Almost a quarter of EU enterprises already sell online, according to Eurostat (2024 data): 18% through a website or app only, the rest via EDI or both. For most of them, the problem is no longer being online — it is operating online with a margin. Catalogues that grow faster than the team, marketplaces with different rules, customer service that doesn't scale, returns eating the profit, stock that is either missing or piling up.
It is on this ground — operations — that AI is producing results today. Not in "immersive experiences" or "hyper-individual personalisation", but in repetitive, data-heavy work that a person did slowly and AI does fast — provided someone has designed the process around it.
This article lists nine opportunities, ordered from least to most implementation effort. For each one: the problem, what AI does, what you need to have, how you know it is working and where it usually fails. It applies to a Shopify store in Lisbon, to a seller with accounts on Amazon, Worten and Fnac, and to an industrial distributor selling by catalogue to other businesses.
1. Catalogue enrichment
The problem. Incomplete titles, missing attributes, descriptions copied from the supplier, product sheets without translations. A catalogue of two thousand SKUs with weak listings is a catalogue that doesn't show up in searches and doesn't convert.
What AI does. Generates titles, structured attributes (material, dimensions, compatibilities), channel-specific descriptions and translations from the supplier sheet, an image or a URL. Amazon itself offers this to sellers: from a few words or a photo, the tool produces the full listing, and Amazon reports that more than 900,000 sellers already use it and accept the generated content with little or no editing about 90% of the time.
What you need. A source of truth for the product (PIM, spreadsheet or database) and clear rules per channel — what Amazon requires in a title is not what Zalando requires.
Sign it's working. Share of complete listings, impressions and conversion rate per listing before and after.
Where it fails. Generating descriptions in bulk without anyone reviewing the first fifty. AI invents plausible attributes (a power rating, a dimension) when the source sheet doesn't have them. Review samples and lock technical fields against free generation.
2. Feeds and marketplace compliance
The problem. Every marketplace has its own format, categories, mandatory attributes and content rules. A seller on five channels maintains five versions of the catalogue and receives rejections nobody has time to read.
What AI does. Maps categories and attributes between channels, detects the reason for rejections and proposes the fix, and checks listings against each platform's rules before sending them. Marketplace platforms such as Mirakl already integrate this kind of assisted mapping for sellers and operators.
What you need. Your current feeds and the history of rejections — that is what teaches the system what each channel accepts.
Sign it's working. Rejection rate per channel and time between creating a product and having it live on every marketplace.
Where it fails. Letting AI "fix" categories without validation: a product in the wrong category may sell, but with the wrong commission or outside the channel's rules.
3. Pricing and competitor monitoring
The problem. The right price changes every day and depends on what competitors do, the stock you hold, your minimum margin and the season. Doing this by hand for hundreds of SKUs is impossible; doing it with fixed rules leaves money on the table.
What AI does. Collects competitor prices, suggests adjustments within defined limits (minimum margin, maximum price, target position) and explains why. On marketplaces, it helps win the buy box without entering price wars nobody wins.
What you need. Real cost per SKU (including logistics and commissions), written margin rules, and a person who approves adjustments above a threshold.
Sign it's working. Margin per order, not just volume. A pricing algorithm that sells more at lower margin is failing.
Where it fails. Automating without limits. Price wars between algorithms are real and end with two sellers selling below cost.
4. Pre- and post-sales customer service
The problem. "Where is my order?", "does size M fit?", "how do I return this?" — 60 to 80% of a store's contacts are variations on a dozen questions, and the team answers each one as if it were the first.
What AI does. An agent connected to the order system, the catalogue and the returns policy answers these questions with real data (not generic text), in several languages, at any hour, and hands over to a person when it doesn't know or when the customer is angry.
What you need. Integration with order status and a written returns policy. Without access to the data, the agent is an FAQ with better grammar.
Sign it's working. Share of contacts resolved without human intervention, first-response time and — importantly — satisfaction measured on the contacts the agent resolved.
Where it fails. Switching on the agent without defining what it cannot do (promise refunds, change addresses, give discounts) and without an easy way out to a person. A customer stuck in an automated loop is a lost customer and a negative review.
5. Demand forecasting and replenishment
The problem. Too much stock in SKUs that don't move, stock-outs in the ones that do. On marketplaces, a stock-out costs twice: you lose the sale and you lose ranking position.
What AI does. Forecasts demand per SKU from sales history, seasonality, campaigns and external signals, and proposes purchase orders and transfers between warehouses.
What you need. At least a year of clean history per SKU and per channel, real supplier lead times, and the costs of holding and of stock-outs.
Sign it's working. Days of stock, stock-out rate and capital tied up, measured before and after.
Where it fails. Trusting the forecast for new SKUs or those with short history — there, AI knows no more than a simple rule. And forgetting that a forecast is an input to a decision, not the decision.
6. Returns and fraud
The problem. Returns are the hidden cost of e-commerce, and a small fraction of customers is responsible for a large fraction of them. On marketplaces with generous policies, abuse is a structural cost.
What AI does. Classifies return reasons from text and photos, detects abuse patterns (same customer, same products, same reasons) and flags products with an anomalous return rate — often due to an error in the listing, not the product.
What you need. A returns history with reasons, and a clear policy on what happens with a suspicious pattern (the decision has to be a person's).
Sign it's working. Return rate per SKU and per customer, and processing cost per return.
Where it fails. Blocking customers automatically. A false positive is a legitimate customer treated as a fraudster, and that reaches social media within minutes.
7. Reviews and customer questions
The problem. Hundreds of reviews and questions a month, across several platforms and languages, that nobody reads systematically — and that contain the best information available about the product and the listing.
What AI does. Summarises reviews by theme, detects recurring problems (wrong size, missing part, confusing instructions), suggests replies and flags the ones that need a human answer. On marketplaces, answering questions quickly affects visibility.
What you need. Access to reviews via API or export, and an owner for each type of problem detected (the listing belongs to marketing, the missing part to logistics).
Sign it's working. Response time to questions and the number of listing or product fixes originating from reviews.
Where it fails. Replying to everything automatically in the same tone. A negative review deserves a reply from a person.
8. Site search and recommendations
The problem. The store's search doesn't understand "trainers for running in the rain" or "part compatible with the 2019 model X". The customer can't find it, and leaves.
What AI does. Semantic search (it understands intent, not just words), recommendations based on the session context and not only on history, and — increasingly relevant — listings prepared to be read by AI assistants that shop on the customer's behalf.
What you need. A well-structured catalogue (point 1 again) and browsing and purchase events recorded.
Sign it's working. Rate of searches with no results, conversion from search and average order value with recommendations.
Where it fails. Putting semantic search on top of a catalogue with missing attributes. AI doesn't guess what isn't in the listing.
9. Fulfilment operations
The problem. Picking, packing, carrier selection, delivery promises and incident management — the work that decides whether the promise made on the page is kept.
What AI does. Optimises picking routes, chooses the carrier per order (cost, lead time, incident history by area), predicts delays before the customer asks and drafts the communication. In B2B, it manages delivery terms that differ by customer and by contract.
What you need. Integration with the WMS or order system and with carriers, and incident data by carrier and by area.
Sign it's working. Cost per order shipped, share of deliveries within the promised window and "where is my order" contacts per thousand orders.
Where it fails. Optimising shipping cost without looking at incidents: the cheapest carrier in an area may be the one generating the most returns and contacts.
B2B and industrial marketplaces: the same, with more attributes
The conversation about AI in online commerce is almost always about B2C. But industrial catalogues — components, tools, electrical material, spare parts — have exactly the same problems at a larger scale: tens of thousands of SKUs, technical attributes that decide the purchase, compatibilities, datasheets in PDF, prices by customer and by volume.
This is where catalogue enrichment (point 1) and search (point 8) are worth most: extracting attributes from PDF datasheets and making them searchable, answering "which part replaces reference X" with data, and preparing the catalogue for the B2B marketplaces that large distributors and retailers are opening to third-party sellers. An industrial buyer doesn't browse — they search by reference and by specification, and if they don't find it, they go elsewhere.
Where to start
If the store or seller is starting out with AI, the order that usually works is: catalogue (1) → customer service (4) → reviews (7). They are the three with the most data available, the least integration and the most visible results. Pricing (3) and forecasting (5) require clean data and written rules before any tool. Fulfilment (9) is where the margin is, but also where integration is heaviest.
In any of the nine, the rule is the same: one process, one metric defined beforehand, one person responsible and small deliveries every week. The tool is the easy part.
Frequently asked questions
Do I need a technical team to use AI in e-commerce? For points 1, 4 and 7, no: the platforms (Shopify, Amazon, the marketplace operators) already include tools that plug into what you have. For pricing, forecasting and fulfilment you need to integrate data from several systems, and there it helps to have someone who has done it before.
Will AI generate content that marketplaces penalise? Marketplaces penalise false content or content outside their rules, not AI-generated content — Amazon itself provides the tool. The risk lies in invented attributes; the fix is reviewing samples and locking technical fields.
How much does it cost? It depends on the point. Catalogue enrichment and customer service have low, predictable usage costs; what weighs is design and integration time. To estimate usage cost, see our AI cost guide and the token calculator.
Does this apply to a small store? Yes, and with an advantage: a small store can review the generated listings, tune the service agent and measure results within days. What doesn't scale is doing everything by hand from a few hundred SKUs upwards.
What about AI shopping assistants — will customers stop visiting the store? A share of purchases will go through assistants that search and compare on the customer's behalf. For those, the structured, complete product listing is the shop window — which reinforces point 1 and links to optimisation for AI engines.
Sources
Writes about applied AI, operations, GEO/SEO and how to turn companies into machines that keep running even when no one is watching.