What Is an AI Agent? Definition, Types and Pitfalls
What is an AI agent? The difference from a chatbot, how agents work, which types exist, the build options and the pitfalls — plus what the AI Act requires.
Gepubliceerd door NordX Consulting — AI bureau voor enterprise bedrijven in Nederland.
An AI agent is a software system that is given a goal rather than a script, and works out for itself which steps to take to reach it. It can call tools, look things up, judge results and adjust its approach — without every step being written out in advance. That is what separates it from a chatbot, which answers, and from classic automation, which follows a fixed path.
AI agent, chatbot or automation: the difference
These three get used interchangeably, which leads to wrong expectations and wrong quotes.
| Receives | Does | Fails at | |
| Classic automation | A fixed path | Exactly the same steps, every time | Any exception |
| Chatbot | A question | Answers, sometimes looks something up | Multi-step tasks |
| AI agent | A goal | Chooses steps, uses tools, adjusts | Poorly bounded goals |
A chatbot answers "what is the status of my order?". An agent is given "resolve this customer complaint" and looks up the order itself, checks the delivery, proposes a resolution and queues the refund.
That makes an agent more powerful and more risky. A chatbot that gets it wrong gives a wrong answer. An agent that gets it wrong takes a wrong action.
How an AI agent works
Four components, present in every serious agent:
1. The model — the language model reasoning about what the next step should be.
2. Tools — the things the agent can actually do: query a system, send an email, create a record. Without tools an agent is a chatbot.
3. Memory — what happened earlier in the task, so step four knows what step two produced.
4. Boundaries — what the agent may not do, and where a human must sign off. This is the component most often missing.
The loop is always the same: look at the goal, choose a step, use a tool, judge the result, repeat. Until the goal is met or a boundary is hit.
Types of AI agent in business
- Customer service agents resolve recurring requests end to end: status, returns, address changes, simple refunds.
- Sales agents follow up leads, enrich records, schedule meetings and keep the CRM current without anyone typing it in.
- Voice agents take calls, qualify and route them. The technology has become usable in recent years, but expectations still tend to run ahead of practice.
- Data agents pull information from several systems, combine it and deliver an answer or report.
- Internal process agents do the work between systems: onboarding, invoicing, completing case files.
Building an AI agent: the options
There are broadly three routes, and they differ mainly in who maintains the result.
No-code and workflow tools. Platforms such as n8n, Make or Zapier let you build agents by connecting blocks, with a model as the reasoning step. Fast for a first version and good for processes with clear steps. You hit limits on complex logic, and you depend on what the platform supports.
Agent frameworks. Libraries for developers to build agents in code. Anything is possible, including proper error handling and tests. Requires development capacity, including after delivery.
Off-the-shelf agents. Vendors offering an agent for one task, usually inside their own platform. Fastest start, least customisation, and your data and logic live with them.
The choice depends less on the technology than on who will still be looking after it in a year. For what an agent of your own costs and returns, see building or commissioning an AI agent.
Where agents come unstuck in practice
A goal that is too vague. "Improve customer service" is not a goal for an agent. "Handle status enquiries through to sending the reply" is.
No boundaries. An agent without explicit limits on amounts, volumes and irreversible actions is a liability. Build in the points where a human must confirm, and build them from the start.
No logging. If you cannot reconstruct afterwards which steps the agent took and why, you cannot investigate an error or demonstrate that you were in control.
Starting with the hardest case. The exception is tempting to start with and returns the least. Start with the volume.
What the AI Act makes of this
An AI agent is an AI system within the meaning of the AI-verordening, so the ordinary duties apply. Two matter immediately:
- Transparency (Article 50) — if the agent talks to people, they must know it is AI.
- AI literacy (Article 4) — the people working with the agent must understand what it does and where it can go wrong.
Deploy an agent for recruitment, credit scoring or comparable decisions about people and you enter the high-risk regime, with considerably heavier requirements. The full overview is in our AI Act compliance guide, and what you record in your AI register applies to agents too.
In short
An AI agent is given a goal and chooses its own steps; a chatbot answers a question and classic automation follows a fixed path. The power sits in multi-step tasks across systems, the risk in actions you did not anticipate. Start with a bounded goal that has real volume, build boundaries and logging from day one, and choose your build route on who maintains it — not on what demonstrates fastest.
Sources
Frequently asked questions
What is the difference between an AI agent and a chatbot?
A chatbot receives a question and gives an answer. An AI agent receives a goal and works out which steps are needed, using tools such as querying systems or creating records, adjusting as it goes. An agent takes actions; a chatbot talks.
What do you need to build an AI agent?
Four things: a model that reasons, tools that let the agent actually do something, memory of what happened earlier in the task, and explicit boundaries on what it may not do. That last one is most often missing and is precisely what makes an agent safe.
Can you build an AI agent with n8n?
Yes. Workflow platforms such as n8n, Make or Zapier let you assemble agents by connecting blocks with a model as the reasoning step. That works well for processes with clear steps. With complex logic or strict error handling you run into the platform's limits.
Does an AI agent fall under the AI Act?
Yes, an agent is an AI system under the AI Act. If it talks to people, the transparency duty (Article 50) applies; the AI literacy duty (Article 4) applies regardless. Deploy it for decisions about people, such as recruitment or credit scoring, and you fall under the high-risk regime.
Related articles
Meer weten? Bekijk onze andere artikelen op het NordX blog of neem contact op via nordx.ai.