What an AI chatbot in a store is
An old scripted bot walks the customer through answers that someone wrote in advance. For simple things that is enough, because the customer clicks “delivery” and gets the shipping rates or the right contact person. It gets worse with a question outside the script. The conversation does not have to end there, but the bot has only what its author foresaw.
An AI chatbot runs on a language model, so the customer can write in their own words: “looking for a warm jacket for a child for autumn, under PLN 150”. The model alone knows nothing about your shop, though. That is why a store bot starts by searching the catalogue and a knowledge base with your delivery and returns rules, and only then builds its answer from what it found there.
Judge understanding of the question and accuracy of the product separately. A bot can read perfectly well that the customer wants a jacket under PLN 150 and still pick a model that does not fit, simply because the catalogue description was thin. In CartBot the instructions forbid the bot from guessing parameters it did not find in the data. That lowers the risk of a mistake but does not remove it. If the jacket listing says only “for any weather”, the bot has no basis to promise it will stay dry in a downpour.
What a chatbot for an online store does in a conversation
In the chat the customer sees product cards with a photo and a price, and CartBot adds the chosen product straight to the store’s cart on WooCommerce and Shopify, as well as on PrestaShop and Shoper. Variants are another matter. Much depends on the integration and the theme, and sometimes the customer still picks the size or colour on the product page.
The second kind of question is about store rules, meaning delivery and returns, and the bot takes those answers from your knowledge base. It is good if the bot tells packing time apart from the time a parcel spends with the carrier, because “we ship within 24 hours” says nothing about whether the order will arrive by Saturday.
One more thing is easy to overlook when choosing. Conversations stay in the panel, and after a few weeks of reading them you can see what people cannot find in your shop, from a missing dimension on a product card to a returns rule that someone keeps asking about.
What the bot will not handle for you
A disputed complaint or a request for an exception to your terms is a job for a person who checks the circumstances and takes responsibility for the decision. In CartBot the operator sees the conversation history and takes it over in the panel, so the customer does not have to explain everything again.
It is harder when nobody is there. You leave the warehouse in the evening after the last parcel, and conversations on the site keep coming. The bot then has to tell the customer plainly that the team will reply later, and ask for contact details. The same goes for missing data. A sensible answer to a question about weight that is absent from the description sounds roughly like this: “The weight of this model is not in the description. I can pass the question to the team.” If the bot instead quotes the weight of a similar product, it sounds helpful and sells the customer a guess.
How to install a chatbot in your store
The chat window itself is switched on by a script pasted into the page, much like an analytics snippet, and that is the shortest part of the job. Connecting the catalogue and order tracking takes more work, because every platform has its own permissions and its own theme quirks. On Shopify, for example, measuring sales from the chat needs a separate pixel.
The bot takes the catalogue best straight from the store’s API. If that is impossible, an XML or CSV feed at a URL will do, and as a last resort a plain file. A file has one drawback. It does not refresh itself, so after a price change someone has to upload a new one, and API sync also runs on a schedule, so a new price reaches the chat with a delay.
Before you set a start date, ask the vendor for a list of the access they need. Also write down the questions customers ask you most often, together with your answers. They are the best material for acceptance testing, because you know them by heart and will spot at once where the bot strays from the truth. We do the CartBot setup ourselves, as part of the subscription.
How to choose a chatbot for your store
In a demo the vendor chooses the questions. Come with your own and ask for a live demonstration, ideally on your catalogue, because everything looks good on someone else’s. Below are five questions to start with.
On the cart question, do not take anyone’s word for it. Ask for a red jacket in size M, add it through the chat and look inside the cart, because the card in the conversation looks the same whether the red or the black one ended up there.
- What will the bot answer when a customer asks about a feature missing from the description, say the weight?
- Will it add a specific size and colour to the cart?
- How does a customer ask for a person, and what will they see on a Sunday evening when nobody is at the computer?
- How long until the bot sees a new price?
- Will you show me a test order in the report, so I can compare it with the store panel?
How to tell the bot pays off
For the first weeks, read the conversations to the end. Check whether the customer got the answer they came for and whether cases passed to a person got a reaction. While you are at it, note how much time you spend replying now, so you can compare it with the month before launch.
A report of orders linked to conversations shows which purchases went through the chat and for how much. It cannot tell you how many of them would have happened without the bot. For the payback calculation, set the cost of the service against what is left of those orders after costs, and add the hours you no longer spend replying. If you want to see this on your own products, write to us. We will show you the bot on your catalogue before you pay anything.