement is too small to be worth the effort.
Can we forecast in Excel if that is what our team uses?
Often, yes. If the result needs to land in a spreadsheet your planners already work in, we build the model to write into it. Forcing a team into a new tool is the most reliable way to make a forecast go unused.
Find out what your data can already tell you
A free 30-minute data readiness assessment. We look at what you have, tell you what it can support today, and name the first thing worth building. No preparation needed.