Retail AI wins at the shelf edge: demand, price, and the next best offer
AI in retail & e-commerce
The question: Where does AI actually move revenue and margin in retail — and where does it quietly destroy trust?
Retail is AI's oldest proving ground: forecasting, pricing, recommendations, service. It is also where the traps live in the wild — cannibalisation, promo confounding, personalisation creep — and you have already met its datasets in the labs.
What the lesson covers
Demand forecasting is the backbone: better forecasts cascade into less stock-out, less markdown, less waste. The craft is granularity (SKU × store × week) and honest error measurement against a naive baseline — a model that cannot beat "same as last year, seasonally adjusted" is decoration. Promotions, weather and cannibalisation (a promoted cola steals from the unpromoted one) are the hard parts.
Pricing and promotion: elasticity varies wildly by category, and the observational trap from your pricing lab is everywhere — best-sellers get the displays, so naive analysis overstates promo lift. Retail leaders run continuous experiments (which the data you analysed could not replace) and guard margin with floors and competitor-aware rules.
Recommendations and search are the revenue surface: "customers also bought", semantic product search, and increasingly conversational shopping assistants grounded in the actual catalogue (RAG over products — with retrieval-time stock and price checks so the bot never sells what is not there).
Service and returns close the loop: the Klarna arc taught the hybrid pattern; returns fraud and sizing prediction are quiet money-savers. And personalisation runs on the trust budget from Lesson 12 — the "would we say it to their face?" test decides which segments and triggers are legitimate.
Underneath all four of those sits one operational truth that decides whether any of it survives contact with a store: retail AI runs on data that is generated by people under time pressure. A shelf count is taken by somebody at the end of a shift, a return reason is chosen from a dropdown by a cashier with a queue, a promotion is logged by whoever remembered. That is not a criticism of the people — it is the process, and it means the input distribution is systematically different from what an analyst imagines. The practical consequence is that any retail model needs a data-quality gate before it needs an accuracy target, and the gate has to be checked against the process rather than against the table.
Key points
- Forecast quality is measured against a naive seasonal baseline — at SKU × store granularity.
- Promo analytics without experiments overstate lift (displays go to best-sellers; cannibalisation and pull-forward hide).
- Catalogue-grounded assistants need retrieval-time stock/price checks.
- Personalisation spends the trust budget — segment triggers must pass the face test.
- Retail data is generated by people under time pressure — shelf counts at shift end, return reasons picked with a queue waiting. A **data-quality gate tied to the process** comes before any accuracy target.
Framework — Retail AI value chain
forecast → buy/allocate → price/promote → recommend/search → serve/returns. Diagnose any retail AI pitch by which link it strengthens and which metric (stock-out %, markdown %, conversion, AOV, NPS) proves it.
The lab
Apply the value chain to real retail data you already have.
Open this lesson, its lab and its quiz
Sources and further reading
- The State of AI — consumer/retail cuts — McKinsey
- Dunder Mifflin & Pricing labs — this academy