A frequent mistake with an AI bot on a site looks like this: the bot was connected, given a couple of pages of text, and it briskly answers any questions. Including those for which there is no answer in those pages. A buyer asks about delivery to another city, and the bot confidently names a deadline that nobody promised. They ask about a discount — the bot invents one. A model is built that way: it picks a plausible answer, and if it is not limited, the plausible easily becomes invented.
So a conversation about an AI bot for a site is worth starting not with the choice of a service, but with the question: what the bot should answer, where it takes the answers from, and what it does when there is no answer. Below is what such a bot can do, how it differs from a familiar bot with a script, how the knowledge base is built, which integrations there are, and what the cost of implementation is made of.
What an AI chatbot on a site can do
Answer questions from the knowledge base
The main work of an AI consultant is to answer visitors’ questions from the business’s materials: delivery and return conditions, opening hours, what the services consist of, how to book. Unlike a “Questions and answers” section, the bot understands a question asked in the visitor’s own words and finds the answer even if it is worded differently in the base.
A well set up bot answers only from the knowledge base, and when there is no answer it honestly says so and offers to contact a person. That is the main thing that distinguishes a useful bot from a risky one.
Pick products and services
If the bot is connected to the catalog, it can help with the choice: clarify what the product is needed for, what the budget is, what the limits are, and suggest suitable items. For services it is the same: from a description of the task the bot will suggest which service fits and explain the difference between the options.
Here it is important that the bot works with current data. If the catalog was updated and the bot keeps offering what is not in stock, there is little use in the picking.
Pass a request to a manager
The bot can collect contacts and the gist of the question and pass them to a person: into a CRM, to email, into Telegram. This is convenient when a visitor writes in the evening or on a weekend: the request is not lost, and the manager sees the whole correspondence and does not ask the same questions again.
The decision of when to pass the conversation to a person also has to be made in advance. For example, always for questions about a refund, for complaints, for a request for a non-standard order.
It is useful to decide in advance which data the bot collects before passing the conversation on: the name, a convenient way to get in touch, the gist of the question, and for a store — the products of interest. The better the request is filled in, the shorter the manager’s first conversation with the client. At the same time the conversation should not be turned into a questionnaire: if the bot asks ten questions in a row, the person will close the window before they get to the contacts.
How an AI bot differs from a bot with a script
A bot with a script works by a tree set in advance: buttons, answer options, transitions. It will not make a mistake within the script, but it will not understand a question that did not get into the script. An AI bot understands free text and formulates the answer itself, and that is why it needs limits.
| Feature | A bot with a script | An AI bot |
|---|---|---|
| How it understands the question | By buttons and keywords | By the meaning of free text |
| Where the answer comes from | Written in advance | Formulated from the knowledge base |
| Risk of inventing | None | There is, if it is not limited to the base |
| New questions | Does not handle them | Answers if the answer is in the base |
| Setup | A scheme of the script | Knowledge base, rules, a check |
| Suits | Bookings, simple surveys, a menu | Consultations, picking, answers from documents |
In practice these approaches are often combined: a script for typical actions such as a booking or submitting a request, and AI answers for free questions.
If visitors’ questions come down to a few typical ones, you can start with a scripted bot or an ordinary form. An AI bot is justified when there are many questions, they are varied, and the answers to them are in the business’s materials, but visitors do not find them.
How the bot learns about your business: the knowledge base
A language model by itself knows nothing about a particular company. For the bot to answer to the point, it is given a knowledge base: page texts, conditions, a price list, instructions, answers to frequent questions.
Usually this works by a scheme called RAG — retrieval followed by generation of the answer. If you explain it simply: when a visitor asks a question, the system first looks in the knowledge base for suitable fragments, and then the model formulates an answer based on them. The model does not hold the whole base in its head: it receives only what was found for the question.
Three practical conclusions follow from this.
The first: the quality of the answers depends on the quality of the base. If the base has outdated delivery conditions, the bot will confidently name them. If the same information is written in two places differently, the bot may take either variant.
The second: search through the base can miss. The question is worded differently from the text in the base, and the needed fragment is not found. Then the model will either say it does not know, or — if it is not limited — invent. So the rules for the bot state directly: answer only from the materials that were found, and if there are none, offer to contact a person.
The third: the base has to be maintained. Prices, conditions, hours changed — they have to be updated in the knowledge base, otherwise the site and the bot will say different things.
A good knowledge base for a bot consists of short thematic fragments: one question — one answer, or one topic — one section. Long documents, where what matters is hidden in the middle, the bot uses worse.
It is convenient to check the base on real questions. I take questions that clients asked by email, in messengers and by phone, and run them through the bot. In each answer two things matter: whether such a fact is in the base, and whether the bot added something of its own. Questions the bot could not answer show gaps in the base. Answers with invented details show that the rules need to be tightened.
Integrations: CRM, Telegram, catalog
By itself the bot answers questions. It starts bringing benefit to the business when it is connected to other systems.
CRM. The bot creates a deal or a lead with contacts and the history of the correspondence. The manager sees what the person asked about and continues the conversation from the place where the bot stopped.
Telegram. The simplest option for a small business: a notification of a new request comes into a work chat or personal messages. It suits if there is no CRM or it is not needed yet.
Catalog. The bot gets access to products, prices and availability and can pick items. Here it matters how often the data is updated: if the bot works with an export from a week ago, it may offer what is no longer there.
Tables. Requests, or questions the bot could not answer, can be collected in Google Sheets. Such a list is useful on its own: it shows what is missing in the knowledge base and on the site.
Each integration is a separate setup and a separate point that has to be checked: whether the request arrived, whether the fields were filled in correctly, what happens if the CRM is temporarily unavailable.
How to add a bot to a site and to Tilda
Almost all AI bot services connect the same way: the service gives a widget code, it is inserted on the site’s pages, and a chat window appears in the corner of the screen. Setting up the bot itself — the knowledge base, the rules, the integrations — happens in the service’s account.
On a site with its own code the widget is added to the shared page template, so the chat is available everywhere, or only on the pages that need it.
On Tilda there are ready blocks for popular chats. For example, Tilda’s help describes the T133 block for JivoChat: the widget code is inserted into it. For the widget to work on all pages of the site, Tilda recommends adding it to the page that is used as the shared header. If there is no ready block for the service you need, the widget code is inserted into the T123 “HTML code” block. This is the only case when I use an HTML block on Tilda: to output third-party widgets.
When choosing a service for the bot it is useful to find out several things in advance:
- whether answers can be limited to the knowledge base only, and whether you can set what the bot says when there is no answer;
- in which formats the base is uploaded: texts, files, site pages, tables;
- which integrations are ready, and which will have to be set up separately;
- whether the history of dialogues is visible, so the bot’s mistakes can be taken apart;
- where the correspondence and the users’ contacts are stored;
- how the widget looks on a phone and whether it can be adjusted to the site’s design.
The answers to these questions are usually in the service’s documentation. If they are not there, that is an answer too.
After connecting it is worth checking three things: the chat does not cover important elements of the page on a phone, does not get in the way of buttons and forms, and does not slow down the loading of the page.
Personal data and consent
A bot that collects a name, a phone or an email processes personal data. In Russia this is regulated by Federal Law 152-FZ “On personal data”.
Under article 9 of the law, consent to the processing of personal data must be specific, subject-related, informed, conscious and unambiguous. From September 1, 2025, after the changes introduced by law No. 156-FZ, consent must be drawn up separately from other documents and information that a person confirms or signs.
For a bot in practice this means: before contacts are asked for, the person must see why they are needed and who will receive them, and give consent separately. Besides that, it is worth understanding where the bot service stores the correspondence and the data. The concrete wording of the consent and of the data-processing policy is better agreed with a lawyer: the requirements depend on which data is collected and where it is passed.
What the cost of implementation is made of
The cost of an AI bot for a site is made of several parts, and not all of them are obvious at the start.
A subscription to the service. Most platforms for AI bots work by subscription. Plans often depend on the number of dialogues or messages, the number of bots, the integrations available.
Use of the model. In some services payment for the work of the language model is included in the subscription, in others it is counted separately by the volume of requests. This is worth finding out before choosing a service: as the number of dialogues grows, the expenses can grow.
Preparing the knowledge base. Collect the materials, split them into topics, remove contradictions, write what is missing. Often this is the most labor-intensive part: it turns out that half the answers to frequent questions are not written down anywhere.
Setting up rules and scripts. What the bot does when it does not know the answer, when it passes the conversation to a person, which questions it does not discuss.
Integrations. Connecting a CRM, Telegram, a catalog, tables. The more systems, the more setup and checking.
The check. Run the bot through real questions from clients, find where it is wrong or invents, fix the base and the rules. Without this stage it is risky to launch the bot.
Support. Update the base when conditions change, look through the questions the bot did not answer, add materials. A bot without support starts answering with outdated information over time.
The most frequent underestimate is preparing the knowledge base and the check. Connecting a widget can be done quickly. Getting the bot to answer correctly and not invent — that is the main work. A good sign of readiness to launch: on test questions the bot either answers exactly from the base, or honestly says it does not know and offers to contact a person. If it invents even sometimes, the knowledge base or the rules are not ready yet.