E-shop
Chatbot for online shops: 24/7 AI support that sells
What an AI chatbot for an online shop can really do, where its usefulness stops and how to set it up so it answers correctly out of hours.

Most of the questions an online shop receives are not difficult. They are the same ones, many times a day, across four channels at once: “when will it arrive?”, “do you have it in a size 8?”, “can I pay with Klarna?”, “do you deliver to Northern Ireland?”. The problem is not the answer. It is that someone has to give it at eleven at night, when the customer is sitting with their phone in hand and a half full basket.
A chatbot for an online shop is worth it when it does two things: it answers product and policy questions correctly, however customers type them and whatever time they arrive, and it turns the answer into a buy link inside the conversation. The “24/7” refers to the assistant. Anything that needs a person waits until the morning, and the bot has to say so.
What a chatbot answers well and what it does not
An AI chatbot with access to the shop's catalogue and policies answers reliably in three categories:
- Product availability and features: sizes, colours, materials, compatibility.
- Policies: delivery times and costs, returns, payment methods, VAT invoices.
- Help choosing: “I'm looking for a present for an 8 year old, up to £30”.
It does not answer well on anything that needs judgement or access to systems you have not given it: complaints about a damaged parcel, exceptions to the returns policy, questions about an order that is not connected to the order system. In these cases the right thing is to collect the details and hand over to a person, not to improvise.
The quality of your data decides the quality of the answers
The chatbot is only as good as the catalogue behind it. If product descriptions are two words long and the returns policy exists only in the owner's head, the bot will answer vaguely or wrongly. Before you switch anything on, write down:
- One page with the ten most common questions and the exact answers.
- Your delivery and returns policies in plain English, with numbers (days, pounds, limits).
- What the bot must never promise, for example a delivery date to the Scottish Highlands and islands.
Search by photo and voice
On a phone, typing is the obstacle. A customer who saw a shoe on a friend and has a photo, or who wants to say “I want the same in black”, will not type a product code. Newer tools accept an image or a voice message and return the closest products in the catalogue, with a direct checkout link. This is the point where a chatbot stops being just support and starts working as a salesperson.
Language, the way customers actually write
Customers write with typos, without punctuation, in abbreviations and sometimes in a mix of languages. Test the bot with real messages from your Messenger before you put it live. Also ask for it to reply in the same tone your shop uses: an online shop for children's clothes does not talk like one that sells tools.
Where to be careful legally
The bot will collect names, phone numbers and emails. You need to tell people how their data is processed and, if you plan to send offers later, get explicit, recorded consent. A conversation about an order is not, on its own, consent for marketing. Keep that question separate and store the answer with a date.
How to measure whether it is worth it
Not with the percentages suppliers promise. Measure two things in the first month: how many conversations closed without a person stepping in, and how many ended in an order within the same session. If the second is zero, the bot answers but does not sell, and that is usually fixed with better product data, not a better model.
Step by step: setting it up on an online shop
- Connect the catalogue through a feed or API, so prices and stock update by themselves. A bot with a copy of last month's catalogue sells products you do not have.
- Upload the three pages from the previous section: common questions, policies with numbers, forbidden promises.
- Set the handover rules: a complaint, a damaged parcel, a request for an exception, a question about a specific order with no connection to the order system. In all of these the bot collects the name, order number and description, and says when a person will reply.
- Start with web chat, where you control the environment, and add Messenger, Instagram and WhatsApp after two weeks.
- Test with 30 real messages from your history before you open it to customers. The ones answered wrongly show what is missing from the data, not from the model.
- Decide who reads the conversations, every day in the first month and every week after that.
Hypothetical example: an online shoe shop
A hypothetical example with invented numbers: an online shoe shop receives 400 messages a month across four channels. 250 are “do you have it in a size 8”, “when will it arrive” and “can I exchange for another size”. 150 of the 400 arrive after 21:00 or at the weekend and are currently answered on the next working day.
With the bot connected to stock, the 250 routine messages are answered instantly, and every availability answer comes with a link to that exact size. If only 15 of the 150 evening messages become orders that would otherwise have been lost, with an average basket of £70 that is £1,050 a month. At the same time, the team now handles 150 messages instead of 400: the 150 that really needed a person. Your own number comes from your own messages: count them for a month before you decide.
Common mistakes
- Giving a delivery date. “2 to 4 working days, depending on the area” is an answer. “Thursday” is a promise you do not control.
- Not connecting the order system. “Where is my order?” is the most common question after a purchase. If the bot cannot see it, the customer writes in again by email.
- Pretending to be a person. A name and the label “assistant” in the first line are enough. A customer who works it out on their own loses trust in the shop.
- Asking for an email “to continue”. The conversation has to answer first. Consent for offers comes afterwards, separately.
When not to automate
If the online shop receives fewer than 50 messages a month, a person with a phone notification covers the flow better and more cheaply. If the catalogue does not have correct sizes and stock, fix that first: a bot built on wrong data gives wrong answers with confidence. And if most of your messages are complaints, the problem is in delivery or the product, not in customer service.
With PandoraBot
SalesBot answers your customers on your website, WhatsApp and Instagram, recommends products and hands the conversation to a human when needed. See how it works and pricing, or start with a free scan of your website.
FAQs
Will the chatbot replace customer service?
No. It will take over the repetitive questions and the hours when the shop is closed. Complaints and exceptions have to reach a person, and the bot needs to know when to hand over.
How long does it take to set up?
The technical connection to the catalogue is the easy part. The time goes into writing the answers and policies properly: allow a few hours of your own work before you switch it on.
What happens if the bot gives a wrong answer?
It will happen, especially in the first weeks. Read the conversations every day at the start, fix the data that caused the mistake, and define explicitly the topics where the bot must refer to a person.
What should I measure at 30 and 90 days?
At 30 days: the percentage of conversations that closed without a person, and orders placed through the conversation. At 90 days: how many messages per order the team now receives compared with before, and the return rate on orders closed by the bot. If it is higher, something is missing from the product descriptions.
Do I need a separate bot for each channel?
No. The same catalogue and the same policies feed web, Messenger, Instagram and WhatsApp. What changes per channel is the length of the answers, not the content.



