What is an AI agent? Explanation, examples and pitfalls

In brief

An AI agent is software with a language model that chooses the steps towards a goal by itself. It looks up information, uses tools and checks its own intermediate results.
The difference with a chatbot: a chatbot answers one question, an agent completes a task in several steps.
Several agents can work together, each with its own role. One agent directs the others.
An agent makes mistakes the way a language model does. That is why a person should approve the plan and check the result.
Start with a well-defined task, real sources, limited permissions and a fixed moment at which a person decides.

What is an AI agent?

An AI agent is a program that uses a language model to carry out a task on its own. You set a goal, and the agent decides which steps are needed. It can request information, read documents, run a calculation and write a draft.

An agent has four parts:

A language model that reasons and writes text.
Tools: connections to sources and systems, such as a legislation database, a case management system or a calendar.
A goal and instructions: what should come out, and within which limits?
A loop: the agent takes a step, looks at the result and chooses the next step, until the goal is reached.

The Model Context Protocol (MCP) is the open standard an agent uses to call tools. In December 2025 MCP moved to the Agentic AI Foundation of the Linux Foundation.

What is the difference between a chatbot, a copilot and an AI agent?

The difference lies in who chooses the steps.

	Chatbot	Copilot	AI agent	Classic automation
What does it do?	Answers one question	Helps within one task, such as writing or summarising	Completes a task in several steps	Runs a fixed sequence of steps
Who chooses the steps?	The user	The user	The agent, within limits	The person who built the process
Uses tools?	Usually not	Limited	Yes	Yes, configured in advance
Handles exceptions?	No	Limited	Yes, but can get it wrong	No

Classic automation (such as RPA or a fixed workflow) is predictable but rigid. An agent is flexible but less predictable. In practice you often combine them: fixed steps where possible, an agent where judgement is needed.

What can an AI agent do for your organisation?

An agent mainly helps with work that involves a lot of searching, reading and combining. Examples:

Legal: finding relevant statutory provisions and case law and making a first analysis of a case file.
Construction: setting a building plan against the rules of the Dutch Buildings Decree (Bbl) and pointing out open issues.
Care: retrieving the client record and preparing a draft report.
Government: setting a council proposal against earlier decisions, the statutory text and the budget.

An agent takes over search and preparation work. A person checks and decides. That is not a formality: a language model can invent something that sounds logical. See also who is liable when an AI agent makes a mistake.

How do multiple AI agents work together?

For a larger task you split the work across agents that each have their own role, as in a team. A common pattern:

An orchestrator receives the assignment, makes a plan and divides the work.
Specialists each carry out part of it, for example searching sources, calculating or writing.
A counter-role looks at the work from the other side, such as the opposing party or the regulator.
A reviewer checks the whole for errors and missing sources.
A person approves the plan and assesses the final result.

How to decide where in a process the person decides and where the agent works is covered in Human in the loop: hybrid agent workflows.

What should you watch out for when deploying AI agents?

Start small. Choose one task with a clear beginning and end, and measure whether it works.
Connect real sources. An agent without sources fills gaps with whatever sounds likely.
Grant as few permissions as needed. An agent that may read does not need to be allowed to send or delete.
Fix the decision point. Decide in advance which steps a person approves, especially anything that goes outside or changes something.
Log what happens. Which sources were used, what did the agent do and who approved it?
Build knowledge. Since 2 February 2025, organisations that use AI must take measures for the AI literacy of their people (Article 4 AI Act). More in AI literacy under the EU AI Act.

How it works at Prudai

Prudai builds AI colleagues per field: LEO for legal work (with LEO Fiscaal for tax advisers), VERA for construction and the Dutch Building Quality Assurance Act (Wkb), ZIA for care and welfare, BEVER for government, IRMA for governance, risk and compliance, and MAIA for management information. LEO, VERA, ZIA, BEVER and IRMA work on one shared source registry with more than a hundred validated sources.

In LEO, larger tasks run as a workflow with fixed roles:

How a workflow runs in LEO: the user confirms the intake, LEO proposes a plan, a person approves, adjusts or stops it, specialists do the work, a counter-role and a reviewer check it, and the user receives the result.

A person starts it. A workflow only begins when the user confirms the intake. LEO can pre-fill the intake, but never starts a run by itself.
Plan first, then approval. LEO first proposes a plan. In most workflows LEO then waits for a person's approval; that person can approve, adjust or stop. The decision is recorded.
Counter-role and reviewer. Many workflows have a counter-role that looks at the work from the perspective of, for example, the opposing party or the tax inspector. A reviewer then checks.
Checking sources. In the chat, a gate checks every answer for ECLI numbers. If an answer cites a ruling that was not retrieved in the conversation, LEO tries to retrieve it. If that fails, the answer carries an "unverified" notice. The lawyer reads the ruling.

For the specialist

The agent loop. Technically, an agent is a language model that calls tools in a loop (tool calling): the model chooses a tool with arguments, the system runs it and returns the result, and the model continues reasoning. Anthropic distinguishes workflows (predefined code paths) from agents (the model directs its own process), and advises starting with the simplest pattern that works.
Prompt injection. An agent that reads external documents or web pages can come across hidden instructions. Limit the agent's permissions and have a person approve any action that writes or sends.
AI Act. An agent is generally an AI system within the meaning of Article 3(1). For high-risk systems, Article 14 requires effective human oversight. Article 50(1) requires providers to ensure that people know they are interacting with an AI system.
Context. An agent that works for a long time collects a lot of text. What no longer fits in the context window must be retrievable again. That argues for sources with a fixed location, see One fact, one place.

Frequently asked questions

What is an AI agent in one sentence?

An AI agent is software with a language model that chooses the steps towards a goal by itself, using sources and systems along the way.

What is the difference between an AI agent and ChatGPT?

ChatGPT is basically a chat assistant that answers your question. An agent completes a task in several steps and uses tools on its own. Many chat assistants now gain agent features, so the line is shifting.

Does an AI agent replace staff?

An agent takes over search and preparation work, so staff have more time for judgement and contact. Responsibility for the result stays with a person.

Is an AI agent reliable?

As reliable as its sources and the checks on them. An agent can make mistakes or invent things. Real sources, limited permissions and a person who approves reduce that risk.

What does it cost to start with AI agents?

That depends on the task and the sources. Start with one well-defined process, measure the result and expand from there.

Sources

What a language model is and how generative AI works is explained in What is generative AI?.

Would you like to know how Prudai's AI colleagues work in your field? See Prudai's AI colleagues.

Updated on 6 October 2026

What are AI agents and how do they help you move forward?