AI agent in business — what it is and when it makes sense

An AI agent is a program in which a language model itself decides which step to take next and reaches for tools on its own: a calendar, a CRM, mail, a document base. A chatbot answers a question; an agent receives a goal and plans the way to it. That is how Anthropic, the maker of the Claude models, draws the line in its text "Building effective agents" of 19 December 2024. In a workflow, models and tools are orchestrated through predefined code paths. In an agent, the model dynamically directs its own process and tool usage.

OpenAI draws the line from the other side. In its guide "A practical guide to building agents" (2025) it writes that applications which integrate language models but do not use them to control workflow execution are not agents. Its examples are a simple chatbot, a single-turn model call and a sentiment classifier.

Cost, risk and legal duties depend on that distinction. Below: the three conditions under which an agent makes sense in a business, a calculation of token costs, the rules after the July 2026 change to the AI Act and data on how many European companies use AI at all. State of the law, prices and data on 8 October 2026.

What is an AI agent

An AI agent is a system that receives a goal, plans steps itself, calls tools and decides what to do next on the basis of their results, with varying involvement of a person. Four companies that build or sell such systems define it in a similar way:

• Anthropic: an agent is a system in which the model directs its own process and tool usage, maintaining control over how it accomplishes the task.
• OpenAI: agents are systems that independently accomplish tasks on your behalf.
• Google Cloud: software that uses AI to pursue goals and complete tasks on behalf of users, showing reasoning, planning and memory and having a level of autonomy to make decisions (“What is an AI agent?”, updated 2026).
• IBM: a system that autonomously performs tasks by designing workflows with available tools (IBM Think, “What are AI agents?”, 3 July 2024).

The common core of these definitions: a goal instead of a single question, a language model that plans, tools through which the agent acts in other systems, a decision loop (take a step, check the result, choose the next) and autonomy: the model chooses the order of steps, not the programmer. All four definitions come from companies that earn money from agents. The law does not define an agent: the AI Act has no such definition.

"Agentic AI" is the same family of terms in adjectival form: agentic systems, that is systems built from agents or acting like an agent. "AI agents" is the plural of the same concept.

The boundary of the concept is not sharp. In an expert report prepared under the European Data Protection Board's Support Pool of Experts programme (I. Barberá, "AI Privacy Risks & Mitigations — Large Language Models", 10 April 2025), the author writes that "the concept of agentic AI remains an evolving and not yet fully defined domain". The report expresses the views of its authors, not the official position of the EDPB. The practical conclusion for a buyer: the word "agent" on a product guarantees nothing. You check who chooses the next step in it.

Chatbot, assistant, automation, agent — four things under one name

The simplest test: who chooses the next step — a person, a workflow written in advance or the model? The answer tells you whether you are buying a chatbot, an assistant, automation or an agent.

• A chatbot answers questions. A person leads the conversation and the bot does nothing beyond answering. According to OpenAI it is not an agent. Chatbots in an online shop's support are covered in the article on ecommerce customer service.
• An AI assistant helps with work: it summarises, drafts, searches documents. In its press release of 26 August 2025 Gartner calls assistants the precursor to agentic AI: they simplify tasks and interactions, but depend on human input and do not operate independently. In this role they appear in office suites as Copilot in Microsoft 365 and Gemini in Google Workspace, which we compare in Google Workspace vs Microsoft 365.
• Automation (a workflow) runs a flow written in advance: a form came in, so save the contact in the CRM and send a confirmation. A language model can be one of the steps, for example it reads data from a document, but the order is decided by what somebody programmed.
• An agent receives a goal and decides itself which tool to call, in what order and when to stop.

In the same press release of 26 August 2025 Gartner writes that the most common misconception is referring to assistants as agents, and calls it "agentwashing". In its press release of 25 June 2025 it describes "agent washing" as the rebranding of existing products, such as AI assistants, robotic process automation (RPA) and chatbots, without substantial agentic capabilities. It also estimates that only about 130 of the thousands of agentic AI vendors are real. That is an analyst estimate, not a register.

Anushree Verma of Gartner gives a simple rule of division in the same release: use AI agents when decisions are needed, automation for routine workflows and assistants for simple retrieval.

How integration between systems differs from an RPA robot and from automation with AI is covered in the article on business process automation. Here one thing is enough: an agent is automation in which the model, not a rule, chooses the way.

How an AI agent works: model, tools, loop and limits

An agent works in a loop: the model plans a step, calls a tool, reads the result and decides on that basis what next, until the task is done or a stop condition is met. Anthropic describes agents as language models using tools in a loop, guided by feedback from the environment. At each step the agent should check the ground truth, that is the result of the tool call, to assess its progress. It may pause for human feedback at checkpoints or when it hits a blocker.

A tool is in practice the API of one of your systems. Anthropic gives a customer-support example: tools fetch customer data, order history and knowledge-base articles, and actions such as issuing a refund or updating a ticket are performed programmatically.

Tools are permissions. Every tool you give an agent is access: to customer data, to sending messages, to changing an order. OpenAI stresses that guardrails should be coupled with robust authentication and authorization protocols, strict access controls and standard software security measures. An agent connected to a mailbox and a drive sees everything in them, so before you connect it, check where your company's data really lives and who should have access to it. The same goes for tools attached through MCP (Model Context Protocol), a standard for connecting AI applications to company systems and data: through it the model can do as much as the account the server uses allows.

Without context an agent guesses. In its press release of 11 May 2026 Gartner quotes Rita Sallam: without a clear understanding of the specific relationships and rules within an organisation's data, AI agents cannot operate accurately and are far more likely to hallucinate, that is to state falsehood as fact. Companies approach this with caution. In a Gartner survey of 30 September 2025 only 19% of 360 IT leaders had high or complete trust in their vendor's ability to provide adequate hallucination protection, and 74% believed AI agents represent a new attack vector into their organisation. The survey covered organisations with at least 250 full-time employees in North America, Europe and Asia/Pacific, with no breakdown by country.

A common way to give an agent context is RAG (retrieval-augmented generation): before the model answers, the system searches matching fragments of company documents and gives them to the model along with the question. The quality of answers then depends on the quality of those documents.

A loop without a step limit is an open bill

The loop that gives an agent flexibility is also its biggest cost risk. When an instruction is unclear or a tool returns an error, the model can keep trying: call the tool again, get another error, try differently. Every turn of the loop is another model call, and you pay for each by the number of tokens. Anthropic says outright that the autonomous nature of agents means higher costs and the potential for compounding errors.

Both vendors give the same safeguard: a hard stop condition. Anthropic writes in the same text that the task usually terminates upon completion, but it is also common to include stopping conditions such as a maximum number of iterations to maintain control. OpenAI's guide lists reaching a maximum number of turns among typical exit conditions. Its toolkit for building agents (the Agents SDK) has a max_turns parameter for this: once it is exceeded, the SDK raises a MaxTurnsExceeded exception (OpenAI Agents SDK documentation, read 8 October 2026). The limit can be switched off, so whether it exists is decided by the implementer, not the tool.

The conclusion for a buyer: an agent's specification must contain a step limit per task and a monthly budget limit, and once either is exceeded the case goes to a person. What one step costs, we calculate below.

When an agent makes sense — three conditions

An agent makes sense where ordinary automation fails: decisions with exceptions, rules that cannot be maintained and data that cannot be laid out in a table. OpenAI advises starting with workflows that have previously resisted automation and gives three criteria, each with its own example:

1. Complex decision-making. Workflows involving nuanced judgment, exceptions or context-sensitive decisions. OpenAI's example: refund approval in customer service.
2. Difficult-to-maintain rules. Systems that have become unwieldy due to extensive and intricate rulesets, making updates costly or error-prone. Example: performing vendor security reviews, that is assessing whether a new supplier meets the company's requirements.
3. Heavy reliance on unstructured data. Interpreting natural language, extracting meaning from documents or interacting with users conversationally. Example: processing a home insurance claim.

OpenAI adds a precondition: before you commit to building an agent, validate that your use case clearly meets these criteria. Otherwise a deterministic solution may suffice, that is one that always does the same thing with the same data.

Anthropic describes the same area from the side of the task. An agent suits open-ended problems where it is hard or impossible to predict the number of steps and you cannot hard-code a fixed path. It adds the most where the task combines conversation with action, has clear success criteria, enables a feedback loop and includes meaningful human oversight. As an example it gives customer support, where success can be measured clearly by resolutions.

The fourth condition concerns limits: a person approves what cannot be undone. In the same guide OpenAI writes that actions that are sensitive, irreversible or high-stakes should trigger human oversight until confidence in the agent's reliability grows. It lists cancelling customer orders, authorising large refunds and making payments. An agent can therefore prepare a refund, gather documents and propose a decision, but the "approve" button stays with an employee.

A test for your process: can you write down the rule? Describe the process as a list of "if… then…". If the list closes and covers nearly every case, you need automation, not an agent: it will be cheaper, faster and predictable. If the list grows with every new case and people settle the exceptions anyway on the basis of emails and documents, you have a candidate for an agent. On one condition: the systems the agent is to use must have APIs.

When an agent does not make sense

In many business tasks an agent is not needed — the model vendors say so themselves. Anthropic recommends finding the simplest solution possible and increasing complexity only when needed, "which might mean not building agentic systems at all". For many applications, it adds, optimising single model calls with retrieval and in-context examples is usually enough. Agentic systems often trade latency and cost for better task performance, and you decide whether that trade makes sense.

In its press release of 25 June 2025 Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. That is an analyst prediction, not a measurement. More important is the sentence from Anushree Verma in the same release: many use cases positioned as agentic today do not require agentic implementations.

An agent does not make sense when:

• the process can be written as a rule, in which case automation is enough, for example in n8n or Make, tools we cover in the article on low code and no code;
• you only need an answer to a question based on documents, in which case one model call with retrieval is enough, with no loop and no action tools;
• the systems have no API and the data sits in spreadsheets and mailboxes, so the agent has nothing to act with; how many European companies have systems that can be connected this way is shown with Eurostat data in the article on business process automation;
• the effect of a mistake is irreversible and nobody can approve the decision before it is carried out;
• you cannot say when the task is done: without a success criterion the agent does not know when to finish, and you do not know whether it works.

What an AI agent costs — a token calculation

You pay for an agent by tokens, and an agent uses many times more of them than a chat, because at every step the model receives the whole course of the task so far. A token is a fragment of text in which vendors count what you send to the model and what the model returns; token counts differ between languages. Anthropic quotes prices in dollars per million tokens (Claude API pricing, read 8 October 2026), and OpenAI bills in the same unit (OpenAI API pricing, read 8 October 2026).

We calculated an example on Anthropic's price list: the Claude Sonnet 5.5 model costs $2 per million input tokens and $10 per million output tokens. Anthropic publishes prices in US dollars; we leave them in dollars rather than convert. The token counts are our assumptions, not a market average.

The calculation step by step (our arithmetic):

• Chat reply: 3,000 input tokens (the question, the instruction and document fragments) × $2 per million = $0.006; 300 output tokens × $10 per million = $0.003. Total $0.009.
• Agent task: 10 model calls, on average 8,000 input tokens per step, because the context grows with the results of successive tools, and 500 output tokens. One step is $0.016 + $0.005 = $0.021. Ten steps are $0.21, about 23 times more than a chat reply.
• Month: 1,000 such tasks come to $210.

Anthropic's price list also gives lower rates for batch processing and for data from the cache; our calculation does not include them. On the other hand, an agent that gets stuck in a loop pays for every further turn: in our example about $0.021 per step, and in practice more, because the context keeps growing. The step limit from the previous section sets the upper bound on the cost of one task.

Gartner confirms the order of magnitude. In its press release of 17 August 2026 it writes that routing a task to an agentic reasoning model increases provider inference costs by at least five times compared with a basic chatbot interaction, and often much more as task complexity grows. That is a cost on the model provider's side, not a price for the customer. Running costs already hold companies back: in McKinsey's "The state of AI in 2026" (25 August 2026, 1,719 respondents from 97 countries, self-reported) about 20% said AI-related operating costs, including token costs, constrained their AI use.

Tokens are only the running cost. The cost of building an agent, that is integrations, permissions, tests and an observation period, depends on the number of tools and systems, and we found no independent study of these prices in Europe. One rule holds regardless of price: the budget should be built into the agent itself, not checked on an invoice after a month. In the design of the chat for our own website, exceeding the monthly token budget is one of the conditions after which a person takes over the conversation.

AI agents, the AI Act and the GDPR

The AI Act does not define an agent. An agent falls under the regulation when it meets the definition of an AI system, and the duties depend on what you use it for, not on its being an agent. Art. 3(1) of Regulation (EU) 2024/1689 defines an "AI system" as "a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments". An agent built on a language model should therefore be checked against that definition. The Commission, in its non-binding guidelines on the definition of an AI system, notes that it is not possible to determine automatically which systems fall within the definition, and that only some AI systems are subject to regulatory obligations.

Telling people they are talking to an AI (Art. 50(1)). From 2 August 2026 providers of AI systems intended to interact directly with natural persons must design them so that the persons concerned are informed that they are interacting with an AI system, unless this is obvious. The information has to be given at the latest at the time of the first interaction (Art. 50(5)). The duty rests on the provider of the system. The definition of a provider (Art. 3(3)) covers not only whoever develops an AI system but also whoever "has an AI system developed" and places it on the market or puts it into service under its own name or trademark. We know of no official interpretation for a company that orders an agent from a contractor and runs it under its own brand; in a specific rollout a lawyer decides. In its guidelines on Article 50 of 20 July 2026 the Commission states plainly that AI agents are covered by Article 50(1) if they are capable of interacting with the persons instructing them or with other natural persons in the execution of their tasks, for example making bookings, managing correspondence or executing purchases, and that agents must disclose both their artificial nature and the person on whose behalf they are acting. The guidelines are non-binding. A breach of Art. 50 can bring a fine of up to EUR 15 million or up to 3% of total worldwide annual turnover, whichever is higher (Art. 99(4)); for SMEs the lower amount counts (Art. 99(6)). How this looks in a shop's chat is described in the article on ecommerce customer service.

Staff competence (Art. 4) after the July 2026 amendment. Amending Regulation 2026/1744, in force since 27 July 2026, eased this duty. Providers and deployers of AI systems "shall take measures to support the development of AI literacy" of their staff, and the provision states expressly that the duty does not require them to guarantee any specific level of AI literacy of any individual. In its questions and answers on AI literacy (updated 27 July 2026) the Commission adds that Article 4 does not entail an obligation to measure the knowledge of employees.

High risk from 2 December 2027. An agent becomes a high-risk system through its use. Annex III lists, for example, AI systems intended to be used for the recruitment or selection of natural persons, "in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates". After the July 2026 amendment the obligations for Annex III systems apply from 2 December 2027, and for Annex I systems from 2 August 2028.

Who supervises, and how penalties are set. Each member state designates its market surveillance authorities and lays down the rules on penalties (Art. 70 and 99(1)), so who inspects you and how depends on the member state. The Commission publishes a list of the national single points of contact for market surveillance, updated on 7 September 2026.

GDPR. The European Data Protection Board says in its Opinion 28/2024 (17 December 2024) that supervisory authorities should take into account whether the controller deploying a model conducted an appropriate assessment to ascertain that the model was not developed by unlawfully processing personal data. Hallucinations do not release anyone from the GDPR: in its ChatGPT taskforce report (23 May 2024) the Board recalls that, because of the probabilistic nature of the system, a model may produce made-up output, and that in any case the principle of data accuracy must be complied with.

The roles of provider and deployer, all deadlines and fines for SMEs are explained in the article EU AI Act for business.

This describes the rules; it is not legal advice.

AI agents in European companies — what the data say

There is no data on agents in European companies; there is only a statistic on the use of AI in general, and it shows wide gaps by company size and by country. According to Eurostat, in 2025 AI technologies were used by 19.95% of enterprises with at least 10 employees: 17.00% of small, 30.36% of medium-sized and 55.03% of large ones. Eurostat measures AI technologies, not agents; there is no agent indicator in its table.

The widest gap is by size: large enterprises use AI about three times as often as small ones. Member states differ almost as much: by our calculation from the Eurostat dataset the total runs from 5.21% to 42.03%, so the figure for your country can be far from the EU average. Among EU enterprises that had ever considered AI but did not use it, the most common reason was a lack of relevant expertise (70.89%), followed by lack of clarity about the legal consequences (52.52%) and concerns about data protection and privacy (48.83%), according to Eurostat. Two of the three main reasons are law and data, the topics of the previous section.

Only global surveys speak about agents themselves. In McKinsey's 2026 survey 40% of respondents from large organisations (revenue above $1 billion, 27% a year earlier) and 22% of those from smaller ones, unchanged year on year, reported scaling AI agents. "Smaller" here means revenue below a billion dollars, so these are not SMEs in the European sense.

We bring the wider picture of AI in companies, from assistants in office suites to the rules, together in the section AI in business.

How to start: one process, a person in the loop, a cost limit

Start with one process and the simplest solution that can handle it; add an agent only when a simpler flow does not fit. OpenAI puts it briefly in its guide: start small, validate with real users and grow capabilities over time. In practice that is six steps:

1. Choose a process that meets the conditions from the section on when an agent makes sense: decisions with exceptions, rules that cannot be maintained, a lot of text and documents.
2. Build the workflow first. Write down in a workflow everything that can be written down, and insert the model only where a rule is not enough. That is how we build process automations: the flow first, the model only in the step that needs it.
3. At the start give the agent read-only tools: searching documents, viewing an order, free time slots.
4. Let a person approve irreversible actions: a refund, a payment, a cancellation, sending something to a customer.
5. Set limits: steps per task and a monthly token budget, with the case handed to a person once they are exceeded.
6. Review conversations and logs regularly before you widen permissions.

This is how we design the AI chat for our own website. We are deploying it on our site and there are no results yet, so we describe the assumptions, not the effects. The bot's first message says that an AI is answering. During working hours a person takes over the conversation in six situations: when the visitor asks for it, when the bot finds no answer in the knowledge base or has low confidence, when the visitor is rated as a high-value prospect, for a complaint and for legal or financial topics beyond what we have published, when the writer is already our client, and when the monthly token budget is exceeded. Outside working hours the bot does not hand the conversation over automatically, says so plainly and offers a free meeting slot. Once a person has taken over, the bot stays silent. The CRM receives a summary of the conversation, and the full transcript stays in the support mailbox. After launch we will observe the bot for at least four weeks and check every week where it made something up; only then will we propose a similar solution to clients.

By Anthropic's definition this chat is closer to a workflow than to an autonomous agent: it answers from a knowledge base and uses a few tools, such as booking a meeting, but within a fixed flow. It is a deliberate choice, in line with the model vendors' own advice: as much autonomy as the process requires, and not one step more.