An AI agent is a program that uses a language model to complete a task in several steps on its own. It receives an input, decides what to do next, calls tools to do it and delivers a result that either continues on its own or is presented to a person.
What sets an agent apart from other solutions
The term is often used for things that are not strictly agents. Three solutions are regularly confused, and telling them apart helps you with every offer.
Chatbot
A chatbot holds a conversation. It answers questions but usually changes nothing in your systems. It can be part of an agent, but the agent itself needs no chat interface.
Automation
An automation follows fixed rules and always does the same for the same input. It is predictable and cheap to run, but it cannot cope with wording nobody thought of.
AI agent
An agent decides in each case which steps are needed. It costs more to run and is never fully predictable, which is why it is used where fixed rules are not enough.

How to choose between the last two is covered in detail in AI agent or automation.
How an agent works
Every run goes through the same four steps.
- Perceive. The agent receives an input, for example an email, a document or a message, and reads it.
- Decide. The language model determines what it is about and which step makes sense next.
- Act. The agent calls tools, for example checking the CRM, looking up a price or creating an entry.
- Report back. It writes a result that is used directly or presents the case to a person for approval.
What this is made of and how to build a first agent from it is covered in how to build an AI agent.
Examples from business
Sorting enquiries
The agent reads incoming emails, recognises service, location and urgency, creates the contact in the CRM and writes a draft reply that you only need to approve.
Reading receipts
Invoices and delivery notes arrive in changing formats. The agent reads out amount, date and supplier and prepares the booking.
Answering support questions
The agent suggests answers from your documentation and hands over to a person as soon as it is unsure.
Summarising reports
From several sources comes a text with the most important changes of the week, which you review and forward.
In all four cases the start is unstructured text or a document, and for delicate cases a person decides in the end. That is no coincidence but the typical pattern of agents that work in everyday use.
What AI agents cannot do
- They are sometimes wrong. Language models sound convincing even when the answer is incorrect, and without review or handover that goes unnoticed.
- They are not fully predictable. The same input can lead to different results, and where that must not happen, a fixed rule belongs.
- They cost money on every run. Unlike a rule, every decision incurs a cost for the model, which adds up for frequent processes.
- They only know what you give them. Without access to prices, history or policies, the agent guesses.
An agent does not replace a person but takes over the part of a task that consists of reading and presorting. The decision in delicate cases stays with you.
When an agent pays off for you
An agent pays off for processes that occur often, contain unstructured text at one point and where a mistake is either not expensive or caught by a person. Which processes pay off first is shown in the order in which to automate. And if you want an agent built, you will find the service under AI agents.
- An AI agent uses a language model to complete a task in several steps on its own and uses tools to do so.
- A chatbot holds conversations, an automation follows fixed rules, an agent decides in each case.
- A run consists of perceiving, deciding, acting and reporting back.
- Typical use cases start with unstructured text and end with a person for delicate cases.
- Agents are sometimes wrong, not fully predictable and cost money on every run.
- Whether an agent pays off depends on frequency, type of input and the cost of a mistake.
Common questions
Is an AI agent the same as ChatGPT?
No. ChatGPT is an assistant you talk to. An agent can use the same language model but works on a defined task and accesses your systems without anyone typing in a chat.
Do AI agents replace employees?
They mainly take over reading, presorting and entering from employees. The decision in delicate cases and the conversation with customers stay with people, and there is more time for them.
What does an AI agent cost?
There is a one-off cost for building it and a running cost for the language model that grows with the number of runs. The prices for building are on the service page, and the running costs can be estimated reliably after a test phase with real cases.
Is using one compatible with GDPR?
It can be arranged if you plan it from the start: processing in the EU, a data processing agreement with the provider and a decision on which data the agent may see at all. Your data protection adviser makes the binding assessment.
How do I know whether my process suits an agent?
If the start is free text or a document and the process cannot be described as an if-then chain, an agent pays off. In all other cases an automation is simpler and cheaper.



