On the Kazakhstani market, three completely different tools are sold under one banner: "artificial intelligence." RPA, chatbots and AI agents are three different classes of technology, with different prices, different capabilities and different limitations. Confusing them means either overpaying several times over, or buying a tool that cannot handle the job. In this article we break down how they differ and how to choose the right one for your specific task.
Why the confusion costs real money
The cost of choosing wrong is not abstract. Here are two scenarios that repeat on the market every month.
A large retailer orders an "AI chatbot" for its website for 12 million tenge. In reality, 92% of customer queries are "where is my order," "opening hours," "how do I make a return." A static bot for 400 thousand would have handled that. The difference in cost of ownership over three years is around 40 million tenge, for the same result.
The opposite case: the CFO of a holding company orders an AI agent to move payments from the bank-client system into 1C. The AI "reads the screen" like a human and delivers 87% accuracy, meaning an error in every eighth transaction. The accounting team is horrified. An RPA robot for one and a half million would have delivered 99.9% here. They picked the wrong class of tool and lost six months.
Both failures happened because no one explained the difference. Let's explain it.
RPA: a robot that repeats actions
RPA (Robotic Process Automation) is a software robot that repeats a person's routine actions inside existing systems: it clicks, copies, moves data between programs. It does not understand meaning; it simply reproduces a memorized sequence.
RPA is strong wherever the scenario is the same every time: migrating data between systems, filling out standard forms, exporting reports, working with legacy programs that have no API. Its weakness is that it breaks with any change to the interface and cannot handle unstructured data such as emails or contracts.
A good example from Kazakhstan: every morning an accountant downloads a statement from the bank-client system, opens 1C, imports it and reconciles it against the payment register. 45 minutes of identical actions every day. An RPA robot does this in 3 minutes, saving around 150 accountant hours per year.
The classic chatbot: scripted answers
A chatbot answers questions along a pre-written tree of scenarios. It recognizes keywords or offers a button to press from a menu, but it does not understand the meaning of the query.
A bot is strong at narrow, standard tasks: opening hours, order status, how to reset a password. It is predictable, cheap and quick to launch. Its ceiling: any question outside the script leads to a dead end, and the customer moves on to a live operator or closes the chat.
Example: a courier service receives 2000 requests a day, 90% of which are "where is my order" and "when will it arrive." Six buttons in a menu cover 85% of requests. AI is not needed here; it is more expensive and would add no value.
The AI agent: it understands meaning and chooses its own actions
An AI agent is built on a large language model (GPT or Claude). Unlike the first two, it does not follow a rigid script: it understands a task in natural language, breaks it into steps and reaches out to the right sources on its own – searching an archive, reading a database, generating a document.
An agent is strong wherever the data is unstructured and the wording is different every time: parsing documents and contracts, free-form customer requests, call analysis, searching archives. It even understands mixed Kazakh and Russian. Its limits: it is not suitable for tasks that demand 100% accuracy (accounting, payroll) and it can "hallucinate" if the architecture is built incorrectly.
Example: the legal department of a gas operator receives tender documents running 500 to 2000 pages. A lawyer spends 8 to 12 hours parsing a single tender. An AI agent with a RAG architecture does the initial parse in 15 minutes with 92 to 95% accuracy, and the lawyer only checks the key clauses.
Comparison table
| Parameter | RPA | Chatbot | AI agent |
|---|---|---|---|
| What it understands | Screen structure | Keywords | Meaning in natural language |
| Data type | Structured | Short queries | Any: text, documents, speech |
| Flexibility | Low | Low | High |
| Startup cost (KZ) | 1.5–8 million tenge | 0.3–1.5 million tenge | 3–15 million tenge |
| Accuracy | 98–100% | 70–85% | 90–95% (with RAG) |
| Best scenario | Data transfer | FAQ and navigation | Documents, calls, search |
Decision tree: 4 questions for your task
To choose a tool without vendors and presentations, answer four questions about your specific task.
What type of data comes in? If it's structured (forms, cells), that's RPA. Short queries from a limited set of topics: a chatbot. Free-form text, documents, calls: an AI agent.
Does the wording of the task change from case to case? Always the same: RPA. Within 5 to 20 scenarios: a chatbot. Every case is unique: an AI agent.
Do you need understanding of meaning, or is repetition enough? Repetition is enough: RPA. Recognize the topic: a chatbot. Understand the context and give a recommendation: an AI agent.
What is your tolerance for error? You need 0% errors (accounting): RPA. 5 to 15% is acceptable: a chatbot. 5 to 10% is acceptable with a human check: an AI agent.
A hybrid works most often
In real corporate deployments, a single tool is rarely enough. The most common and effective option is a combination of two or three technologies, where each does what it is strong at.
Take tender processing, for example: an AI agent parses the documentation and extracts requirements, RPA automatically monitors the platform and pulls new lots, and a chatbot in Telegram notifies the people responsible. Or a contact center: an AI agent analyzes calls for quality, RPA fills in the CRM from the transcript, and a classic bot handles standard customer questions. Not "one tool for everything," but the right combination for the process. That is exactly how an AI system is built: it orchestrates agents and tools into a single process that launches itself on an event.