An AI chatbot for an online store: what it can do, where it goes wrong and how to judge its value

October 13, 2026

A shopper opens a product page, doesn’t find the delivery time to their city and closes the tab. A manager would have answered in a minute, but it is eleven at night. An AI assistant exists for such cases. Below: what it takes on, where it makes mistakes and how to check it is worth having.

How it differs from a chatbot with buttons

A button bot follows a script: “Choose a section”, “Choose a question”. Someone has to think through and draw every path. If the shopper writes “do you have the same one, but darker and under fifty dollars”, the script breaks.

An AI assistant receives the question in free form, finds the relevant fragments in the store’s data and composes an answer. You don’t write scripts. In return you are responsible for the data: if your site has no returns policy, the bot won’t invent one. If a price in the catalog is out of date, it will state the old one.

Which questions it closes

In most stores the questions repeat. A typical set:

Terms: shipping within the city and to other regions, timing, payment, returns, warranty, opening hours.
Availability and price: “how much is this set”, “is it in stock”.
Recommendations: “suggest an inexpensive pillow”, “what would suit a gift under thirty dollars”.
Navigation: “where are your contacts”, “how do I place an order”.

Budget recommendations work well if the bot searches the catalog like a database: filters by price, sorts, takes a category. If it merely “reads” pages, a request like “the most expensive” turns into a guess.

Where it goes wrong

This is the most useful thing to know before buying. The main causes of mistakes:

The catalog was read incompletely. The store has tens of thousands of items and the bot sees part of them. It won’t find a product it doesn’t have and will say it isn’t for sale.
Outdated data. Prices and availability change faster than the bot’s knowledge base is refreshed.
Variable products. The price is set on a size or color, not on the main product. The bot may not see the price.
Questions about a specific order. Where is my order, why wasn’t I refunded. The bot doesn’t see your CRM and shouldn’t guess.
An ambiguous request. “Something pretty”. A good bot asks a follow-up, a bad one offers something random.

Good behavior when it doesn’t know looks like this: the bot says “I don’t know” and offers to leave contacts. Bad behavior is a confident answer that isn’t in the data. How we tested this on real stores is described in a separate article.

What such a bot consists of

Four layers, and a mistake can hide in any of them:

Data. The site’s pages and the catalog. The best catalog source is an XML (YML) feed; page markup is worse.
Search. Finds the right fragments and products among thousands.
The language model. Words the answer from what was found.
Rules. Don’t make things up, don’t reveal instructions, hand hard cases to a human.

How to judge its value in two weeks

Don’t wait for a vendor’s pretty reports. Use simple numbers.

Before launch. Write down 20-30 questions your managers receive most often. That is your test set for the bot. Ask them to the bot and mark where the answer is right.

After launch, the first two weeks. Look at:

the share of conversations in which the shopper got an answer without a human;
the number of leads that arrived outside working hours;
the questions the bot didn’t answer. This is a list of what your site is missing;
the number of mistakes that you or customers noticed.

A simple estimate: routine questions per day × manager time per answer. If that is less than an hour a day, payback will be modest. If it is several hours, and some questions arrive at night, there is something to talk about.

When you don’t need a bot

The product is complex and every shopper needs expert advice: industrial equipment, custom orders.
The store gets a few inquiries a week.
Your site lacks the basics: shipping, payment, returns. Write them first; people need them too.

What next

If you are going to try one, start by preparing the data: check that the site has shipping and payment terms, that products have prices and that the catalog can be exported. See how BotNow works and which platforms we support on their own pages.

A comparison with other ways of answering shoppers: AI assistant, live chat or a form. What the bot does when it doesn’t know the answer: read on.

Frequently asked questions

How is an AI assistant different from an ordinary chatbot with buttons?

A button bot leads the shopper along a pre-drawn script. An AI assistant understands a free-form question and looks for the answer in the store's data. You don't write scripts, but the quality of the answer depends on how complete that data is.

Does a store with little traffic need a chatbot?

If a couple of dozen people visit a day and you manage to answer them yourself, a bot adds little. The value shows when questions arrive in the evening and at night, when the same ones repeat, or when the catalog is so large that shoppers get lost in it.