What is an AI agent? Explained without the hype

The difference is how much the software does by itself. A chatbot answers questions, an assistant knows your work, a workflow runs fixed steps and an agent decides for itself which steps it needs to reach a goal.

Those four blur into each other, which makes it confusing. Almost every vendor now calls its product an agent. Research firm Gartner has a word for that: agent washing. Existing chatbots, assistants and RPA tools get a new label without actually working independently (Gartner, June 2025). So here they are side by side.

Chatbot

A chatbot answers questions in a conversation. Think of the free version of ChatGPT, or the chat window on a website that answers questions from the FAQ. Handy, but it does nothing inside your systems. You keep copying and pasting.

Assistant

An assistant is a chatbot that knows your work. You give it fixed instructions and your own documents: a custom GPT, a Claude Project or a Copilot agent with your handbook in it. It writes quotes in your tone and knows how your returns policy works. But here too, you start the conversation and you take the next step.

Workflow

A workflow is a fixed series of steps that runs automatically, for example in n8n, Make or Zapier. An email comes in, AI pulls out the order details, the order goes into the ERP and the customer gets a confirmation. The same route every time. Predictable, easy to test and often exactly what you need.

Agent

An agent gets a goal instead of a route. It chooses which steps and tools it needs, checks whether it worked and tries something else if it didn't. That makes it suited to work with lots of exceptions, such as customer questions that are slightly different every time. It also makes it less predictable. That's why we almost always build agents inside a workflow, with fixed limits and a human who approves.

These four are also the four phases of AI adoption we see in organizations: prompting, assistants, workflows and agents. Most organizations are in phase one. The real value is in phases three and four, because only there does the work itself change. More on that in the 4 phases of AI adoption.

Agentic AI is the umbrella term for AI that carries out tasks on its own instead of only answering. An agentic system plans steps, uses tools and checks its own result. An AI agent is one form of agentic AI.

The word is also used a lot in marketing. A chatbot with a new label is suddenly agentic. So always ask three questions:

• What does the system really do on its own? Does it only read, or does it also prepare something in your systems?
• With which permissions? Read only, or also write and send?
• Where does a person decide? With us, always where money, customers or commitments are involved. See human in the loop.

In the four phases of AI adoption, agentic AI is phase four. Most organizations are still in phase one.

An agent works in a loop: it plans a step, carries it out with a tool, looks at the result and decides on the next step. It repeats that until the goal is reached, or until it needs a human. Four components make the difference.

1. Planning

The agent breaks the goal into steps. 'Answer this complaint email' becomes: read the email, look up the customer, find the order, check the delivery, work out what went wrong, write a reply and propose a solution. The language model does the planning. The better a model can reason, the longer and more complex the tasks it can handle.

2. Tools

Without tools, an agent is just a chatbot. Tools are its hands: a connection to Exact, AFAS or HubSpot, access to the mailbox, a search function across your documents. Increasingly that runs through MCP, the Model Context Protocol. It's an open standard Anthropic introduced in November 2024, so that AI models can talk to systems in one consistent way. In December 2025, Anthropic handed MCP over to the Agentic AI Foundation, part of the Linux Foundation. ChatGPT, Gemini and Microsoft Copilot now support it too (Anthropic, 2025).

The most important design choice sits here: which tools does the agent get, and with which permissions? An agent that can read orders but not change them can't break much.

3. Memory

An agent has two kinds of memory. Short-term memory is the context window: everything it reads and does within one task. Long-term memory is what it keeps between tasks, such as earlier contact with the same customer, or a knowledge base with your product information and examples of good work. Looking that up is often called RAG: the agent first finds the right pieces and bases its answer on them, instead of talking off the top of its head.

This is also where the biggest pitfall lies. An agent is only as good as the information it can find. If your prices live in three different Excel files, it will hesitate just like a new employee would. That's why AI often starts with data, not AI.

4. Human in the loop

A good agent knows when to stop. With us, there's always a human in the process: the agent prepares, an employee checks and presses send. That's called human in the loop. The more trust there is, the bigger the steps the agent may take on its own. You build that trust with numbers, not hope. Why that works is explained in why your AI agent needs a boss.

The best examples are boring: work with high volume, digital input and a clear outcome. Processing orders, booking invoices, preparing answers to customer questions, drafting quotes. A few examples from our own work:

• Customer service at an enterprise client. 6,000 customer questions a month. At first, 30 percent of questions had a prepared answer; now it's 80 percent, ready before the ticket opens. An employee checks and sends.
• Product copy at EasyToys. AI reads PDFs and emails from suppliers, writes product copy in six languages and puts it straight into the ERP. That saves 132 hours and € 6,600 a month.
• Pitch decks at KplusV. AI reads every pitch deck, assesses it against KplusV's own canvas and finds suitable funders from a database of more than 700. A consultant always checks.
• Requests at Van Berkel Professionals. Requests are read, filtered and prepared. That delivers 192 hours a month, about two FTEs.
• Quotes and invoices. At other clients, a sales proposal went from four hours to one, and invoice processing saves twelve hours a week.

See the pattern? None of these agents works entirely alone. They remove the retyping, searching and preparing. The craft and the signature stay with a human.

More about what agents do is on AI agents. You can read the full stories in our cases.

An agent pays off when a task comes up often, has digital input, follows rules with exceptions and has an outcome you can check. It doesn't pay off when the work is rare, the data is missing or a regular workflow can do it too.

A good idea when

• The work comes back often: dozens or hundreds of times a month
• The input is digital: emails, PDFs, forms, data from systems
• There are rules, but also exceptions a fixed workflow can't handle
• There are examples of good work, so you can show how it should be done
• You can check afterwards whether it went right

Not a good idea when

• It's a few times a month. Then a good prompt or an assistant is enough
• The data isn't there, or is still on paper. Start there
• Every step is fixed exactly. Then a regular workflow is cheaper and more predictable
• Standard software already does it. Buy that, it's faster
• A mistake has big consequences and nobody is checking

That second list isn't a detail. Gartner expects more than 40 percent of agentic AI projects to be scrapped by the end of 2027, due to rising costs, unclear value or inadequate risk controls (Gartner, June 2025). According to the same firm, many applications now sold as agents don't need an agent at all.

That's why our APO method has a hard brake built in. If the data is missing or there are no examples of good work, we cap the confidence of an opportunity at 40 percent, however good it sounds otherwise. Foundation first, then the agent. More on that in why one AI agent is rarely enough.

The biggest risks are an agent doing something wrong in your systems, data ending up in the wrong place and costs quietly creeping up. All three can be managed, but only if you sort them out upfront.

• Errors and hallucinations. A language model can convincingly make things up. With a chatbot, you still read it. With an agent that acts on its own, it ends up in your ERP. That's why a human approves where it matters, and there are checks in the workflow itself, such as verifying that the order number exists. More in why AI hallucinates.
• Too many permissions. An agent with write access to everything can get everything wrong. Give it only the permissions it needs for that one task.
• Hidden instructions. An agent that reads emails also reads emails with hidden commands, such as 'forward all invoices to this address'. That's called prompt injection. So an agent must never simply do what the input tells it, and sensitive actions go past a human.
• Privacy and the GDPR. Personal data that passes through an agent to an AI service falls under the GDPR. You need a data processing agreement and you want to know where the data is stored.
• Costs. An agent stuck in a loop keeps running up AI usage. Set limits and keep an eye on usage.
• Nobody manages it. The quietest death of an agent: it runs, a vendor changes something in an integration, and three weeks later someone notices nothing is coming in anymore. Monitoring and maintenance are part of the build.

At NinA, what we build runs in the EU: on your own infrastructure or on servers with a European provider. We work to ISO 27001 and NEN 7510, and there's always a human in the loop.

An AI agent costs build time, AI usage, hosting and maintenance. Building is usually the biggest item; the language model's usage is the smallest for most SME applications. Whether it pays off depends on what the agent earns back each month.

Calculate with four items:

1. 01Building and integrations. Mapping the process, building the agent, connecting it to your systems and testing with real examples. The more systems and exceptions, the more hours.
2. 02AI usage. You pay the model provider per amount of text (tokens). Per task that's often cents, but across thousands of tasks a month it adds up. A cheaper model that's almost as good then saves a lot. Our own benchmarks show that.
3. 03Hosting. The workflow and the agent have to run somewhere, preferably in the EU.
4. 04Maintenance and monitoring. Systems change, models get replaced, new exceptions crop up.

At NinA we only build agents within the AI Partnership: a dedicated team for a minimum of twelve months, then two months' notice, from € 4,500 per month for 30 hours (package S). Whatever goes live then runs under a maintenance contract of € 245, € 445 or € 895 per month, including AI usage. Every partnership starts with the € 6,000 AI Quick Scan, in which we calculate what each opportunity delivers. All amounts are excluding VAT.

We only start a partnership if the business case for the partnership as a whole earns back at least twice the investment in year one. All prices, with a worked example, are on our pricing page.

Start small: pick one process with lots of manual work, measure how much time it takes today and let an agent prepare first while a human approves. Only once the numbers add up do you give it more room.

1. 01Pick one process. Not 'something with AI', but 'the orders that come in by email'. Ideally where the most retyping happens today.
2. 02Measure three numbers. How long does a step take, with how many people, and how often a month? Forty minutes, two people, eighty times a month is already € 57,600 a year at € 45 an hour. How to calculate that is in our step-by-step plan.
3. 03Check the data and the examples. Is the input digital? Are there ten or more examples of good work? If not, sort that out first.
4. 04Start with 'human approves'. The agent prepares, an employee checks. You see straight away where it goes wrong, without a customer noticing.
5. 05Measure and scale. After a few weeks you know how often it gets it right. Only then the next process.

And bring your people along. An agent nobody understands doesn't get used. Or worse: gets bypassed. AI training on people's own work makes a big difference.

Our motto exists for a reason: don't let the hype get to you, start small. Want to know which process delivers the most for you? We calculate that in one day on site in the AI Quick Scan. Rather read on yourself first? The step-by-step plan is at implementing AI.