The problem with a fixed minimum

Standard reordering rules work like this: keep at least 50 in stock, and when it drops below, order up to 200.

Simple, and it has two weaknesses.

The numbers were set once. Usually during implementation, from last year’s demand, and never revisited. Demand moves; the rule does not.

They treat every week the same. A minimum that is right in a quiet month leaves you short in a busy one.

AI forecasting addresses both. Instead of a fixed number, it estimates what you will actually need over the coming lead time, from your own sales history.

What it looks at

Four things a static rule cannot:

Trend. Sales climbing or falling over months.

Seasonality. Recurring patterns — busier in December, quiet in August, a spike every quarter-end.

Lead time variability. Not just the average, but how unreliable a supplier is. A vendor who is sometimes ten days late needs a bigger buffer than one who is always seven days.

Relationships between products. If two items sell together, the forecast for one informs the other.

What good looks like

Two numbers tell you whether it is working, and they pull against each other.

Stockouts. How often you could not fulfil because you had nothing. Every one is lost revenue and sometimes a lost customer.

Stock value held. Cash sitting on shelves. Every unit held is money not doing anything else.

Anyone can improve one by worsening the other. Order everything in bulk and you will never stock out — you will just have your working capital in a warehouse.

The test of a good forecast is improving both at once, or holding one steady while the other improves meaningfully.

Measure both before you change anything. Without a baseline you cannot tell whether it helped.

Where it works well

Products with real history. Two years or more of consistent sales gives something solid to learn from.

Stable, repeating demand. Regular sales with recognisable seasonal shape.

Many products. The value is partly that it handles thousands of items you could never review individually. Reviewing fifty by hand is feasible; reviewing five thousand is not.

Reliable data. Accurate stock records and complete sales history.

Where it does not

Four situations. Being honest about these is what separates a useful tool from an expensive disappointment.

New products. No history means no forecast. It will produce a number anyway. That number is not information.

Volatile demand. If your sales are driven by a few large unpredictable orders, past patterns will not predict future ones. A forecast implies a regularity your business does not have.

Anything shifting structurally. A new competitor, a changed price, a discontinued line, a lost major customer. The model is learning from a world that no longer exists.

Inaccurate stock records. If your system says 40 and the shelf has 12, the forecast is built on fiction. This is the most common real cause of a forecasting project disappointing — and it has nothing to do with the AI.

Keep the human review

Forecasting should propose, not decide.

Things it cannot know:

• You are running a promotion next month
• A major customer has signalled a large order
• A supplier is about to raise prices
• You are discontinuing the line in six weeks
• Your competitor has just gone out of business

All of these change what you should order, and none of them are in your sales history.

A workable arrangement: the system proposes quantities, a person reviews the list weekly, and overrides where they know something the data does not. That is faster than deciding everything manually and safer than automating it entirely.

Introducing it sensibly

1. Fix your stock accuracy first. Cycle counting, not an annual count. Nothing downstream works without this.

2. Measure your baseline. Stockouts per month and average stock value, for three months.

3. Start with your fast movers. The top twenty percent of products by volume, where the history is best and the impact largest.

4. Run it in parallel. Let it propose while your existing rules still operate. Compare for a couple of months.

5. Expand gradually. Add product groups as confidence builds.

6. Keep the weekly review. Permanently, not just during the trial.

What to watch for

Over-confidence on thin history. A forecast for a product with six months of data will look as authoritative as one with six years. It is not.

Ignoring the override. If your team stops reviewing because the system is usually right, the month it is wrong will be expensive.

Chasing the wrong metric. Reducing stockouts alone is easy and expensive. Watch both numbers.

Blaming the forecast for a data problem. If proposals look wrong, check stock accuracy before questioning the model.

The short version

Inventory forecasting is genuinely useful where you have history and accurate stock records. It handles a scale of product range no person can review, and it adapts as demand moves.

It is unreliable on new products, volatile demand, and any market that has just changed.

Measure stockouts and stock value together. Keep a weekly human review. And fix your stock accuracy first — most forecasting disappointments are really counting problems.