What the AI in Wameed actually does, and what it does not

"AI" in retail software usually means a chart with a trend line. Here are the four specific jobs it does in Wameed, what each needs to work, and where each one stops.

1. Demand forecasting

What it does: predicts how much of a product you will sell over a coming period, using sales history, day of week, season and the patterns it has learned from your own data.

What it needs: history. A meaningful forecast needs several months of clean sales data, and it improves substantially after a full year, because that is when it can see seasonality. A forecast built on eight weeks is extrapolation wearing a confident face.

What it does not do: predict things that are not in the data. A new competitor opening across the road, a road closure, a viral post — none of these are visible to a model that only sees your sales.

Use it for: ordering quantities, production planning in a bakery or kitchen, and staffing levels. Seestaff scheduling.

2. Anomaly detection

What it does: flags patterns that differ from your established norms — an unusual concentration of voids at one terminal, refunds clustered at the end of one person's shifts, a product whose stock variance moves differently from everything around it, a discount rate that has quietly drifted.

What it needs: a few months of normal operation to learn what normal is.

What it does not do: conclude that anyone has done anything wrong. It produces a list of things that are statistically unusual. Most unusual things have innocent explanations — a new staff member, a change in process, a promotion nobody logged.

Use it as: a prompt to look, never as a finding. The investigation order inshrinkage still applies: rule out administrative error first, then supplier shortfall, then damage, then external theft, and only then consider internal loss.

We are deliberate about this. A tool that tells a shop owner an employee is stealing, when the real cause is a carton entered as a piece, does real harm to a real person.

3. Invoice OCR intake

What it does: reads a photographed or scanned supplier invoice and turns it into a goods receipt — supplier, products, quantities, prices — matched against your open purchase orders.

What it needs: a legible image and products it can match to your catalogue.

What it does not do: read handwriting reliably, or match a product it has never seen without help. First time you receive from a supplier, you will do some matching by hand; after that it learns the mapping.

Use it for: removing the most tedious data entry in a shop. A delivery of forty lines is a photograph rather than fifteen minutes of typing. Seepurchase orders.

Always check before posting. It is intake, not authority — the numbers still need a human's eye before they become your stock and your payable.

4. Smart reorder

What it does: combines the forecast, your current stock, supplier lead times and your reorder points to propose a purchase order with quantities.

What it needs: accurate stock, which means the disciplines ininventory basics, and lead times that reflect what suppliers actually do rather than what they promise.

What it does not do: send the order. It proposes; you review and approve. Automatic ordering without review is how a shop ends up with eighty units of something because of a data error.

The honest limits

Garbage in. Every one of these depends on your data being reasonable. Wrong cost prices, duplicate products, unrecorded waste and stock that has never been counted will produce confident and wrong output. Fix the data first; the AI is a multiplier of whatever you feed it.

Cold start. A new shop gets little value for the first few months. This is unavoidable, and a vendor who claims otherwise is selling you something that is guessing.

Not a manager. These tools surface things. Deciding what to do is still yours, and should be.

What we deliberately do not do

• Automatic price changes. Your prices are a commercial and reputational decision, not an optimisation target. Seepricing strategy.
• Automatic ordering without review.
• Anything that presents a statistical anomaly as a conclusion about a person.

These are product decisions rather than technical limits, and we would make them the same way again.

Getting value from it

1. Get your data right — cost prices, stock counts, recorded waste
2. Give it a season of history before judging the forecast
3. Treat anomalies as questions, not answers
4. Review every proposed order for the first month
5. Check every OCR intake before posting

See theWameed AI feature page.