For two years now, every management system has advertised artificial intelligence. The demos are impressive, the promises broad, and owners leave those meetings without knowing what will actually change in their week.

The useful question is not "does this software do AI". It is: which decision, currently made blind or made too late, will be made better or earlier? If there is no answer, the feature is a sales argument, not a tool.

Here is the sort, use by use.

What already produces a measurable result

Demand forecasting. By far the most mature use. From sales history, seasonality and recent trends, a model estimates upcoming quantities per reference. It does not replace the buyer: it gives them a quantified starting point instead of a mental average. The gain shows up directly in two numbers — fewer stockouts and less overstock, at the same time.

The condition is strict: you need at least twelve to eighteen months of clean history. On a young or poorly maintained catalogue, forecasting produces elegant, wrong numbers.

Anomaly detection. A model learns what is normal in your data and flags what is not: an unusual discount, an out-of-range stock variance on a family, a stock movement at an improbable hour, an order ten times a customer's usual size. This works well because it does not ask the model to predict the future, only to recognise a departure from the past. That is far easier, and far more reliable.

Natural-language search. Asking "which references have not moved in three months in the Casablanca warehouse" and getting the list, without building a filter. The gain per query is not spectacular, but it opens analysis to people who would never have run it otherwise. An underrated use.

Related-product recommendations. From real baskets, the system identifies what gets bought together and suggests it at the point of sale or in a follow-up. On a wide catalogue, the effect on average basket size is measurable within weeks.

What is still fragile or marginal

Automatic price setting. Technically possible, commercially risky. Price depends on the customer relationship, local competition and strategy — most of which is not in your data. A model will optimise short-term margin and damage a long relationship without ever noticing.

Forecasting on short history. New product, new shop, shifting market: there is nothing to learn from. The model will still output a number — which is exactly what makes it dangerous. A number derived from nothing looks like a number derived from something.

Assistants that "run your business". A phrase that should trigger one simple question in the demo: on exactly which data, and what happens when the answer is wrong? The answers to those two questions separate the tool from the pitch.

The one condition that decides everything

No model compensates for incomplete data. This is the one rule with no exceptions.

In practice, useful AI assumes that sales are recorded line by line rather than as totals, references are unique and stable over time, stock movements are entered when they happen, and returns, breakage and credits are tracked like everything else.

A company that already keeps clean data gets a result within weeks. A company where half the movements go through a paper notebook will get nothing — whatever the model, whatever the vendor.

There is good news in that, from an unexpected angle: the work required to benefit from AI is exactly the work that improves your management without it. No investment is wasted.

Questions to ask in a demo

  1. On which of my company's data does this feature rely, precisely?
  2. How much history is needed before it becomes reliable?
  3. Can I see why it proposes this result, or is it a black box?
  4. What happens when it is wrong — who sees it, and when?
  5. Can I switch the feature off without losing the rest of the software?
  6. Is my data used to train a shared model with other customers?

Question 3 is decisive in a management context. A replenishment proposal that cannot be explained will not be followed by an experienced buyer, and they will be right not to follow it.

Question 6 deserves an answer written into the contract, not given verbally.

Where to start

  1. Pick one decision you make regularly with poor information — most often: what to reorder, and how much.
  2. Check that the data behind that decision is clean over the last twelve months. If it is not, that is the work to do first.
  3. Run the feature in parallel with your current method for a month, changing nothing in your actual orders.
  4. Compare: on the references where model and buyer disagreed, who was right?
  5. Only extend after that comparison.

Step 3 is the only honest way to find out whether AI brings you anything. It costs one month of attention and prevents years of misplaced confidence.

Where the software fits

Useful AI in commercial management is not the kind that decides for you. It is the kind that makes you notice on Tuesday what you would otherwise have discovered next month.

In G-stock, assistance covers demand forecasting per reference, replenishment proposals explainable line by line, anomaly detection on sales and stock movements, and natural-language search across your own data. Every proposal states what it is based on, and still requires approval.