A forecast is only as good as the history behind it. Most inventory teams that ask for AI actually need clean stock records, consistent units and a way to separate real demand from stockouts.
The model is the last step, not the first.
Forecasting is a data problem
If an item was out of stock for two weeks, the sales history shows low demand. A model trained on that history learns the wrong lesson. Before any forecast, mark the stockout periods and reconstruct what demand would have been.
The same applies to unit changes, pack sizes and product replacements. A new SKU that replaces an old one needs a linked history, or the model sees a brand-new product with no demand.
The minimum history
A practical starting point is 18 to 24 months of daily or weekly sales, purchase receipts, stock on hand, lead times, returns and price changes. For seasonal categories, two full cycles are better.
If that history does not exist, forecasting software will produce confident numbers with no basis. Fix the capture first: every sale, return, transfer and stock adjustment should write a timestamped record.
Seasonality and lead times
Seasonality is not only festivals. It includes weather, school terms, salary cycles, local events and the purchasing pattern of your largest customers. A model that only sees month numbers will miss those.
Lead time matters as much as demand. A forecast that says “sell 400 units next month” is useless if replenishment takes six weeks. The output should be a reorder recommendation with a safety stock, not just a demand number.
Promotions and returns
Promotions create spikes that are not baseline demand. Mark them in the history so the model can separate the lift from the underlying trend. The same applies to bulk orders, dealer schemes and marketplace campaigns.
Returns need their own treatment. A high-return SKU has a different net demand from its gross sales, and ignoring returns overstates what to buy.
Where AI helps
AI is useful when the pattern is complex: many SKUs, short lifecycles, promotions, substitutes and uneven lead times. It can combine signals that a spreadsheet cannot hold at once.
It is less useful when the catalogue is small and stable. A simple moving average with a safety stock rule may be easier to explain and just as accurate.
A pilot that proves accuracy
Pick 50 to 100 representative SKUs. Hold back the last three months of history, forecast it, and compare with what actually sold. Measure error by SKU and by value, not just an average percentage.
Run the current method on the same period. If the model does not beat the existing rule, do not roll it out. A forecast that is harder to explain and no more accurate is a cost, not an improvement.
Safety stock and service level
A forecast is only half of the replenishment decision. The other half is safety stock, which depends on lead-time variability, demand variability and the service level the business wants to offer.
Calculate the two variabilities separately. A supplier with a stable lead time needs less buffer than one that occasionally ships late. A fast-moving SKU with volatile demand needs more than a steady one.
Set the service level by category. A critical spare part may justify a higher level than a slow-moving accessory. The number should come from the cost of a stockout, not from a default in the software.
Review the buffer monthly. A safety stock that never changes is usually a number someone set once and forgot.
FAQ
How much sales history do we need?
Eighteen to twenty-four months is a practical starting point. Seasonal categories need at least two full cycles, and stockout periods must be marked.
Can AI forecast a new product?
It can use similar products, price and launch plans, but the uncertainty is higher. Keep a manual review for the first few cycles.
Do we need a data warehouse first?
Not always. A clean transaction history in the inventory system is enough for a pilot. A warehouse helps when multiple sources and long histories are involved.
For the system side, see inventory management and inventory management software cost. For a custom forecasting layer, start with AI app development cost.
