AI agent specialists in Belgium: who actually builds agents?

If you are looking for an agency to build an AI agent for you, the problem is not that there is too little supply. The problem is that the word “agent” gets attached to two very different things, and that you only notice the difference once it is built.

I searched Belgium for agencies that demonstrably offer AI agents as a service. There are eleven. I also found six that use the word while their own service description is about something else: a chatbot, a voicebot, or a pre-drawn automation. What an AI agent actually is and how it differs from a chatbot or a workflow is fully worked out in what is an AI agent. This article is about the supplier, not about the concept.

How was this list put together?

One testable criterion: the agency demonstrably offers AI agents as a named service on its own site, and the service description is about carrying out tasks independently, not about answering questions. Every entry was checked on the agency’s own site on 9 August 2026.

The disclosure, right away and not in a footnote: ClickForest is on this list and has an interest in you choosing us. That is why there is no quality ranking in this article, only a grouping by profile, and why there is a section at the end about who we are not the logical choice for.

What this overview deliberately does not do. It does not judge results, because none of the agencies on this list publishes verifiable figures from client projects and neither do we. It contains no links to the other agencies, because their name is enough to find them. And being on this list means an agency demonstrably offers the service, not that it is good at it. What you do get here is the test to make that distinction yourself.

Which Belgian agencies actually build AI agents?

Eleven, from founder-led agencies to an IT group of nine thousand people. That spread is wide enough to mention separately: having an agent built by a group that size is a different project from having one built by an agency of two founders. The order is a grouping by profile and explicitly not a ranking.

Two agencies were registered as borderline cases and are deliberately not on the list: they do offer an agent service, but they mix it into the same package with conversational agents, in one case topped up with orchestration in n8n and in the other with a vaguely described process automation, which makes it impossible to see from the outside what you are buying.

And two candidates were rejected because they are not Belgian. Both present themselves with a Brussels landing page; one has its head office in Richmond, Virginia, the other in the United Kingdom with a British phone number. That is exactly the kind of entry an unchecked list falls for.

There is overlap with the overview of AI chatbot agencies, but less than you might think: of the nine agencies there, three also offer an agent service. So an agency can appear on both lists, and that is not a contradiction. They are two different services, and that is exactly the point of the next chapter.

How do you tell from an offer whether you are buying an agent or a chatbot?

By one question: does the thing carry out actions in your systems itself, or does it give answers? That distinction does not come from me. The model builders themselves are unusually explicit about it, and that is the yardstick you can lay against any offer.

OpenAI writes in its own guide to building agents: “Agents are systems that independently accomplish tasks on your behalf.” And then, without hedging: applications that use a language model without letting it control workflow execution are explicitly not agents in OpenAI’s view. It names simple chatbots, single-turn language models and sentiment classifiers as its own examples.

Anthropic makes the same distinction, but between workflows and agents: “Workflows are systems where LLMs and tools are orchestrated through predefined code paths.” Agents are different: “Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.” So ask who decides the order of the steps: the system itself, or the plan the agency drew up in advance. You can see that difference in a quotation without having to read any code. It is not about the tooling. Make and n8n are perfectly good building blocks, and we use Make.com ourselves where it makes sense, but if the pre-drawn flow is the entire product, you are not buying an agent.

From that follow three questions you can ask in a first conversation, and that require no technical knowledge.

Which systems is it allowed to touch? An agent that can reach nothing can also do nothing. NIST sums it up in its AI Agent Standards Initiative: “the real-world utility of agents is constrained by their ability to interact with external systems and internal data.” If you get no concrete answer about connections to your accounting, your CRM or your order processing, you are buying a chat window.

“Always ask which systems the agent is allowed to touch. One that cannot do anything inside your systems is a demo, not an employee.”

— Frederiek Pascal, CEO and founder of ClickForest

What happens when it gets something wrong? An agent that acts independently can cause damage independently. Dimitri Van Landuyt, associate professor of business informatics at KU Leuven, put that risk sharply in a conversation about software companies pushing AI into their products:

“You don’t want the explanation for a mistake to be: ‘It was a hallucination of our AI tool’”

— Dimitri Van Landuyt, associate professor of business informatics at KU Leuven

And the third: what changes about the way the work gets done? McKinsey analysed more than fifty of its own agent projects and made the most important lesson a section heading in the report: it is not about the agent, it is about the workflow. “Often, however, organizations focus too much on the agent or the agentic tool. This inevitably leads to great-looking agents that don’t actually end up improving the overall workflow.” An agency that only talks about the agent and not about your process is building a demo.

How many agencies carry the label without the service?

Six of the nineteen Belgian providers that use the word agent, so roughly one in three. They call the offer an agent and then describe it as something else, usually in the same paragraph and without any bad intent. The word has simply become vague, and someone who builds a chatbot that works well is still selling something useful. It is just not what you thought you were buying.

One French-speaking agency calls its service “agents IA conversationnels”, conversational AI agents, and works it out as a WhatsApp chatbot and a voice agent for incoming calls; the actual automation sits in a separate service that does not fall under the agent label. Another describes, also in French, an agent that “answers your calls and messages, qualifies the requests and transfers them to the right people”, and that is a virtual receptionist. A third puts the heading “What is an AI agent?” above an offer that then delivers “AI workflows” with Make, n8n and ChatGPT.

I am not naming them. They are not lying, they are using a word the market itself has made unclear, and you can moreover see it yourself in two minutes with the questions above. But it happens often enough to ask about.

There is a name for it too. Gartner calls it agent washing, the relabelling of existing products such as assistants, RPA and chatbots without substantial agentic capabilities. Its own estimate, and that is an American estimate from June 2025 that you cannot simply transfer to Belgium: of the thousands of vendors calling themselves agentic, roughly one hundred and thirty are real. Analyst Anushree Verma is just as direct about it:

“Most agentic AI propositions lack significant value or return on investment (ROI), as current models don’t have the maturity and agency to autonomously achieve complex business goals or follow nuanced instructions over time. Many use cases positioned as agentic today don’t require agentic implementations.”

— Anushree Verma, Senior Director Analyst at Gartner

That last sentence is the most important one in this article: many applications sold as agentic today do not need an agent at all. A good agency tells you that, and then delivers the simpler solution.

Is there a certification or ranking for AI agent agencies?

No. No certification, no independent ranking, no body that could issue one today. Anyone claiming to be accredited should be asked by whom exactly and what that accreditation actually tests. Usually the answer turns out to be a commercial arrangement.

What does exist, and what is not a quality judgement:

• Partner networks of the model builders. OpenAI launched a partner network in June 2026 with tiers and a specialisation around agents; Anthropic has a comparable programme with a public partner directory. OpenAI’s launch partners are global consultancies such as Accenture, McKinsey and PwC. In Anthropic’s public directory I found no Belgian company on 9 August 2026. So a tier is a commercial relationship with a vendor, not a judgement about your project.
• Commercial directories that publish “top ten” lists without published selection criteria. Several were not even reachable during the check, which leaves their methodology unconfirmed.

At ClickForest, an AI agent for SMEs in Flanders is therefore never a badge but a working system: what you buy is an agent that does something inside your own systems, not a label.

What can you expect from an agent project, and what not?

A working system for one clearly defined task, not a digital employee. That last phrase circulates widely, including in the AI answers search engines give about Belgian agencies today, and it raises expectations an agent does not currently meet. An agent does not take over a job; it takes over an action you repeat yourself today.

The figures help you calibrate your expectations, but read the label with them. Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027, due to rising costs, unclear business value or inadequate risk controls. That is a prediction by a commercial analyst house, not a measurement. McKinsey did ask: 62% of respondents say their organisation is at least experimenting with agents, but for no single business function do more than 10% of them say the organisation is scaling them. It is self-reporting, not a measurement of running systems. Michael Chui of the McKinsey Global Institute sums it up soberly:

“AI agents have been the subject of intense buzz and excitement. Already, about a quarter of our survey respondents report that they have started scaling at least one agentic AI system, but usually only in one or two business functions. […] When it comes to agents, it takes hard work to do it well.”

— Michael Chui, Senior fellow at the McKinsey Global Institute

The most honest advice comes from Anthropic itself, which would stand to gain from selling you an agent: “we recommend finding the simplest solution possible, and only increasing complexity when needed. This might mean not building agentic systems at all.” Agents trade latency and cost for better task performance, and they compound errors. For a fixed, predictable task an ordinary automation is cheaper and more reliable, and that is the territory of marketing automation.

And the Belgian context, with the right scope attached: 34.5% of Belgian enterprises with ten or more employees used at least one AI technology in 2025, against a European average of 19.95%. Among small enterprises in the EU, those with ten to forty-nine employees, it is 17%. A Belgian figure specifically about AI agents does not exist: the official statistic measures whether you use AI, not whether that system acts independently.

One figure from another study puts that in perspective. The European Central Bank surveyed around six thousand firms across twelve euro area countries and found that 80% of them underestimate how many peers of their own size and in their own country already use AI, by eighteen percentage points on average. So there are probably more of them than you think, and that is exactly the reason to ask the question now rather than next year.

How do you recognise an agency that fits your scale?

By what it asks before it proposes anything. An agency that starts with the tool, or with the agent, skips the step that according to the McKinsey research makes the difference: which workflow you want to change and who does that work today.

In practice you can distinguish three profiles in the list above. Large IT houses such as Cegeka, Savaco and AE have the capacity for a rollout across several departments and the governance that comes with it. AI engineering houses such as ML6 and Faktion build bespoke and sit close to the research. Small, founder-led agencies deliver quickly on one clearly defined task, with the person who sold it also doing the work. None of the three is better; they suit a different question and a different budget.

If you want to explore it yourself first before bringing anyone in, you can: how to build an AI agent yourself with Claude Code and MCP explains the building blocks, what MCP actually is covers the connection to your own systems, and agentic engineering in practice shows what that work looks like day to day. If you would rather have your own people learn it than outsource it, AI training for your team is the route. And if you are not really looking for an agent but for someone to set your AI course, AI consultancy for SMEs is the better entry point.

Who is ClickForest the right choice for, and who not?

For SMEs and scale-ups in Flanders that want to automate one clearly defined task with an agent that genuinely works inside their systems, with the person who sells it also at the keyboard. ClickForest builds founder-led from Bonheiden near Mechelen, with Claude Code and with connections to your existing systems.

What that looks like in practice is in the order reporting we built for Bastiano and on the page about bespoke AI agents.

And honestly, three situations where you should look elsewhere. If you want a rollout across several departments at once, with formal governance and a support organisation of its own, a house like Cegeka or AE suits you better than a founder-led setup. If you are in a project where research and model development are the centre of gravity rather than integration, ML6 or Faktion are more obvious. And if the first conversation shows your task is fixed and predictable, I will not deliver you an agent but an ordinary automation, because that is cheaper and breaks less often.

If you are actually looking for something else, comparable overviews exist for AI chatbot agencies and for GEO agencies in Belgium. If you sell online and want to know how agents affect your findability, agentic commerce is the starting point.

Conclusion: which question do you ask in that first conversation?

These two: which systems is it allowed to touch, and what happens when it gets something wrong. They require no technical knowledge, they can both be asked in a single conversation, and together they expose the difference between a real agent and a chatbot with a new label. An agency that gives no concrete answer to either has probably not built the system yet.

What struck me most while compiling this is not that the supply is thin. There are eleven, more than I expected. What stands out is that roughly one in three uses the word differently from how the makers of the technology define it themselves, and that no certification exists to help you with that. ClickForest builds AI agents for SMEs in Flanders, so read this with that in mind. The terms in this article are explained in the glossary.