How to use AI in retail and e-commerce: where to start
In retail and e-commerce, AI helps with routine order and returns questions, product pages built from supplier data, supplier invoices and review analysis. A good first step is usually an assistant for your support agents, not a chatbot left alone with customers. A person approves what reaches the customer.
When people talk about AI in retail, almost everyone thinks of a chatbot. Customer support matters, but it is only one part. An online shop or a multichannel retailer also has the product catalogue, supplier documents, reviews and stock, and just as much time is lost there. This guide goes through all of them and tells you where to start.
How do you use AI in retail and e-commerce?
You use AI on repetitive work with text and data: answers to customer questions from your own policies and orders, product pages from supplier data, invoices and delivery notes checked against the order, reviews grouped by theme. A person approves what reaches the customer or the accounts. The table sums up the common problems.
| Problem | What AI can do today | What it does not do well |
|---|---|---|
| Questions about orders, delivery, returns | Answers from your policies and order system, with a clear handover | Does not resolve a difficult complaint on its own |
| The product catalogue | Writes product pages and fills attributes from supplier data, translates | Does not invent specifications the supplier did not give |
| Supplier invoices, delivery notes and price lists | Extracts the data and checks it against the purchase order | Cannot reliably read a poor scan |
| Reviews, returns and feedback | Groups reasons by theme and product, flags new problems | Does not replace the conversation with the supplier |
| Stock and demand forecasting | Helps forecast if you have clean history | Cannot guess a season it has never seen |
1. Customer questions: orders, delivery, returns
This is usually where most messages are, and many repeat: where is my order, how do I return it, when is the courier coming. An AI agent connected to your policies and order system answers the routine ones and passes on the difficult ones with full context. We cover the implementation in how to deploy an AI support agent, and the advisory view is on Vlad Tudor's page on AI for retail and customer support.
What it cannot do: stand in for good people. Klarna announced in 2024 that AI was doing the work of hundreds of agents; in 2025 its CEO said "We went too far" and the company started hiring people back into support (Forbes, May 2025). The lesson is simple: measure the quality of answers, not only how many tickets the system closes.
2. The product catalogue
Every new product needs a title, description, attributes, a category and sometimes a translation for other markets or marketplaces. The data comes from suppliers in files that all look different. AI can read the supplier sheet, fill the attributes in your shop's structure, suggest the category and write a description in your brand's tone. A person reviews it before publishing.
What it cannot do: know what the supplier did not write. If dimensions or composition are missing, the system must mark them as missing, not invent them. For products with strict rules (cosmetics, supplements, food, toys), claims are checked by someone who knows the rules.
3. Supplier documents
Invoices, delivery notes and price lists arrive from dozens or hundreds of suppliers. AI extracts the data, checks it against the purchase order and the goods receipt, and flags differences: a missing quantity, a changed price, a substituted product. In Romania, the RO e-Factura XML is a good source of truth for invoices. The pipeline is described in how to automate document processing with AI.
4. Reviews, returns and feedback
Reviews and return reasons tell you what is wrong with a product, but nobody has time to read them all. AI groups them by theme (size, quality, packaging, delivery), by product and by supplier, and flags when something new appears. The buying team gets a short list, not an export with thousands of rows.
5. Stock and demand forecasting
Demand forecasting is a classic machine-learning case, not a chatbot one. It works if you have clean sales history over at least a few seasons, with promotions and stock-outs marked. Without history, any model is guessing. If the data is not ready, start there: how to get your data ready for AI.
What should a retailer automate first?
It depends on where you lose the most time, but in most companies the order is this:
- Routine customer questions, starting with an assistant for your agents rather than straight to customers. Agents check every answer, and you see the quality before letting it speak on its own.
- Product pages, if you launch many new products or sell on several marketplaces.
- Supplier invoices and delivery notes, if goods-in and accounts check them by hand.
- Review and returns analysis, once the first flows work.
Why an assistant for agents is a good first step: the only peer-reviewed field study on the subject found that access to an AI assistant raised support agents' productivity by 14% on average and by 34% for novice and less experienced agents, with minimal effect on the most experienced (Brynjolfsson, Li and Raymond, Quarterly Journal of Economics, 2025; United States, 5,179 agents, 2020–21 data). For you, the real figure is the one measured in your own pilot.
A hypothetical example: an online shop with a small support team would start with an assistant that drafts answers to returns questions from the returns policy and the order. For a month, agents approve or correct every answer. Then you compare time per ticket and the share of answers that needed correcting.
Data and compliance: GDPR, the AI Act, consumer protection
A few things to know. This is not legal advice; for your case, talk to your lawyer or DPO.
- GDPR: order history, addresses and conversations are personal data. Personalisation based on behaviour is profiling and needs a legal basis and clear information for customers. More in how to stay GDPR-compliant when using AI.
- AI Act: an assistant that talks directly to customers has to say it is AI (Article 50 of Regulation (EU) 2024/1689, applicable since 2 August 2026). AI-generated synthetic content (images, video, audio) must be marked; for systems on the market before 2 August 2026, the grace period ends on 2 December 2026.
- Consumer protection: the agent must not say anything other than your terms and the law. Whether your returns policy is more generous than the 14-day statutory withdrawal right or not, the answer has to come from the right text, with the source behind it.
How do I start?
- Tech Call, free: a short conversation where you show us where the team loses the most time and we tell you honestly what is worth automating. Book it here.
- Tech Audit: a short, paid audit on real tickets, product sheets or invoices and on your systems (shop platform, ERP, helpdesk). You get in writing the use case, the data it needs and what "working" means in numbers.
- Pilot: one flow, with your people approving every result.
- Production: integration with the shop platform, ERP and helpdesk, monitoring, and a handover to a person. The code and documentation stay with you.
What we can show: we delivered an AI customer-support agent for retail, integrated with ChatGPT, for a client that went on to raise a $4.7M seed round. What we build is on the AI agents and AI automation pages.
Sources
- Brynjolfsson, Li and Raymond, "Generative AI at Work", Quarterly Journal of Economics 140(2), 2025 (United States, deployment data 2020–21)
- Forbes, "Klarna reverses on AI, says customers like talking to people", 18 May 2025
- Regulation (EU) 2024/1689 (AI Act), Article 50
Access to an AI assistant raised support agents' productivity by 14% on average and by 34% for novice agents, with minimal effect on the most experienced (Brynjolfsson, Li and Raymond, QJE 2025; United States, 5,179 agents).
Frequently asked questions
Is an AI chatbot worth it for a small online shop?
It depends on volume. If you get few messages, a ready-made widget configured on your policies may be enough. A custom agent makes sense when you have many repetitive questions, order and returns systems it has to connect to, and a need to control exactly what it says.
Can AI write product descriptions?
Yes, from supplier data and in your brand's tone, including other languages. It cannot know what the supplier did not write, so missing specifications must be marked, not invented. For cosmetics, supplements, food or toys, claims are checked by someone who knows the rules.
How much does an AI agent cut ticket volume?
We do not publish a percentage: the deflection rates in circulation come from vendors measuring their own products. The only peer-reviewed study found +14% productivity on average and +34% for novice agents (United States, 2020–21). Your real figure is measured in the pilot, on quality and time.
Do I have to tell customers they are talking to AI?
Yes. Since 2 August 2026, Article 50 of the AI Act requires people to be informed when they interact directly with an AI system, unless it is obvious. Synthetic content, such as generated images or video, must be marked too. This is not legal advice.
Can AI forecast my stock?
Yes, if you have clean sales history over several seasons, with promotions and stock-outs marked. Forecasting is a classic machine-learning case, not a chatbot one. Without good history, any model is guessing, and the first project is getting the data ready.
Want to discuss a project?
Book a free discovery call with the Sapio team.