What is an AI agent? How AI agents work, with examples and limits
An AI agent is software that is given a goal and decides for itself which steps to take to reach it: it uses a language model to plan, uses tools such as search, databases or business systems to act, checks the result and repeats until the task is done or a person needs to decide. This guide explains how an AI agent works, what it is made of, the main types, what agents are good and bad at, the risks, and the EU rules that apply.
Eurostat reports that in 2025, 19.95% of EU enterprises with ten or more staff used at least one AI technology, and 5.35% used AI to automate workflows or assist in decision making.
The short version
- A chatbot answers; an AI agent acts.
- An agent runs a loop: understand the goal, plan, use a tool, look at the result, decide the next step.
- It is useful where tasks vary too much for fixed rules but follow a recognisable pattern.
- The main risk is an agent with more access or freedom than the task needs.
How an AI agent works
- Goal. A person or another system gives the agent a task: "find out why this invoice is unpaid and prepare a reply".
- Plan. The model breaks the task into steps and decides which tool to use first.
- Act. The agent uses a tool: searches a database, reads a document, queries an ERP, drafts an email.
- Observe. It reads the result and compares it with the goal.
- Repeat or stop. It chooses the next step, asks a person for approval, or finishes and reports what it did.
Anthropic's engineering article "Building effective agents" (19 December 2024) draws a useful line: workflows are systems where models and tools are orchestrated through predefined code paths, while agents are systems where the model dynamically directs its own process and tool usage. Many useful business systems are workflows with one agentic step, not fully autonomous agents.
The components of an AI agent
- A language model that reads, reasons over and writes text.
- Instructions that define the role, the goal, the rules and when to stop or escalate.
- Tools: functions the agent may use, such as search, reading records, creating drafts or updating a system.
- Context and memory: the task history, retrieved documents and facts from earlier steps, often supplied by retrieval (RAG); see RAG vs fine-tuning for when to retrieve and when to train.
- Permissions: which data and actions the agent may reach, ideally with the rights of the user it acts for.
- Human approval points for actions that are costly or hard to undo.
- Logs and evaluation: every step recorded, and a test set of real cases to measure accuracy.
Types of AI agents
- Assistant agents inside office, browser and coding tools that carry out multi-step tasks for one user.
- Customer service agents that look up orders, bookings or accounts and prepare an answer or action; see customer service automation.
- Research agents that search several sources, compare them and write a summary with references.
- Process agents in the back office: invoice questions, supplier onboarding, document checks, order exceptions, inside company systems.
- Multi-agent systems where several agents with narrow roles hand work to each other, with one coordinating.
The difference from a chatbot is explained in chatbot vs AI agent, and the difference from rule-based robots in RPA vs AI.
Benefits and limits
Agents handle work that is too varied for fixed rules: emails in any wording, documents in any layout, cases that need a few lookups before anyone can act. They can collect the facts across several systems in seconds and hand a person a prepared case instead of a blank screen. They work outside office hours and scale with volume.
The limits matter as much. Language models can be wrong with confidence, so accuracy must be measured on your own cases, not assumed. Each extra step adds a chance of error, so long autonomous chains are less reliable than short ones. Agents are slower and more costly to run per task than a fixed rule, which is why stable, rule-based steps should stay rules or API integrations. And an agent is only as good as the data and systems it can reach.
Risks and how to contain them
The OWASP Top 10 for LLM Applications (2025) names "Excessive Agency" (LLM06) as the vulnerability that enables damaging actions in response to unexpected, ambiguous or manipulated model output. It lists three root causes: excessive functionality, excessive permissions and excessive autonomy. The practical answers follow from that:
- Give the agent only the tools the task needs, and narrow tools rather than general ones.
- Run it with the permissions of the user it serves, not an administrator account.
- Require approval for payments, deletions and messages to customers.
- Treat content the agent reads, such as emails and web pages, as untrusted input that may contain instructions.
- Log every step and review samples regularly.
What the EU rules say
- Definition. Article 3(1) of the AI Act defines an AI system as a machine-based system designed to operate with varying levels of autonomy that infers from its input how to generate outputs such as predictions, content, recommendations or decisions. An AI agent normally fits that definition.
- Transparency. Article 50(1), applying from 2 August 2026, requires systems that interact directly with people to be designed so that people know they are dealing with an AI system, unless that is obvious.
- AI literacy. Article 4 has applied since 2 February 2025. As replaced by Regulation (EU) 2026/1744, it requires providers and deployers to take measures to support the development of AI literacy of their staff and others operating AI systems on their behalf, without requiring a specific level for any individual.
- High-risk uses. Agents used in areas listed in Annex III, such as recruitment or creditworthiness, face the high-risk requirements, which apply from 2 December 2027.
- Automated decisions. Under GDPR Article 22, people have the right not to be subject to a decision based solely on automated processing with legal or similarly significant effects, with limited exceptions.
Our AI governance guide covers the internal rules a company needs around these duties.
When to get a custom AI agent
Agent features in ready-made tools suit personal productivity. A custom agent is worth considering when the task runs inside your own systems, needs your data and permissions, has to follow your approval rules, and must be measured against your own cases. Typical candidates are high-volume back-office tasks where people today collect facts from several systems before they can act.
Our AI agent development service designs and builds these, with EU hosting, human approval where it matters and the code owned by you. Read AI agent development cost for the budget side and how to choose an AI development company before you compare offers. Every guide is in the guides library.
Questions people ask
What is an AI agent in simple terms?
An AI agent is software that is given a goal and works out the steps to reach it by itself. It uses a language model to decide what to do next, uses tools such as search, databases, email or business systems to do it, checks the result and repeats until the task is done or it needs a person to decide.
How is an AI agent different from a chatbot?
A chatbot answers in a conversation; the person does the work. An AI agent acts: it reads data, uses systems and completes steps on its own, often without a chat window at all. Many products combine both, with a chat front end and agent capabilities behind it. Our chatbot vs AI agent guide compares them side by side.
What are examples of AI agents?
Common types are assistants inside office and coding tools that carry out multi-step tasks, customer service agents that look up orders and prepare refunds, research agents that collect and summarise sources, and back-office agents that handle invoices, supplier questions or document checks inside company systems with a person approving the final step.
Are AI agents safe to use in a business?
They can be, when the design limits what they may do. Give an agent only the tools and permissions the task needs, require human approval for actions that are hard to undo, such as payments or messages to customers, log every step, and test it on real cases before launch. OWASP lists excessive functionality, permissions and autonomy as the main causes of damage.
Does the EU AI Act apply to AI agents?
Yes, an AI agent is normally an AI system under the AI Act definition. Which duties apply depends on the use: transparency when it interacts with people (Article 50), AI literacy measures for staff (Article 4), and stricter requirements if it is used in a high-risk area listed in Annex III, such as recruitment or credit decisions.
Sources
- Eurostat: Artificial intelligence by size class of enterprise (isoc_eb_ai), 2025 (updated 15 June 2026, read 13 November 2026)
- Anthropic: Building effective agents (published 19 December 2024, read 13 November 2026)
- OWASP Top 10 for LLM Applications 2025: LLM06 Excessive Agency (read 13 November 2026)
- Regulation (EU) 2024/1689 (AI Act), Articles 3(1) and 50(1), Annex III (read 13 November 2026)
- Regulation (EU) 2026/1744 amending the AI Act, new Article 4 and Article 113 (read 13 November 2026)
- Regulation (EU) 2016/679 (GDPR), Article 22 (read 13 November 2026)