The word "bot" today covers three different products, and the gap between them is tenfold in price and a hundredfold in the kind of tasks they close.
Level one: the scripted bot
A tree of answers written in advance. "Press 1 to check your order status." There is no neural network inside.
It replies instantly, costs little, and behaves predictably down to the last character. It breaks on the first question that is not in the tree: the user writes in plain human words, and the bot replies "I didn't get that, please pick a menu item."
It works when you have fewer than twenty scenarios and they do not change.
Level two: the bot with a language model
Inside sits a neural network that understands a question phrased any way at all. Behind it is the company knowledge base: regulations, product descriptions, answers to common questions.
The user asks "and if the size doesn't fit, what do I do", and the bot finds the return rules in the base and replies in coherent prose.
What it can do: understand, search, explain. What it cannot do is act. It will walk you through the return procedure, but it will not create the request.
Level three: the AI agent
The same model plus tools. The agent has access to functions: look up an order in the database, create a request, send an email, assign a task, write off from stock.
The model itself decides which tool to call and in what order. The user writes "refund order 4412", and the agent finds the order, checks the return window, creates the request, assigns a task to accounting, and replies to the person.
This is where the word "agent" earns its place: the system acts, and its result changes the data in your systems.
Where each one breaks
Scripted. On an unfamiliar question. Cured by widening the tree until the tree becomes unmanageable.
With a model. On hallucinations: if the answer is not in the knowledge base, the model is prone to inventing a plausible one. Cured by a hard instruction to answer "I don't know" and by showing the source under every reply.
Agent. On consequences. A scripted bot's mistake is an inconvenience. An agent's mistake is an email that went out, stock that was written off, or a refund that was created. Cured by limiting permissions and requiring human confirmation on irreversible actions.
What it costs per month
Development is a separate conversation. Below is about running costs, which people ask about less often and where the surprises are bigger.
| Type | Cost per request | 10,000 requests per month |
|---|---|---|
| Scripted | 0 | Hosting only |
| With a model | ₽0.5–3 | ₽5,000–30,000 |
| Agent | ₽3–15 | ₽30,000–150,000 |
The spread is wide because the price depends on context length. An agent that gets the dialogue history, descriptions of ten tools, and chunks of the knowledge base slotted into every request spends many times more on a single answer than a bot handling a short question.
The practical takeaway: look at the cost of one request for your average dialogue. The model's price per thousand tokens tells you nothing on its own. The difference between a well-assembled and a badly-assembled context is a fivefold difference on the invoice at the same answer quality.
How to choose
Answer two questions.
First: how many different phrasings does one question have? If users write in free text and there are dozens of variants, a scripted bot will not do, however many branches you add.
Second: does the system need to change anything in your data? If the job is only to answer, stop at level two. An agent is more expensive to build, run, and maintain, and taking one just to answer questions makes no sense.
On moving between levels
The levels stack on top of each other. A bot with a knowledge base is built over the same infrastructure as a scripted one, and an agent adds a layer of tools onto the bot.
Starting at level two is sensible almost always: it clears out the bulk of repetitive requests and gives you the statistics that show which actions are worth handing to an agent. Starting straight at level three means designing access rights and error handling for scenarios whose frequency you know nothing about yet.