An AI chatbot for an e-shop pays off when it answers from your own records and policies, not from the model's memory. Most e-shop support is the same handful of questions: where is my order, how do I return this, is it in stock, when will it arrive. An AI agent connected to your orders and your policy pages can answer those in seconds, in Greek and in English, at any hour, and hand everything else to a person with a drafted reply. Two rules come with it: in the EU the customer must be told they are talking to an AI, and the shop remains responsible for every answer the bot gives.
What can an AI chatbot do for an e-shop?
An AI chatbot can handle the routine questions that fill an e-shop's inbox, as long as the answer exists in a record the shop controls:
- Order status: where the order is, with the courier's tracking link, read from the order itself.
- Returns and exchanges: what the policy allows and how to start, quoted from the policy page.
- Product questions: sizes, compatibility, materials and availability, from the product data.
- Delivery: costs, times and options, from the shipping settings.
- Evenings and weekends: the hours when many orders are placed and no one is at the desk.
What it should not do alone: refunds, exceptions to policy, complaints and anything involving money or a dispute. Those go to a person, with the conversation and the order details already attached. Support is often the first place AI pays back in a small business, as covered in AI use cases for small businesses.
What is the difference between a chatbot and an AI agent in customer support?
A chatbot answers questions from text. An AI agent can also look things up and act: check the order in the shop's system, read the courier status, open a return and draft a reply with the right tracking link. The difference matters in an e-shop because the most common question, "where is my order?", cannot be answered from any document. It is answered only by the live order and courier data. A bot that can only read the FAQ page will politely fail at the question customers ask most. How agents decide their next step is explained in what agentic AI is.
Why does the connection to your orders matter more than the chatbot?
The connection to your orders matters more because it decides what the bot can answer, and in Greece it is usually the real project. When I audited 505 Greek online shops in August 2026, 16.0% carried a WooCommerce signature and 70.5% showed no recognisable commerce platform at all, which means a custom build or a platform that hides its fingerprint (The Greek e-shop technical audit). A WooCommerce shop exposes orders through a documented interface (WooCommerce REST API). A custom platform needs that connection built, along with the courier's tracking. Scope that part first, read-only, before choosing any chatbot tool.
What do real deployments show?
Real deployments show large gains on routine volume and a clear limit on quality. In February 2024 Klarna reported that its AI assistant had handled 2.3 million conversations in its first month, two-thirds of its customer service chats, the work of 700 full-time agents, with issues resolved in under 2 minutes instead of 11 (Klarna, 27 February 2024). In May 2025 its chief executive said the company would invest again in human support and that customers must always be able to reach a person (CX Dive, 9 May 2025). A field study of 5,179 support agents given a generative assistant found 14% more issues resolved per hour on average and 34% more for the least experienced (Brynjolfsson, Li and Raymond). The pattern: automate the routine, keep a person within reach, and measure quality, not only volume.
How do you stop an e-shop chatbot from giving wrong answers?
You stop wrong answers by never letting the bot answer from memory. In February 2024 a Canadian tribunal held Air Canada liable for its website chatbot's wrong advice on bereavement fares and rejected the argument that the chatbot was responsible for its own words (Moffatt v. Air Canada). The same logic applies to an e-shop that promises a return window or a delivery date it does not offer. The safeguards:
- answers about orders come from the order record, and answers about policy quote the policy page, retrieved at the moment of the question as described in RAG explained;
- every reply is checked against its source before it is shown, so an order number, a date or a tracking link in the reply must match the record;
- when no source supports an answer, the bot says so and passes the question to a person;
- a set of real past questions is re-run after every change, as described in why AI hallucinates.
How safe is an AI agent that reads customer messages?
An AI agent that reads customer messages is safe only if it treats every message as untrusted text and personal data. Anyone can type instructions into a contact form, which is why prompt injection is the first security question for a support agent, and messages carry names, addresses and order details covered by GDPR. My running support demo, on synthetic data, shows the design: it removes personal details before storing a message, routed all 156 test messages as planted, caught all 10 planted injection attempts while letting through the customer who wrote "ignore my previous email", and sent nothing on its own, because it has no way to send (the demos). A person approves what goes out, as described in human in the loop.
What does the law require from an e-shop chatbot?
The law requires three things from an e-shop chatbot in Greece. Under Article 50 of the EU AI Act, in force since 2 August 2026, customers must be told they are talking to an AI unless it is obvious, at the latest at the first interaction (the EU AI Act in Greece). Under GDPR, the messages are personal data, so you need a legal basis, a processing agreement with the AI provider and a retention limit. And under consumer law, what the bot says about returns, withdrawal and delivery is what the shop says, so it must match your terms exactly.
How do you start with AI customer support in an e-shop?
You start small and measured:
- take the last three months of messages and count them by type;
- pick the three to five types that a record can answer, usually order status, returns and delivery;
- connect the orders and the courier tracking, read-only;
- run the agent in draft mode for a few weeks, with a person sending every reply, and measure the share of drafts sent unchanged, the time to first response and the hand-off rate;
- let it reply directly only on the types with a measured record, and keep a visible way to reach a person.
The full service is described in AI for customer support.
Frequently asked questions
How much does an AI chatbot for an e-shop cost?
The model itself usually costs around a cent or less per reply at current prices. Most of the cost is building the connection to orders and couriers, the testing and the first weeks of supervision. The running costs are covered in what an AI agent costs to run.
Can the chatbot understand Greeklish?
Current large models read Greeklish well, but test it: many Greek customers write in Latin characters, mix in English words or write without accents. Put real messages of every kind into the test set, as covered in RAG in Greek.
Should the chatbot replace my support staff?
No. It should take the repetitive questions so your people have time for complaints, exceptions and the customers you want to keep. Klarna's reversal shows the cost of removing the person entirely.
If you run an e-shop, send me the ten questions your customers ask most, without any customer data. Which of them an agent can answer is usually clear from where the answer lives.