Two years into the generative-AI wave, most retailers have tried something and most of it was a demo. Here is what we have seen deliver measurable results in real stores, and what we would not bother with yet.
Works: product content at scale
Descriptions for thousands of supplier products, translated into Latvian, Lithuanian, Estonian, Russian and English, with consistent tone and correct terminology. A model does the first draft from supplier data and images; a human reviews categories that matter. We run this as a queue job inside the store: new products get content automatically, editors only fix exceptions. Result: catalogues that used to have 40% empty descriptions are complete, and search engines have something to index.
Works: PDF and feed extraction
Suppliers still send price lists as PDFs and half-broken spreadsheets. A model that turns them into structured rows (article, name, price, stock, attributes) with a confidence score saves hours a week and catches errors humans miss when copying by hand.
Works: smarter on-site search
Semantic search understands “warm jacket for a toddler” without those exact words appearing in a product name. Combined with a classic index for exact matches and filters, it noticeably lifts conversion for stores with large or badly named catalogues.
Works, with care: support assistants
An assistant that answers “where is my order” and “what is your return policy” from your own data, and hands over to a human the moment it is unsure, reduces ticket volume. An assistant that improvises about prices or availability creates refunds and angry customers. The difference is entirely in how it is connected to your systems and how strictly it is allowed to speak.
Not yet: fully automatic pricing and buying
Dynamic pricing tools are only as good as the data behind them, and most small retailers do not have clean competitor and cost data. Start with rules you understand.
Not yet: AI-generated product photos for physical goods
Customers notice, returns increase, and marketplaces are starting to penalise it. Use AI for backgrounds and consistency, not for inventing the product.
The boring prerequisites
Every useful AI feature above depends on the same things: a clean product database, working integrations, and a store that can run background jobs reliably. Those are the projects we do first. When they are in place, adding the model is the easy part.
Curious which of these would pay off in your store? Ask us; we will tell you honestly if the answer is “none yet”.
