Slow stock and unavailable products: what to automate before using AI

The screen shows stock, but the product is not on the shelf. Meanwhile, another item has not sold for weeks and appears on a new purchase order. Before seeking a more sophisticated forecast, I would check how each piece of information reached the record.

These are hypothetical situations. A system connected to the point-of-sale terminal can deduct sales, but it also needs to capture receipts, returns, shrinkage, reservations and adjustments. Recorded stock and stock available to sell can differ. A fast update does not correct a movement nobody captured.

The GAO counting guide explains that counting is part of control and that quantities must be compared, differences investigated and records adjusted. It is a 2001 guide, not evidence of the current effectiveness of AI agents in Spanish shops. Physical inventory count controls.

Replenish with data you can question

For an initial rule, I would use observed daily sales, actual delivery lead time and a buffer of units chosen by the business. The basic reorder point would be daily sales × replenishment days + buffer, expressed in units. That simplification is not a probabilistic forecast and does not apply equally to seasonal or irregular-demand products.

Hypothetical example: four units a day, five replenishment days and an eight-unit buffer give 4 × 5 + 8 = 28 units. To check whether to reorder, I would also look at outstanding purchase orders and reservations. Twenty physical units, ten due to arrive in time and six reserved give a simplified inventory position of 24 units.

That figure calls for review; it does not by itself give the exact quantity to buy. Missing factors include the coverage target, minimum order lots, delivery days, available cash and, where relevant, expiry dates. A confirmed order arriving too late does not solve a shortage tomorrow either.

Observed demand needs context: low sales because a product was unavailable do not mean nobody wanted it. I would not treat out-of-stock days as normal selling days.

Use a different calculation for slow stock

To detect slow-moving goods, the record can flag items with no sales over a defined period. I would separate recent purchases, seasonal goods and products deliberately kept to fulfil infrequent orders.

In an example of 30 units with a purchase cost of 12 € each, there is 360 € of goods at cost. That is not 360 € of lost margin. Potential recovery depends on the selling price, selling costs, returns to the supplier and whether the product can still be sold.

Nor would I add that amount to supposedly lost sales and call it protected margin. To estimate an unfulfilled sale, you need evidence of demand and contribution margin, allowing for substitutions and postponed purchases. These are different calculations.

Where an agent could help

Sums and defined thresholds can be handled with rules. An agent could help read a supplier's delay notice or prepare a draft explaining the reason for replenishment. I would keep purchasing approval and amount limits outside the model's unrestricted decision-making.

I would first test a few items: counted balances, reconciled movements, reviewed alerts and orders approved by a person. I would measure discrepancies, days without availability and purchases that needed correction. I would expand only when the information and purchasing workflow are manageable.