AI for e-commerce companies with 20–200 employees – from customer service and returns to demand forecasting

A typical e-commerce company with a few dozen employees sells through several channels at once: its own online shop, the Polish marketplace Allegro and other marketplaces. Orders flow into a single system, such as Base (formerly BaseLinker), and from there into the warehouse and accounting software.

Three areas are usually under the most pressure. Customer service answers the same questions about delivery status hundreds of times. The returns team checks requests by hand and keeps track of deadlines. The buyer decides on supplier orders based on a spreadsheet and gut feeling.

This is where AI delivers the most, because the tasks are numerous, repetitive and easy to measure. The table below shows the order in which it is usually worth implementing them and what each one requires.

Most questions to an online shop concern a few topics: where is my parcel, is the product in stock, how do I return an item, when will I get my money back? A chatbot that only knows the FAQ page will give a vague answer and pass the customer to an agent. An assistant connected to the order management system will give the actual status and tracking number.

Calculate the benefit using your own data. Suppose, hypothetically, that the shop receives 1,500 enquiries a week and the assistant closes 40% of them without human involvement. At 4 minutes of agent time per case, that is 600 × 4 = 2,400 minutes, or 40 hours a week.

The assistant must recognise when to hand the conversation to a person: for a complaint, an unusual case or an upset customer. From 2 August 2026, it must also tell customers that they are talking to AI. The same mechanism can handle phone calls as a voice assistant (a voicebot), but you usually start with chat and email.

Complaints and returns take up a lot of time, because each case means checking the order, photos, deadlines and conditions. Then there are the rules: a consumer complaint must be answered within 14 days, otherwise it is deemed accepted. For a return, the shop has 14 days from receiving the customer’s withdrawal notice to refund the money.

AI can identify the type of request, ask the customer for missing photos or the order number, check whether it is within the deadline and draft a reply. The employee receives a ready case and simply approves or changes the decision. The system also tracks deadlines and warns about cases that are running out of time.

Do not hand AI the decision to reject a complaint. A wrongful rejection means a dispute with the customer, a bad review on the marketplace and sometimes intervention by a consumer ombudsman or a court case.

Listing a product across several channels involves a lot of small tasks: description, specifications, category, an offer title that complies with each platform’s rules. AI can prepare versions of the description for your shop, Allegro and other platforms from a single product record. It can also suggest replies to buyers’ questions.

There is one important caveat. If the input data is incomplete, AI may add a feature the product does not have. That is why descriptions should be generated only from the data in the product record, and dimensions, composition and technical specifications must be checked before publication.

Do not use AI to create product reviews. Publishing fake customer reviews is a prohibited commercial practice. You can, however, use AI to draft replies to genuine reviews, which an employee then approves.

Excess stock ties up cash and ends in a low-margin clearance sale. Running out of a bestseller at the peak of the season means lost sales and a weaker position for your offers on marketplaces. An AI model can predict sales based on history, seasonality, planned promotions and manufacturers’ delivery times.

Do not start with the whole range. In most shops, a small proportion of products accounts for a large share of sales, so build the first model for those. The other products can run on simple rules, such as a minimum stock level, for a long time.

The data requires the most work. You need at least a year of history, or two years where seasonality is strong, with promotions and stock-out periods flagged. If the data does not show that a product was not selling because it was out of stock, the model will conclude that nobody wanted it.

Dynamic pricing means automatically changing prices depending on demand, stock levels and competitors’ prices. Personalisation means matching products, content and offers to an individual customer. Both can improve margin and conversion, but they require good data and legal caution.

Announcing a price reduction requires showing the lowest price from the 30 days before the reduction, so the pricing system must keep a price history. If a price is tailored to an individual customer on the basis of automated decision-making, the shop must inform the customer. It is worth consulting a lawyer on the rules before implementation.

Personalisation only makes sense once the customer database is in order and the email marketing system is connected to order data. Without that, recommendations will be random, and customers will get offers for a product they have just bought.

In many online shops, the management board receives a sales summary several days late, pieced together by hand from several channels. When data from the marketplaces, the warehouse and accounting flows into one place, such as Power BI, the board sees sales, margin and returns by channel and category in real time.

AI adds the ability to ask questions in plain language, such as “which categories had the most returns last month?”. It can also draft a commentary on the results. The precondition is the same as everywhere else: the data must be consistent, and terms such as margin must be calculated in one agreed way.

A safe sequence is four stages, with an assessment of results after each one. First, customer service connected to the order management system. Then complaints and returns, which use the same connections, and product descriptions. Next, demand forecasting for the most important products. Finally, reporting, pricing and personalisation.

The cost is made up of licences for AI tools, the work of connecting to the order management system, warehouse and marketplaces, and data preparation. On top of that comes your team’s time spent describing the rules and testing. Before each stage, measure the current state so that after three months you can compare the result with your assumptions.