AI in Retail: Personalization and Supply

Two reports usually land on two different desks and nobody puts them side by side: the list of items that ran out during peak hours, and the list of items that have sat in the warehouse for months. Sales sees the first, supply sees the second, yet both grow from the same root, which is a mistaken demand forecast.

Retail managed these two with experience and the memory of good buyers for years. The trouble starts once the number of stock codes passes a few hundred and a decision has to be made for every item in every branch; at that scale human memory is no longer the right instrument.

Three uses with the greatest effect

• Relevant recommendation: Showing the item that makes sense for this customer and this basket, not the best seller of the whole store.
• Demand forecasting per item and branch: A separate estimate for each combination instead of one blanket number for all outlets.
• Basket analysis: Finding items bought together and rearranging shelves and bundles accordingly.

Why recommendation and inventory are one problem

Many organizations define these as two separate projects, and that is where failure begins. If the recommender promotes an item with a three-week lead time, the result is an unhappy customer rather than extra revenue. Conversely, if the demand model knows nothing about promotions, it treats a sudden lift as noise and leaves the shelf empty. The two need a shared data source and a single decision owner.

What must be correct before any model

Output quality here depends above all on a consistent product catalogue. If one product is registered under three codes and two names, no model can add its demand up correctly. Returns, discounts and out-of-stock days must be visible in the data too; zero sales on a day when the item was unavailable does not mean zero demand, and a model that misses that distinction will lock the shortfall into its forecast permanently.

Why it matters for Iranian retail

Frequent price changes and unstable supply make intuitive estimation harder than in steadier markets. In these conditions a demand model has to separate the effect of price from the effect of season and promotion before the true order quantity becomes clear. Capital locked in slow-moving stock is also more expensive in an inflationary setting, which makes the return on this kind of project more tangible than that of many other initiatives.

The first 90 days

• Day 1 to 30: Choose one fast-moving category and clean its catalogue, merging duplicate codes and inconsistent names.
• Day 30 to 60: Build the demand model on that category and compare its forecast with the buyer's actual orders, without yet replacing the buyer's decision.
• Day 60 to 90: If forecast error beat the current method, extend to the next category and switch the recommender on for those items.

Common mistakes

• Measuring success by click-through on recommendations; the right metric is basket value and margin, not momentary attention.
• Recommending items that are unavailable or carry negative margin; the system must see stock and profitability constraints.
• Training on sales data without accounting for out-of-stock days and past discounts.
• Treating all branches alike; demand patterns in two neighbourhoods can differ completely.

Three actions for this quarter

• Start cleaning the product catalogue; it has the largest effect and requires no model at all.
• Measure your current forecast error so there is a baseline to compare against.
• Define one shared metric for sales and supply so both functions pursue the same goal.

Frequently asked questions

• How many months of sales data are needed to start?
  Seeing a seasonal pattern usually takes at least one full year, though work can begin on fast-moving items with a shorter history.
• Does a small store need these tools?
  Wherever the number of stock codes exceeds what a person can review mentally, yes. Store size is not the deciding factor; product variety is.
• How does personalization relate to customer privacy?
  Most of the value comes from purchase behaviour without identity data; write down your retention and deletion policy from day one.

Takeaway

An empty shelf and dead stock are two symptoms of one weakness: the absence of a reliable demand estimate. Fixing it starts with one category and one clear metric, and the effect shows up in inventory turnover before it appears in the sales report.

Glossary

• Stock code: The unique identifier of a product, the basis of every sales and inventory analysis.
• Demand forecast: An estimate of future sales volume from past sales, season and price.
• Basket analysis: Identifying items customers typically buy together.
• Stockout: Absence of an item at the moment of demand; lost sales that never appear in the sales report.
• Inventory turnover: How quickly stock converts into sales over a period.
• Forecast error: The gap between predicted quantity and actual sales.