At 11:40 p.m. a visitor opens a product page and wants to know one thing: is it in stock, and when will it arrive. The manager went home hours ago. A “leave your details and we’ll call you back” form looks polite, but someone who was just about to buy usually closes the tab.
An AI assistant on the site exists for moments like this. There’s a lot of hype around it, though, so it’s worth being calm about what it does and what it doesn’t.
What it does well
The widget answers visitors’ questions, advises on products and services, helps find and choose a product, books clients for services and takes requests. Those aren’t five separate features but one conversational interface that replaces digging through menus when a person doesn’t know what the thing they need is called.
Say a shopper doesn’t type “impact drill 800 W” but “I need to drill into a concrete wall in my flat, what should I get”. A standard site search has nothing to say to that. The assistant asks what exactly will be drilled and how deep, then suggests a few options from the catalog.
Answers must come from your materials
This is the line between a useful assistant and a harmful one. If it answers from a language model’s “general knowledge”, sooner or later it will invent delivery terms you never offered. So a sound setup relies on your catalog, your delivery and payment terms and your answers to common questions. What isn’t there, it doesn’t claim.
The practical consequence: at launch the assistant is about as good as the material you give it. If your product descriptions are three words long, don’t expect good advice.

Bookings and requests
For services the scenario is different. A salon, a clinic, a repair shop: the client writes in the chat, the assistant clarifies what’s needed and books a free slot or collects contact details for a manager. The request arrives where people work on it, with the context of the conversation attached. The manager doesn’t have to ask again what the client already said.
Handing over to a manager without losing context
The worst thing you can do is make a person repeat their question. When a conversation moves from the assistant to a manager, the whole dialogue should go with it: what was asked, what was answered, what the client has already decided. Then the manager steps into the chat with “I see you were choosing between two models, let’s narrow it down” rather than “how can I help?”.
It also matters that the handover works outside business hours. In the evening the assistant collects a contact and a convenient time; in the morning the manager sees a ready request with the full history.
What an assistant doesn’t do
It doesn’t negotiate discounts unless you set that up. It doesn’t decide whether to refund. It doesn’t know about a promotion that started yesterday until someone adds it to its knowledge.
It can also be wrong. Language models sometimes sound confident even when they don’t know the answer. That’s why there should always be a clear way out to a person: “call a manager” available at any moment, not after five rounds of clarifying questions. That’s the difference between a service customers don’t mind and one that annoys them.
How to test before launch
Take real questions from your email, messengers and calls. Twenty or thirty of them, including the clumsy ones with typos. See what the assistant says. Check the touchy topics separately: warranty, returns, delivery times. If it starts making things up there, either the knowledge base is missing material or the instructions are too loose.

What to watch after launch
In the first weeks, simply reading the conversations is more useful than any analytics. A few dozen chats already show where the assistant gets confused, what material it lacks and what people ask that you never thought about.
Mark the unanswered questions separately. That’s a ready list of what to add to the knowledge base: a new delivery condition, a missing description, an answer to something asked all the time. That’s roughly how an assistant grows: not through a “smarter model” but by you closing real gaps.
And look at how conversations end: with a request, a booking, someone leaving without an answer, a manager call. A lot of exits without an answer means either the knowledge base has holes or the assistant spends too long asking clarifying questions.
What about the cost
It depends on where the widget has to work (website, messengers), what it integrates with (catalog, CRM, booking calendar) and how much material needs preparing. The total is built from those three parts, so for one store it’s a small task and for another a full project. We name a ballpark after a short look at your site.
It’s best to begin with one scenario: either catalog consultation or requests. Once it’s running and you can see real conversations, expanding is easier than designing everything up front.
Read also: What a Small Business Can Automate with AI; AI Employee vs. Chatbot vs. ChatGPT: What’s the Difference.
Related services: Website and CRM integration, 1C integration, Custom CRM.