AI agents for business: what they are and what types exist

An AI agent is a program you give a goal to, and it picks the steps toward that goal itself: it works inside your systems, calls the tools it needs, checks the result and repeats the cycle until the task is closed. A chatbot, in the same logic, only answers a question and stops. The difference is not in how politely it replies, but in who decides what happens next.

Below: what vendors call an agent, what types of agents exist, where they already work in Russian companies, why Gartner predicts that more than 40% of agentic projects will be cancelled, how to tell a real agent from a repainted bot, and who is legally liable for its mistakes.

What is an AI agent in plain terms?

An agent is the combination of a model, tools and a decision loop. A human sets the goal, the model breaks it into steps, calls tools (CRM, database, email, an external API), looks at the result of each step and adjusts the plan. In its technical write-up, Google splits an agent into exactly three parts: the model as the brain, the tools as the hands, and the orchestration layer as the nervous system that runs the observe-reason-act loop.

The most practical dividing line was drawn by Anthropic. A workflow is a system where “LLMs and tools are orchestrated through predefined code paths”. An agent is a system where the LLM “dynamically directs its own processes and tool usage”. In other words, in a workflow the route was set by a programmer; in an agent the model chooses the route as it goes.

One more important caveat from IBM, the kind that saves money in negotiations with a contractor: “tool calling alone does not make an LLM an agent”. If the model can call an API but the order of those calls is hard-coded, what you have is a workflow — and that is perfectly fine. A system becomes an agent at the point where the order of steps is not known in advance.

How is an AI agent different from a chatbot and from RPA?

In three ways: who chooses the sequence of steps, what happens when things fall outside the script, and what kind of data the system can handle. A chatbot follows a dialogue tree, an RPA robot executes an if-then script, an agent builds a plan for a goal and rebuilds it on the fly.

From this comes a working rule of thumb for an owner. If the task is phrased with the verb “answer”, it is a bot. If it is phrased with the verb “do”, it is an agent. Parse an incoming email, pull the order lines out of it, check them against stock, create a deal and write back to the client with a clarifying question — that is “do”.

RPA is not going anywhere, though. It is good where operations are high-volume, the rules are rigid, and the systems do not change for years. A common working pattern is a hybrid: the agent makes the judgment (what this document is, where it belongs, whether a human is needed), and RPA performs the mechanical data transfer.

What types of AI agents are there by design?

There is only one classification with a genuine primary source — from Stuart Russell and Peter Norvig's textbook “Artificial Intelligence: A Modern Approach”, the chapter on intelligent agents. It describes five types in increasing order of behavioral complexity.

1. Simple reflex. A condition-action rule, no memory. An email arrives containing the word “invoice” — file it under “Accounting”.
2. Model-based reflex. Keeps internal state and takes history into account: it remembers that this client already wrote yesterday, and what about.
3. Goal-based. Reasons about the future and chooses actions based on a goal: “assemble the full document set for this deal”, not “process one email”.
4. Utility-based. Judges outcomes not on a “worked or didn't” scale but against a utility function: which route is cheaper, which reply to the client raises the odds of closing.
5. Learning. Changes its behavior from feedback: what the operator corrected by hand, the agent does itself next time.

Be careful with “levels of autonomy”. There is no single standard for AI agents, nothing like the SAE levels for self-driving cars. The elegant ladders from bot to autonomous system that circulate in blogs are an editorial metaphor — useful for explaining the spectrum, but not something you can cite as a standard.

Where are AI agents used in Russian companies?

Above all in document handling and request processing. According to a study by SberAnalytics and Sber Business Soft — the analytics and software arms of Russia's largest bank — (November 2025, 559 respondents), roughly 39% of the companies surveyed use AI agents and assistants. The processes most often automated are document handling and request processing (70%), accounting and financial records (55%), HR (34%) and customer support (30%).

The same study also puts numbers on the effect: AI assistants on the first line of support and in call centers cut labor costs by up to 30–40%, 45% of respondents report routine tasks being done faster, and 36% report fewer errors in documents.

Broken down by function, the picture looks like this:

• First-line support and sales — answers standard questions, qualifies the lead, puts the deal into the CRM. We have a separate breakdown of the economics of this scenario with a payback calculation: when a first-line AI agent pays for itself.
• Document handling — reads incoming primary accounting documents, checks the counterparty details, enters the document into the accounting system. What this looks like on 1C, the accounting and ERP platform most Russian companies run on, is covered in our piece on recognizing primary accounting documents.
• Finance and accounting — reconciling payments against delivery notes, allocating them to the right accounts, preparing reports.
• Analytics — pulls data from different systems and answers the owner's question in words, rather than as an Excel export.
• Procurement and supply — assesses shipments, finds discrepancies in deliveries.

All of these scenarios have one thing in common: an agent pays for itself where there is a stream of repetitive operations on unstructured data and a clear price tag on the manual labor.

Why does Gartner predict that more than 40% of agentic projects will be cancelled?

Because of rising costs, unclear business value and weak risk controls. In a press release dated June 25, 2025, Gartner predicts that over 40% of agentic AI projects will be scrapped by the end of 2027. The analyst's wording is blunt.

“Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied” — Anushree Verma, Senior Director Analyst at Gartner, from the press release of June 25, 2025.

The MIT NANDA report “The GenAI Divide” (August 2025) painted a similar picture: around 95% of corporate generative AI pilots produce no measurable effect on the bottom line. The report does not blame the technology: general-purpose tools are good for individuals but do not fit into a company's workflows. The second finding of the same study is telling — projects where the solution was bought from a specialist vendor reached a result in roughly 67% of cases, while in-house builds succeeded about a third as often.

The way to read this is not “AI doesn't work” but “an agent without a measured process and an accountable owner doesn't work”. Failure is built in before the start: no volume of work, no metric in money, no knowledge base.

How do you tell a real agent from a repainted bot?

By who chooses the order of actions, and by what happens at the edge of the script. Gartner coined the term “agent washing” for the practice of relabeling an old product as “agentic”, and estimates that out of the thousands of vendors calling themselves agentic, only about 130 have genuine agentic capabilities.

Five questions for a contractor that clear away most of the fog:

1. What happens if the client asks a question that isn't in the script? An answer of “we'll add it to the script” means bot.
2. What tools does the agent use, and does it decide for itself when to call them? A list with a single call is an integration, not agency.
3. How does the agent know the task is done? If there is no success criterion, there is no decision loop either.
4. What happens when the model's confidence is low? There must be an explicit handover to a human with the conversation context, not an invented answer.
5. Where does the knowledge base live and who updates it? Without an answer to that, in three months the agent will be quoting an outdated price list.

Answers to these five questions, written into the contract, are what separates a system from a demo.

When do you not need an AI agent?

When a workflow does the job. That comes from the developer of one of the leading models itself.

“Find the simplest solution possible, and only increase complexity when needed. This might mean not building agentic systems at all. Agentic systems often trade latency and cost for better task performance” — Anthropic, Building Effective Agents.

The logic is simple: an agent is more expensive and slower than a workflow, because the model makes several passes instead of one. Paying for that makes sense where rigid rules break down — on unstructured data and in tasks with branching you cannot enumerate in advance.

The second limiting factor is the cost of an error with no human in the loop. Taco Bell rolled out voice AI at more than 500 drive-thrus, then began rethinking that bet after a series of failures: one customer ordered 18,000 cups of water to make the system give up and call a live person. The chain's chief digital officer, Dane Mathews, summed up the experience briefly: “Sometimes it lets me down, but sometimes it really surprises me.” The company's conclusion: at peak hours in busy locations, a human does the job better.

Who is liable if an AI agent makes a mistake?

The business that deployed the agent — even if a contractor built the system. Lawyers at Clifford Chance describe this as the “liability gap”: a vendor's disclaimers shift liability away from the vendor, but they do not shift it away from the company in the eyes of its own client and its regulator.

“Under the terms of many agentic AI technology agreements, the business acquiring the technology (i.e. the customer) ultimately bears the risk of actions taken by AI agents” — Clifford Chance, Agentic AI: the liability gap your contracts may not cover, February 2026.

The practical conclusion for both the contract and the architecture: the limits of the agent's autonomy must be spelled out. What it does on its own, what it does only with an operator's confirmation, where a handover to a human is mandatory. An agent allowed to issue invoices and sign documents without confirmation is not a saving — it is a transfer of risk onto the business owner.

What about personal data and Federal Law 152-FZ?

Sending text containing personal data to an external language model already counts as processing personal data. If the model is foreign, a cross-border transfer is added on top, with its own requirements for notification and for the data subject's consent. Federal Law 152-FZ is Russia's personal data protection law.

“The moment you send text like that to a model, you have started processing personal data and you fall under 152-FZ” — Alexander Eroshkin, article on vc.ru, June 17, 2025 (translated from Russian).

There are three workable options, and they differ in price and in strictness:

1. De-identification before sending. If not a single piece of personal data goes into the model, the cross-border transfer question disappears entirely. The cheapest route, but it takes careful data preparation.
2. A Russian provider. Yandex Cloud signs a data processing agreement, which provides a legal framework. GigaChat, Sber's Russian language model, explicitly prohibits sending personal data in requests — de-identified data only.
3. An open-weights model inside your own perimeter. The data never leaves the company's perimeter, and localization is satisfied by design. More expensive in infrastructure, but it removes almost every question.

A detailed breakdown of how to build a perimeter that data cannot leak out of is in our article “Neural networks without data leaks”.

Where do you start if you do need an agent?

With the process and the numbers, not with the choice of model. The order that saves budget:

1. Find the volume. How many repetitive operations per month, how many minutes each one takes, what an hour of an employee's time costs. If the volume is small, the savings will not cover the cost of maintenance.
2. State the goal in money or hours. “Adopt AI” is not a goal. “Take 70% of standard inquiries off the first line” is a goal with a metric attached.
3. Assemble the knowledge base. Procedures, price lists, answers to frequent questions. Without it, the agent will invent things confidently, and clients will notice before you do.
4. Start with a workflow. If rigid logic covers 80% of cases, you only need an agent for the remaining 20% — and it will cost noticeably less.
5. Define the limits of autonomy and the handover to a human. Before launch, not after the first incident.
6. Appoint an owner for the system. Someone who updates the knowledge base and watches the metric. Without one, the project joins that 40% of cancelled ones.

If you are still deciding what to automate first, start with the overall map — we put one together in “Business automation: where to start”. And if the process is already clear and you need an agent built for it, that is custom development: the system is built around your own procedures, under contract, with data kept in Russia and with support after handover.

Sources

• Anthropic, Building Effective Agents — the line between a workflow and an agent.
• OpenAI, A Practical Guide to Building Agents — the definition of an agent and where it applies.
• IBM, AI agents vs. AI assistants — why tool calling on its own does not make a model an agent.
• Gartner, press release of June 25, 2025 — the forecast that over 40% of projects will be cancelled, and the estimate on “agent washing”.
• SberAnalytics and Sber Business Soft (November 2025, 559 respondents), summary on CNews — the share of Russian companies and the processes being automated.
• MIT NANDA, “The GenAI Divide: State of AI in Business 2025”, summary in Fortune — 95% of pilots with no effect on the bottom line.
• TechCrunch, Taco Bell is having second thoughts about AI at the drive-through, August 30, 2025.
• Clifford Chance, Agentic AI and the liability gap, February 2026.
• Alexander Eroshkin, LLMs and personal data: how to avoid breaching 152-FZ, vc.ru, June 17, 2025.
• Stuart Russell, Peter Norvig, “Artificial Intelligence: A Modern Approach”, chapter 2 — the five types of agents.