Roel Verbeeck

The through-line

The recurring position is a deflationary one about AI agents held by someone who builds them: what works today is narrow, embedded in a workflow, and pointed at a problem that already resists conventional methods [2][3]. The interesting claim is not that agents will one day think for themselves but that the base layer is levelling out beneath everyone. "Al die tools worden een commodity en dus de mensen gaan een verschil maken vanuit hun eigen sterkte" [2]. If the tooling is common property, the differentiation has to come from somewhere else, from people working out of their own strengths, and from products given enough definition to stand on their own rather than sitting inside a portfolio [1][2].

On what agents can actually do right now

The concrete example is a branding engagement for Kan Design, where the work was split into three cooperating agents rather than handed to one general-purpose system: a digital twin of the brand, a market agent, and a design agent that combines the output of both into proposals [3]. That architecture is the argument in miniature. Agents earn their place by being decomposed into narrow roles inside an existing workflow, each doing something specifiable, with the combination step made explicit [2][3].

On simulating the market instead of asking it

The market agent exists because survey data has a known defect. People say they would pick the orange speaker and then take the black one home, an effect visible in a Philips test where respondents chose colorful Bluetooth speakers when asked but walked out with the black one [2][3]. Stated preference is unreliable, so the proposal is to test brand decisions a different way: spawn a thousand synthetic respondents segmented by age, gender and belief, and simulate the market response rather than interrogate it [2][3]. This is agents used precisely where the incumbent method is weakest, not as a cheaper version of the same research.

On tools becoming a commodity and what people are left holding

Once every team has access to the same models and the same tooling, the tools stop being where advantage lives, and the difference comes from what individuals bring out of their own strengths [2]. The employment consequence is not hypothetical or distant. In Belgium it is already visible that fewer developers are needed to do the same work [3], and the sectors expected to feel it first are accounting and Big Four-style consulting, the knowledge work most exposed to agentic automation [2][3].

On the 'Move 37' that has not happened yet

Against the more excitable claims made for agentic AI, the honest assessment is that it has not had its 'Move 37' moment [3]. The comparison is to AlphaGo at the stage where it was still learning by imitating grandmasters, which is roughly where agents sit now [3]. What makes the trajectory credible rather than speculative is the reasoning models, which show a path toward AI producing genuinely novel strategies that humans can then learn from [3]. The point is a sequencing one: imitation first, and originality is a later and separate milestone that should not be claimed early.

On giving a product its own name

When BOSA looked for a partner to renew the Mercurius platform, the winning combination went on to build what is described as the new digital post office for the Belgian public sector [1]. The reasoning about what to do next is a product-organisation argument rather than a technical one: a solution of that kind "verdient geen bijrol in onze productportfolio. Ze verdient haar eigen naam, haar eigen team, haar eigen roadmap" [1]. Hence Dokapi Public launched as a full product rather than a feature line, on the view that a name, a dedicated team and a roadmap of its own are what make something a product at all [1].

Takeaways

• Decompose agent work into narrow cooperating roles rather than one general system: a brand digital twin, a market agent and a design agent that combines them [3].
• Point agents at problems where the existing method is known to fail, such as stated-preference research, where people pick the orange speaker in the survey and take the black one home [2][3].
• Simulating a thousand persona respondents segmented by age, gender and belief is a different way of testing brand decisions, not a cheaper survey [2][3].
• Assume the tooling levels out: "Al die tools worden een commodity en dus de mensen gaan een verschil maken vanuit hun eigen sterkte" [2].
• Displacement is already measurable in Belgium, where fewer developers are needed for the same work, with accounting and Big Four consulting named as next in line [2][3].
• Agentic AI is still at the imitation stage, pre-'Move 37'; reasoning models are the evidence that genuinely novel strategies are coming, not proof that they have arrived [3].
• A solution that matters should not be a supporting act in the portfolio: give it its own name, its own team and its own roadmap [1].