In retail, the AI that pays for itself first is the one that improves product data and demand forecasting, not the one that recommends a storefront. A well-described catalog gets the product found; forecasting per SKU avoids buying what does not sell. Personalization comes later, when there is enough purchase history for the model to learn something.
Almost every conversation about AI in retail starts with personalized recommendations. It is the most visible use case and the one that least depends on a small or mid-sized retailer, because e-commerce platforms and marketplaces already embed some version of it. What the platform does not do is describe your product well or predict how much of it you will sell next month. That is where your own work begins.
Where should retail start with AI: catalog, inventory or personalization?
The short answer is: with the data that already exists and is in bad shape. In retail, that is almost always the catalog.
The Longo Parts case is the most direct example we have. The company sells auto parts on Mercado Livre across Brazil, focused on Nissan, Mitsubishi, Honda and Toyota. More than 2,000 listings had generic titles and descriptions, and that was sinking the store in the marketplace's organic search. In auto parts, the buyer searches by car, by year and by part code. If the listing does not speak that language, it does not show up.
Rewriting 2,000 listings by hand would take months and be inconsistent from one writer to the next. The solution was an LLM classifying the 2,000 products by technical line (steering, brakes, engine, suspension and transmission) and by compatibility, then rewriting titles and descriptions based on semantic analysis and Mercado Livre's search rules. The listings were updated directly on the platform, and all of them now follow the same standard.
What that case teaches about the order of things:
- Product data is the input for everything that comes after. Recommendations, site search, forecasting by category and paid ads all depend on knowing what each item is.
- An LLM solves the scale and consistency problem for text. What it does not solve on its own is the business rule: which attributes matter to the buyer in that category. That has to come from someone who knows the product.
- Results for this kind of project are measured on the marketplace: search position, visits and clicks per listing. Without a before and after, the rewrite is an opinion.
How does demand forecasting avoid overstocking?
Overstock and stockouts are the same mistake with the sign flipped: buying without knowing what will sell. Demand forecasting is the model that reduces that error, and it works best when the buyer understands how the model thinks.
A per-SKU forecasting model uses sales history, seasonality, the promotional calendar and, when available, the price charged. It returns an expected sales range for the coming weeks, per product or per category. Purchasing is then done against that range, with a safety margin set by the supplier's lead time.
Three practical rules so the first round does not disappoint:
- Start with the products that matter. A simple ABC curve points to the items that concentrate revenue. Forecasting 20% of the catalog well is worth more than forecasting 100% badly.
- Compare against the buyer's spreadsheet. The model only earns its place if it errs less than the manual estimate. Measure both errors over the same period and show the number to the team.
- Handle new products separately. With no history, the model has nothing to learn from. A new product borrows the curve of a similar item until it has data of its own.
This is classic machine learning work, not generative AI. It is cheaper to run, easier to explain and easier to audit. Many retail projects fail because they try to solve forecasting with a chat.
When does personalization make sense?
Personalization needs two things: customer behavior history and enough volume for a pattern to appear. Before that, what works is segmenting paths, not recommending products.
The Thaís Joias case shows that segmentation with no model at all. The brand sells gold-plated jewelry to end customers and recruits resellers, and both arrive at the same site with different needs. The customer needs a catalog, prices and the confidence to buy without seeing the piece. The future reseller needs to understand the program and sign up, with a profile review by the brand's team before approval. The store separated the two paths and placed the 1-year warranty against tarnishing, its main trust argument, across the whole site.
Once history exists, personalization starts with the simplest questions: how long until someone who bought earrings buys again, which piece usually goes with another, which reseller is starting to stop ordering. Recurrence and churn models answer those questions with data the store already has. Real-time recommendations on the storefront come last, and often ship with the platform already.
| Retail use case | Data needed | First result to measure | Typical horizon |
|---|---|---|---|
| Catalog rewrite with an LLM (Longo Parts: 2,000 listings, 5 technical lines) | Product sheet, attributes per category, the channel's search terms | Search position and clicks per listing, before and after | 3 to 6 week PoC |
| Demand forecasting per SKU | Sales history, promotional calendar, supplier lead time | Model error against the manual estimate over the same period | First 6 weeks for the model, 90 days for ROI |
| Audience segmentation (Thaís Joias: end customer and reseller) | Customer type at sign-up, visit source | Qualified sign-ups and conversion per path | From site launch |
| Customer recurrence and churn | Order history per customer, interval between purchases | Repurchase rate in the treated segment | 90 days |
| Storefront recommendations | Browsing, cart, cross-purchases | Revenue per session with and without recommendations | Only with traffic volume and history |
The 90-day horizon is not arbitrary. It is the period we use to judge whether an agent or model has really entered operation, and the post on how to measure the ROI of an AI agent in 90 days details the method.
Agent, model or automation?
In retail all three show up together and are easy to confuse. Rewriting a catalog is LLM work with business rules. Forecasting demand is a machine learning model. Updating the listing on the marketplace after the rewrite is automation. When the problem is repetitive and the rule is clear, automation solves it more cheaply; when it requires interpreting text or context, the agent comes in. The difference is in the post on AI agents versus RPA.
Next step
If your catalog is poorly described or purchasing still runs on gut feeling, those are the first two AI projects that usually pay for themselves in retail. The retail and e-commerce page shows the sector's cases, and the machine learning page explains how we build and measure a forecasting model.