In plain terms
Simply put, this program chats with a person on a website or in a messaging app, or talks to them on the phone. It deals with typical requests using ready-made rules or AI. AI-powered ones use natural language processing (NLP)natural language processing. Natural language processing (NLP) is the field of AI that teaches computers to understand, analyze and create human text., which lets them make sense of loose wording and typos. For a one-line version, see the short chatbot definition in our glossary.
Its main job is to spare people from answering the same thing over and over. It responds in seconds, at any hour, and does simple actions itself: it sends reminders, records details, reserves slots and accepts orders. For the customer, it feels like an ordinary chat with a staff member.
How a chatbot works
It works in five steps, from the customer’s message to a response or a handover to a human agent:
- It receives the message: text in a chat, a button press or a voice on a phone call.
- It works out the intent, meaning what the person wants: a delivery date, an appointment, a refund.
- It looks up what to say in a pre-written script or in its knowledge base.
- It answers or does the action: schedules a visit, places an order, sets a reminder.
- It saves the result in your business system, such as a CRM. If it didn’t understand, it hands the chat to a human agent.
What matters inside is how it understands people. A rule-based botchatbot. A chatbot is a program that exchanges messages with a person in a chat window: on a website, in a messenger or in an app. follows branches that people wrote for it in advance, and it understands only button presses and known keywords. The AI-powered type understands free-form text. Older AI bots use machine learningmachine learning. Machine learning (ML) is a field of AI in which a program finds rules in data by itself instead of getting them ready-made from a developer. to recognize the intent, then pick a prepared answer. Newer ones use generative AI: they write new wording each time from the material they were given.
Behind this newer kind sits a large language modellarge language model. A large language model (LLM) is a neural network trained on a huge amount of text. It writes, summarizes and analyzes text on request. (LLM), a program trained on huge amounts of text to write fitting words in response. It shapes two decisions: first, the model runs either on your servers or in the vendor’s cloud service, and that decides who sees customer data. Russian law has rules on this (see below). Second, a language model can give a wrong answer with full confidence, so someone has to verify what it writes.
Here is an example from a clinic. A customer writes that they want to see a doctor tomorrow morning. The program asks for their name and a suitable time, offers two free slots, reserves the chosen time and sends a reminder the day before. If the customer asks about test results, it doesn’t guess: it calls in a receptionist.
Types of chatbots
By the way they understand people, there are three types: rule-based, AI-powered and hybrid.
| Type | How it understands a message | What it can do | When it fits |
|---|---|---|---|
| Rule-based, with buttons | Only button presses and preset keywords | Lead people along fixed branches: bookings, delivery status, common questions | Requests are few and repeat often |
| AI-powered | Free-form messages, with typos and in the customer’s own words | Pick a prepared answer, or write one from a knowledge base | Requests are many and worded in different ways |
| Hybrid | Buttons for frequent topics, AI for everything else | Follow a script and handle anything typed freely | You want both predictability and flexibility |
A rule-based bot is predictable: it says exactly what was put into it. A generative assistant is more flexible but can get the wording wrong, so you have to review its answers more often.
What companies use them for
Businesses use them to serve customers around the clock and relieve human agents of routine work. Customers who find phone calls hard, such as people with hearing lossloss function. A loss function in machine learning is a formula that turns a model’s error into one number: the smaller it is, the closer the model’s answers are to the right ones., can type instead. These are typical jobs:
- Customer support: common questions handled with no queue, including nights and weekends.
- Orders and payments: delivery status, checkout, payment reminders.
- Bookings: a doctor’s appointment, a restaurant table, a meeting with a sales manager.
- An internal help desk for employees: one entry point to internal rules and documents. It covers HR matters such as leave, IT help with system access, and forms.
Here is how I’d work out whether the investment pays off before meeting any vendor. Count how many repeated requests come in a month and estimate, from your own figures, the share the program can close alone. Then price each conversation a human agent takes on: minutes spent on it divided by 60, times an agent’s hourly cost. Monthly saving = repeated requests a month × share closed without an agent × cost per request.
Payback in months = launch cost ÷ (monthly saving − monthly upkeep), with every amount in the same currency. If upkeep eats the whole saving, the project never pays off. If repeated requests are few, the sum won’t work, so read the section below on when to skip the idea.
For scale, Commerzbank says its Bene resolved 70% of requests, in a Google Cloud collection of examples updated in April 2026. It is a single large bank’s own figure in a vendor’s showcase, not a benchmarkAI benchmark. To benchmark AI models means to compare them on one standard set of tasks with known answers. That set is called an AI benchmark. for anyone else.
On a website, the bot talks to a visitor while they are still choosing. In Telegram, in Max (a Russian messaging app) and on social networks, it serves customers who already follow you, sending reminders and delivery updates. On Avito (a Russian classifieds site), it tells buyers about an item, and on the phone line it picks up calls. I’d start small: a single task in the channel where most repeated questions arrive.
I wrote about this in my book “Kak ne stat Bankrotom!?” (“How Not to Go Bankrupt!?”), published on LitRes. “Automating repetitive, low-value tasks frees up time for creative work and personal involvement” (translated). In practice, human agents then deal with complex problems, not the tenth question about opening hours.
Chatbot and AI agent: where the line is
A bot replies inside a conversation and waits to be asked. You hand an AI agentAI agent. By definition, an AI agent is a neural network that does not just answer but carries out a task step by step on its own: it looks for data, uses programs and checks the result. a whole task, as you would an employee. It works through the steps itself: it collects data, prepares an offer and moves a lead through to a deal. For a detailed breakdown with examples, read our AI agent vs chatbot comparison.
I hand the machine more than questions. “If writing a message to a business partner used to take 30 minutes, Artificial Intelligence does it in a matter of seconds” (from the same book, translated). Or, more directly: “I can delegate tasks not only to employees, but also to algorithms” (translated).
A bot serves your customers. An AI assistant works for you, as with that letter. An AI agent is given a whole task, like that lead, and carries it through to the end.
Three cases, three sets of numbers
Three examples show the scale: Moscow’s Alexandra handled more than 1 million passenger questions, Commerzbank’s Bene resolved 70% of requests, and IKEA’s Billie assisted 74% of its users. The three figures measure different things: volume, requests resolved and customers assisted, so they don’t compare directly. In all three, the program absorbs the routine flow, and people handle what calls for a human.
Alexandra, the program that answers passengers for the Moscow Department of Transport (the city’s transport authority), has used AI since 2025. By September 2026 it had handled more than 1 million passenger questions, taking under 2 seconds per reply, according to the Digital Platforms catalog. It runs on two language models and searches its own knowledge base. The details are in our case on the Moscow transport chatbot.
Commerzbank, a German bank, built an assistant called Bene for its customers. By the April 2026 update of that Google Cloud collection, it had handled more than 2 million chats and resolved 70% of all requests.
At IKEA, Billie launched in 2021. From 2021 to 2023 it resolved about 47% of the customer requests it received, Ingka Group, the largest IKEA retailer, said on 29 June 2023. By then, Ingka had retrained 8,500 call center staff, for example for remote interior design advice. By July 2026, Billie assisted 74% of its users, Fortune reports. The magazine counts users helped, while Ingka counted requests resolved, so the figures don’t show growth from 47% to 74%.
When you don’t need one, and why its answers must be verified
Skip the project if requests are few, every question is unusual, or the conversation calls for a decision only a person can make: a discount, a dispute, a complaint.
A bot has three weak spots. It fails on questions beyond its script and knows only what is in its knowledge baseAI knowledge base. An AI knowledge base is a set of selected and prepared company documents that an AI assistant searches for answers: rules, instructions, a product catalog and answers to common questions.. A wrong answer can also cost you a customer. That is why a button or a command that hands the chat to a human agent is a must.
Personal data is a separate question. If a customer gives a phone number, an address or an ID number in the chat, a company in Russia must process it under Russia’s personal data law, 152-FZ. To answer a customer about their own purchase or booking, it can rely on its contract with them (Article 6, part 1, item 5).
Other purposes, such as mailings, usually need the customer’s consent. Since 1 September 2025, that consent must be a separate document, not a line in the terms (Article 9, part 1). If the mailing is advertising, the recipient must also agree in advance to receive ads under Russia’s advertising law, 38-FZ (Article 18, part 1). Whatever the basis, the website must carry a data processing policy (Article 18.1, part 2).
Data on Russian citizens is first recorded and stored in databases inside Russia (Article 18, part 5). Sending the chat to a language model on a server abroad is a cross-border transfer. The company must notify Roskomnadzor, the Russian data protection regulator, beforehand (Article 12, part 3). So the first question for a vendor is where the model runs and where the chats go.
I recommend reading an AI assistantAI assistant. An AI assistant is a helper built on a neural network that, at a person’s request, answers questions, prepares texts and does simple tasks in work programs.’s answers as critically as any other source. In the book, I put it like this: “We need to learn to think critically, check facts, and pass information through the filter of our own mind and morals” (translated).
AI has two sides: it can create deepfakesdeepfake. A deepfake is a fake video, photo or voice recording made by a neural network so that a person seems to say or do something that never happened., but “set up correctly, it works as a filter” (translated). For a customer-facing program, that means three rules. Answer only from a verified knowledge base, show the source behind each response, and hand anything doubtful to a person.
How to build one: two ways
There are two ways: assemble it in a visual builder with no programming, or commission custom development. A builder suits a rule-based bot for a single task: bookings, common questions, collecting leads. Custom development makes sense when the program must connect to your CRM, accounting and other business applications, run on your own servers, or draw on your own documents with AI.
Either way, the work runs in the same sequence. Pick a single process with a flow of repeated questions, and collect real conversations from recent months with what your staff said in return. Write the script or the knowledge base, connect the channel, test it on your staff, and name a human agent who takes over the hard conversations. For a step-by-step example in one messaging app, see the guide on how to create a Max messenger bot.