The short answer
What is an AI agent, really?
An AI agent runs on the same AI you know from ChatGPT, Copilot or Claude. What makes it an agent is what you build around it: what it must do, where it may look and what it delivers.
Give the same system other instructions and other data, and you have another agent. One summarizes quote requests, another prepares purchase invoices for the bookkeeping. The difference lies entirely in what you put around it.
The difference
What is the difference between ChatGPT and an agent?
The difference between ChatGPT and an agent is the difference between a question and a task. A chatbot answers the question you ask and then waits for the next one. An agent gets a recurring task and works through the steps until the job is done.
Take a quote request. With ChatGPT you paste in the email and ask for a summary, every single time. An agent picks up each request itself, takes the prices from your own list, prepares the quote in your format and puts it in front of the salesperson.
A digital employee comes into being: an employee made of software that carries a fixed workflow, or a bounded part of it, from beginning to end.Paul Hoffman
In my book such an agent is called a digital employee. The word agent comes from the tools and the trade press, and it means the same thing. With an employee you naturally think of a job description and of someone who checks the work, and an agent needs those too.
The building blocks
What is an AI agent made of?
An AI agent consists of the AI itself and five things you build around it. In my book that whole is called the scaffolding. It turns general AI into an agent for your company and your work.
- Instructions: what it must do, for whom, in what tone and what to watch for, in plain language.
- Data: the documents and lists it draws on, such as your price list, your terms and earlier quotes.
- Skills: fixed ways of working that several agents can share, such as how your company builds a quote.
- Tools: what it may do, such as read an email, save a file or look at a calendar.
- Connections: the software it talks to, such as your accounting package or your planning.
And you steer it not with code, but with words.Paul Hoffman
That last point is what sets it apart from software as you know it. You do not write in a programming language what the agent must do. You explain it the way you would explain it to a new colleague, and whatever you change applies at once.
My reading
Where does the agent sit on the five levels of AI?
The agent belongs to level 2 of the five levels of AI: one fixed task with fixed instructions, your own data and an owner who keeps an eye on it. On level 1, people still ask an AI assistant separate questions.
On level 2 the work keeps running the way it always did. The agent becomes a new link in the existing chain and takes repetitive work out of it. That is a good first step, but the big gain sits one level higher, when you redesign the work itself.
In my book I call the first months of an agent its apprenticeship. Every mistake your expert corrects goes into the instructions and stays there, even if that expert ever leaves. That time costs more than it returns at first, and yet it is what the first agent leaves you with.
What it can do
How independently do agents work today?
Agents work on their own for longer and longer, but they do not notice their own mistakes. Research organization METR measured that the length of tasks AI can finish on its own has doubled roughly every seven months since 2019.
That measurement covers programming tasks, and METR itself notes that real work is messier than a test assignment. What an agent could not do last year, it can often do now. How often it gets it right differs per task, and you only measure that with your own examples.
It forgets nothing and never leaves, but it never notices when it is wrong. That is why the person signs.Paul Hoffman
According to CBS, Statistics Netherlands, 29.8 percent of small and mid-sized businesses used some form of AI in 2025, and 13.8 percent of companies with two to nine people. Those figures say nothing about agents yet. They do show that most companies still have to start, including with the first step.
In practice
Where do you start with a first AI agent?
Start with one task that comes back every week, that takes time now, and where you can see whether the result is right. Summarizing quote requests or preparing purchase invoices are good candidates.
Write down how that task is done today, as if you were explaining it to a new employee. Collect a set of real examples with the right answer next to them, so you can measure whether the agent gets it right. Appoint an owner who checks everything in the first weeks.
Want to know which tasks in your company pay off most? That is what the BORS scan is for: we go through your work processes together and look at where an agent saves time and where a person must keep signing off.
Frequently asked questions
Is ChatGPT or Copilot already an agent?
No, not the way most people use them. You ask a question, you get an answer, and the program waits for your next question. Most of these programs can now carry an agent: you give them fixed instructions, access to your documents and a recurring task. Setting that up is what makes it an agent, not the subscription.
Do I need to be able to code to build an agent?
No. You steer an agent with instructions in plain language, the way you explain to a new colleague what is expected. Code only comes in when the agent has to talk to software that has no connection for it yet. Most of the work is writing down exactly how the task is done.
Is an AI agent the same as agentic AI?
No, the second is broader. In a 2025 review, Sapkota and colleagues describe an AI agent as a system for one bounded task. In their taxonomy, agentic AI is several agents working together, splitting a large goal into pieces themselves and sharing a memory. For most companies it starts with that one bounded task.
May an agent access my customer data?
Yes, if you arrange it the way you would for a new employee. Give it access only to the data the task needs, use business accounts where your data is not used to train the model, and record who manages it. For personal data the rules of the GDPR apply, also when software does the work.
What does an AI agent cost?
The AI itself is cheap: a business subscription to ChatGPT, Copilot or Claude costs a few tens of euros per user per month. The real cost is the time you spend in the first weeks writing instructions and checking the output. That time is also where the value comes from, because that is where you record how your company does the work.
Sources
The idea, the concept and the corrections of substance are Paul Hoffman’s. AI corrected the language, ordered the text and improved readability. The quotes come from his book De Informatiemotor (2026). The facts were checked on November 2, 2026 against the sources below.
- CBS (Statistics Netherlands), March 16, 2026: Use of AI technology by Dutch micro-enterprises, chapter 2: Use of AI technology (in Dutch)
- METR (Kwa, West et al.), arXiv, February 25, 2026 (version 3): Measuring AI Ability to Complete Long Software Tasks
- Sapkota, Roumeliotis and Karkee, arXiv, September 30, 2025 (version 5): AI Agents vs. Agentic AI: A Conceptual Taxonomy, Applications and Challenges