What AI actually does in a shop, today

"AI in retail" conjures robots and revolution. The reality in an actual shop is quieter and more useful: software that reads more than a human can, forgets nothing, and interrupts at the right moments. Here is the honest inventory of what that means today, from someone who builds it.

What works, now

Pattern reading at scale. Demand curves per product, each customer's personal rhythm, price responses buried in years of till data. Nothing a diligent analyst couldn't do — for one product, for one week. AI's edge is doing it for everything, always. This is the workhorse, and it is thoroughly real.

Drift catching. The slow changes human attention structurally misses: margin eroding a point a quarter, a bestseller's velocity fading, a forecast that always guesses high. Machines are unreasonably good at boring vigilance, which is most of retail vigilance.

Decision challenge. The newest and, we would argue, most valuable layer: at the moment of a decision — this reorder, this discount, this rota — projecting its consequences and pushing back with your own history as evidence. Not "here is a chart" but "the last three times you over-ordered this category, here is what it cost".

Drafting, not doing. Modern AI writes the reorder, the campaign, the transfer note — competently. What it should not do is send them, and we have written about why.

What is mostly marketing (for a small shop, in 2026)

Fully autonomous ordering with no human gate. Computer-vision shelf monitoring at small-shop prices. Chatbots as a substitute for a shopkeeper's knowledge. Each is either economically premature at your scale or an accountability problem wearing an innovation costume.

The one-question filter

Ask any AI vendor — us included — "what specific decision did your last user change because of this, and can you prove the outcome?" Real operational AI has specific, checkable answers. Theatre has adjectives. The technology is genuinely useful; the filter is what keeps you buying the useful part.