In retail, the calendar gives AI projects their first constraint: a forecast is checked at the end of a replenishment cycle, a markdown at the end of a season, a recommendation as soon as an online trial has gathered enough visits. AI in retail thus covers use cases separated by their data needs, their time to proof and their effect on the customer.
This article compares five of them (demand forecasting, recommendation, pricing, customer service and logistics) on the data each requires, the time before a verifiable result and the risk to the customer. Forecasting and logistics work first on items and stock; recommendation and pricing affect the person, through their data or through the price they pay. The choice of these criteria is a proposed method.
AI use cases in retail, side by side
By verifiable result, we mean a result compared with an observed fact (a sale, a lead time, an expected answer) by someone who did not build the system.
| Use case | Data required | Time before a verifiable result | Risk to the customer |
|---|---|---|---|
| Demand forecasting | Sales by item, location and period; stock and stock-outs; prices, promotions, commercial calendar | One replenishment cycle: the forecast is compared with actuals at the order horizon | Indirect: stock-out or item missing from the shelf |
| Recommendation | Browsing and purchase history; catalogue and attributes; available stock | Short online: a comparative trial concludes once traffic is sufficient | Use of personal data; item recommended and then unavailable |
| Pricing | Price and sales history; stock; costs; prices recorded at competitors | One campaign or season for a markdown; shorter for a price adjusted online | Direct: the price paid, and the feeling of unequal treatment |
| Customer service | Terms of sale, returns policy, product sheets; the customer’s orders | Before going live, on a set of questions with known answers; then in production | Inaccurate answer given in the retailer’s name; exchange with a machine not disclosed |
| Logistics | Stock by location; orders; warehouse and transport capacity | A period of operation compared with a reference period | Indirect: delivery promise not kept |
The time follows the cycle of the decision the system informs. An online recommendation or pricing trial can be repeated several times in a season; an end-of-collection markdown can only be judged once per collection. The measures specific to forecasting (error, bias, service level) are detailed in our article on demand forecasting; we do not return to them here.
Shared data, and the data each case adds
Reading down the columns reveals a foundation: the product master data, the sales history and the stock serve several cases at once. Content-based recommendation depends on catalogue attributes. A forecast inherits the gaps in the sales history. An assistant returns the product sheet as it is maintained.
Making this foundation reliable falls under data quality and, for master data shared across channels, under master data management. The same master data would feed a digital product passport, which Regulation (EU) 2024/1781 only makes mandatory for a product group by means of a delegated act.
Each case then adds its own data. Recommendation and the personalisation of customer service require data about people, absent from a forecast built at item and store level. Pricing requires costs and competitor prices, whose collection raises its own questions. Logistics requires warehouse and transport data, which live outside retail tools.
A forecasting, recommendation or pricing model can degrade when the assortment, prices or behaviours change. Putting it into production requires retraining and drift monitoring, a chain described on our MLOps and LLMOps page.
The risk to the customer grows closer to the person and the price
Two cases affect the customer without an intermediary: recommendation, through their data, and pricing, through what they pay. We assert no rule of consumer law here. These points are framed as questions to put to the lawyers and the data protection officer, before the trial.
For recommendation and personalisation:
- on what legal basis does the processing rest, use by use, and how does a customer’s objection propagate to all tools?
- must the criteria that rank the products presented be brought to the customer’s attention, and in what form?
For pricing:
- can a price differ from one customer to another, and on what criteria?
- must the customer be informed when a price has been personalised by automated processing?
- how is the reference price of a discount announcement established when the price has varied in the preceding weeks?
Our article GDPR and artificial intelligence reads the articles of the regulation that these questions encounter, including Article 22 on automated individual decision-making, without qualifying any case.
For customer service, the risk concerns what the assistant asserts: an inaccurate returns condition commits the retailer’s image like any other publication. Our generative AI page describes the evaluation on a test set built with domain experts, which precedes going live. Informing a customer who is interacting with an AI system falls under Article 50, paragraph 1, of Regulation (EU) 2024/1689, presented in our entry on the AI Act.
The first case to take on, based on what the retailer already has
We propose starting from what the retailer already possesses.
- A clean sales history and accurate stock, without a unified customer view: demand forecasting can be taken on without personal data, at the cost of slower proof.
- Sustained online traffic and a catalogue with populated attributes: recommendation allows short trials, once the personal data questions have been examined.
- Ends of season laden with unsold stock: markdown pricing targets the problem, but its proof waits for the end of the season and its legal questions are examined before any trial.
- A customer service team taken up by repetitive requests and documentation kept up to date: the assistant is checked before being exposed to the customer.
- Missed delivery promises even though stock is accurate: logistics is measured on day-to-day operations, against a reference period.
In each situation, a case is set aside when its foundation is missing. A recommendation connected to inaccurate stock suggests absent items; an assistant connected to a returns policy that each channel writes in its own way answers according to whichever version it finds.
The trial itself is placed on a real, reduced scope (a product category, a group of stores, a country) and away from peak periods, as described on our retail and luxury page: a change delivered too close to a busy period waits for the next one.