AI in Retail: Behind the Shop Rather Than in Front of It

Operations before personalisation

Retail AI conversation centres on recommendation and personalisation, which need substantial traffic to produce a measurable return. Most retailers below that scale get far more from operational applications.

Forecasting, markdown timing and product data are unglamorous and they affect margin directly.

Demand forecasting with the caveats

Forecasting works with two to three years of clean history, promotions recorded, and stock-outs flagged so the model learns demand rather than sales.

The stock-out flag is the detail that decides whether a retail forecast is any good. Without it, the model systematically under-forecasts your best sellers, which is precisely the wrong error.

Markdown timing

Deciding when to discount and by how much is a classic optimisation problem with real money attached. Too early and margin is given away; too late and stock is worthless.

This benefits from modelling because the trade-off varies by product, season and remaining stock in ways people struggle to hold in mind across a large range.

Product data enrichment

• Descriptions generated from structured attributes, at catalogue scale
• Categorisation and attribute extraction from supplier feeds
• Duplicate detection across suppliers with different naming
• Image tagging for search and filtering

For retailers with large or frequently changing catalogues this removes a substantial manual burden, and the output is easy to check.

Returns analysis

Free-text return reasons contain patterns that nobody has time to read. Classification surfaces them: a size chart that misleads, a description that overstates, a packaging failure on one route.

Reducing returns is worth more than processing them efficiently, and the reason data is usually already being collected and ignored.

Where personalisation pays

With enough traffic and enough catalogue breadth, recommendation lifts basket value measurably. Below that, the effort is better spent on search quality, which affects every visitor rather than a segment.

Search that understands synonyms and intent is often mistaken for a personalisation problem and is cheaper and more universally useful.

Frequently asked questions