What is an AI agent?
An AI agent selects steps and tools needed to achieve a goal within agreed boundaries. The difference from a chatbot and automation, and when it makes sense.
An AI agent is a system based on a language model that selects the next steps and tools needed to accomplish a given goal. A chatbot primarily holds a conversation; an agent works in a loop: it plans, acts, observes the outcome, and decides what to do next or when to ask a person for approval. This distinction follows the definition of agents used by Anthropic (2026).
That difference affects cost, risk, and how the system is tested: the more decisions a system makes while operating, the more important limits, observability, and manual takeover become.
How is an agent different from a chatbot and automation?
Three things that are all called “AI” in sales conversations, even though they are entirely different tools:
| Automation | Chatbot | AI agent | |
|---|---|---|---|
| Who decides the sequence of steps | you, in advance | there are no execution steps | the model, at runtime |
| Reaction to an exception | follows a rule or escalates | gives an answer without acting | can choose another step or escalate |
| Outcome | data in a different place | an answer | a completed task |
| Cost to maintain | usually lower for a comparable, narrow scope | depends on model and channels | rises with the number of tools, decisions, and safeguards |
| Predictability | high within the scope described by rules | depends on model and context | requires measurement, limits, and supervision |
The conclusion for a first conversation is: if a process can be reliably described with “if A, then B” rules, start with ordinary automation. It will usually be simpler to test and maintain. Consider an agent when the next steps depend on information found along the way and tests show an advantage over a fixed flow.
How does an AI agent work in practice?
We include four elements in a typical production-agent project:
- A goal — expressed as an outcome, not an instruction. “Reply to a customer ticket and close it if the matter is resolved,” rather than “write an email.”
- Tools — functions the agent can call: reading a database, sending an email, creating a CRM task, searching documentation. An agent without tools has no way to do anything.
- Context and access to knowledge — information needed for the current task and, where the process requires it, state retained between steps. This can include RAG, which retrieves material so the model receives an up-to-date basis for an answer.
- Safeguards and supervision — a list of actions requiring human approval, a log of calls and outcomes, cost limits, and a panel that shows the task’s progress.
Without the fourth point, it is impossible to reconstruct an error reliably or determine whether the agent exceeded its scope. The larger the consequence of an action — sending a message or changing a customer record, for example — the narrower permissions should be and the clearer the approval gate needs to be.
Three scenarios in which an agent is worth calculating
Handling a ticket from start to finish. A ticket arrives by email. The agent can determine its category, check the knowledge base and customer history, prepare a response and, for tested routine cases, send it and close the ticket. Non-standard cases go to a person with a summary. The benefit needs to be measured as a change in handling time, quality, and number of escalations — not assumed in advance.
Processing documents with verification. An invoice, agreement, or report. The agent extracts data, checks it against an order, spots a discrepancy, and reports it instead of entering an incorrect amount and continuing. The second part is crucial: extraction without verification is only faster error entry.
Analysing sources and preparing a decision. The agent collects data from several sources, organises it, prepares a comparison, and identifies what could not be established. It does not make the decision — it shortens the route to one.
In these scenarios, it is worth starting with a supporting role: the agent prepares or carries out reversible steps, while a person resolves exceptions and high-stakes actions. The degree of autonomy can increase only after quality has been measured.
What goes wrong in agent implementations
Four risk patterns worth checking — we also see them in our own AI research workflows:
- A goal that is too broad. An agent that is supposed to “serve customers” has nothing actionable to perform. An agent that is supposed to “answer questions about order status” does. The first process must be narrow.
- No data on which the agent can work. If knowledge lives in people’s heads, the agent has nothing to use. The first step is then not an agent but organising sources — a second brain.
- A pilot with no route to production. A prototype may not include observability, cost limits, permissions, or failure handling. Moving to production needs to be priced as a separate scope.
- No measurement. Without a benchmark (“how long did this take before?”), there is no way to tell whether the agent helped. Measure before, not after.
When an AI agent is not a good fit
We say this directly, because it is better to lose an assignment than sell something that will not pay back:
- A process performed rarely and with low manual cost. Time savings may not repay construction and maintenance; costly mistakes, compliance, or strategic process value can change the decision. Calculate it as an investment, not a toy.
- A process that requires certainty as a rule. A model may prepare data for a tax calculation, transfer, or HR decision, but deterministic code or an authorised person should perform the final rule.
- A process without access to data. If an agent lacks the required information, it should refuse, request data, or escalate. A system that produces an answer despite having no basis is not production-ready.
- A company without a process owner. A production agent needs somebody responsible for alerts, exceptions, quality review, and decisions about changes; this does not mean manually reading every log.
What does it cost?
Our ranges as of August 2026: a simple on-site assistant 3,000–8,000 PLN net, plus model and maintenance costs; an agent serving a real business process 8,000–30,000 PLN net; an advanced single agent 30,000–120,000 PLN net. A multi-agent system is a separate scope. Data, integrations, permissions, evaluation, and production requirements move the price.
Frequently asked questions
How is an agent different from a multi-agent system?
An agent is one executor. A multi-agent system divides roles between several agents and can add verification stages. Consider it only when a test shows an advantage of that division over a single agent or a simpler flow. More: multi-agent systems.
Can an agent work on our data without sending it out into the world?
It can, but that depends on the architecture, providers, and agreements. The scope of data, processing location, retention, and logs must be agreed before implementation and subject to suitable legal and security review.
Where should we start if we do not know where an agent would help?
With a process review, not technology. We do this in an AI audit — it produces a list of opportunities ranked by profitability, not a list of fashionable phrases.
We build agents with event logs, limits, and observability. We show our own experience in the AI research-workflows laboratory; we do not present it as a client implementation result. If you have a process that wears you down, describe it to us, or see what our AI agents service includes.
Have a problem to solve?
Let’s find the right first step
Describe your situation in a few sentences. We’ll return with questions or a concrete proposal for what comes next.