In e-commerce a customer chatbot is the most common first AI deployment and the most common first failure. Not because it is bad. Because it is first.
I say this as someone who sells online shops exactly that: a chat with an AI agent. And still I advise starting elsewhere.
The Polish market is mature: 78 in 100 internet users buy online, and 53 in 100 use AI tools in the buying process, 19 in 100 often. But when a customer wants to contact the shop, they choose e-mail (31 in 100), phone (26) and chat with a human (17). Eight in ten people rate contact with a chatbot positively or neutrally. And still choose a human. That is the most important hint from the Gemius study: at the start, AI in a shop should work in the background, on content and data, not in front, with the customer.
Behind the scenes lie the processes with the biggest volume and the smallest risk:
- Product descriptions. The most common and safest use. Vendors report conversion increases from over ten to over twenty percent after replacing manufacturer descriptions with their own content, and a two-thirds reduction in content production time. That is data from the best deployments, so treat it as an upper bound. But even half of that is a lot at 500 SKUs.
- Catalog translations when entering new markets. The same process, a different language, a human checks the terminology.
- Categorization, tagging and normalization of supplier feeds. Attribute mapping, extracting features from catalog sheets. Pure operational work, zero customer contact.
- Classification of return and complaint reasons. It gives knowledge about the product, not just savings.
- Suggestions for the consultant. AI proposes a reply, a human sends it.
- A customer chat on the website. Only when there is a mature knowledge base and integrations.
Why the chat comes last. Because "where is my parcel" requires an integration with the courier's status, and "can I return it" requires the current terms in the knowledge base. Without that the bot answers nicely and off topic. Independent measurements say the bot handles 10 to 25 percent of cases on its own in a typical deployment, a median of 41. The rest goes to a human. NVIDIA's survey, in which 87 in 100 retailers declare revenue growth thanks to AI, is a technology vendor's survey.
So why do we sell chat to shops? Because point six looks different once you have gone through points one to five. Then the knowledge base is current, descriptions are consistent, and the agent passes an exam of 100 questions about delivery, returns and availability before it speaks to a customer. And when a customer writes "do you have this in size 44?", the agent sees the page context, and when they write "I want to order 40 units on invoice", it hands the conversation to a human.
Limitation: the order above assumes you have a catalog with hundreds of SKUs. A shop with twenty made-to-order products has no volume in descriptions, but it has volume in customer questions. Then the chat can be first, provided a human is in the loop.
Check for yourself: open the shop's inbox and count how many of the last 50 customer e-mails are questions whose answer is on the website.
Want to see how the agent handles your customers' questions? Enter your website address and e-mail in the demo form. In 20 minutes you get an agent trained on your website's offer, ready to be quizzed.
Sources
- Gemius, E-commerce in Poland 2025 (CAWI, n=1,629, July 2025)
- NVIDIA, State of AI in Retail and CPG 2025 (vendor survey): https://www.nvidia.com/en-gb/lp/industries/state-of-ai-in-retail-and-cpg/
- Zendesk ProductLab 2025; Zendesk CX Trends, Salesforce, Forrester 2025 (deflection benchmarks)
- Case studies from Shopify, Kensium, PageFly, Migros, Wayfair (vendor and integrator data, inflated)