Most agentic AI talks are either a demo reel or a warning. This one is neither. In two and a half hours we show you what an agent is under the hood, where in your organisation it pays off, and what the evidence actually says about productivity. No formulas, no hands-on, no pitch for a tool.
It is built for executives, managers and decision makers who need to make investment calls this year, and who would rather understand the mechanism than trust a vendor.
What we cover
Foundations. What a large language model actually is: a next token predictor with frozen weights and no memory. We walk through attention on one real sentence, in five steps, without a single formula. Everything after this follows from it.
From chatbots to agents. Where the loop comes from, and the three roles in the system: the model, the harness that gives it tools and a loop, and the human who stays in the saddle. You control the model through the agent, never directly.
Where agents can run. On a laptop, in the cloud, or embedded inside your own product. Each choice puts your data somewhere different, and that is a governance decision before it is a technical one.
Where the value actually lives. Developers are a small slice of most organisations. The same loop applies in legal, operations, sales, analysis, support, finance and HR, and that is where the volume of the gain sits.
What to realistically expect. Real cases from our own work, each with the catch named out loud, next to the measurements that disagree with the hype.
Risk, the EU AI Act and HR. Data egress, shadow AI, vendor concentration and audit trail. Risk classification and human oversight duties. What changes in hiring and training when everyone has an agent.
Why this one is different
We cite the measurements, including the inconvenient ones
Most AI training quotes the number that sells. We show the spread, because the spread is the actual finding.
- A randomised trial found experienced developers believed they were 20% faster while measuring 19% slower. A follow-up a few months later flipped the sign, and the researchers flagged their own result as unreliable. Both belong on the slide.
- A large Stanford dataset finds real gains on greenfield work and often negative effects in mature, complex codebases.
- A controlled study of consultants found strong gains inside the model’s competence and a 19 point rise in wrong answers outside it.
- A UK government trial measured 26 minutes saved per day. Unglamorous and real.
- Enterprise surveys keep finding that most pilots produce no measurable effect on the bottom line.
The punchline is not that AI does not work. It is that the gains are real at the task level, and the firm level gap is about workflow redesign, adoption discipline and governance rather than the model.
The cases come with their catch attached
We use two of our own, both outside engineering. An agent that prepares a manager for a one to one in thirty seconds instead of twenty minutes, and the fact that the resulting briefing is employee personal data, which puts it near the high risk line in the AI Act. An agent that drafts public tender submissions quickly and cheaply, and the customer feedback that cheap scale floods a channel with applicants who cannot explain why they are in the room.
Useful, and worth doing. Neither is free of consequences, and we say so.
We explain the mechanism, not the vibe
Security is in the news every week, so the room already knows the vocabulary. What most people have never been shown is why prompt injection cannot simply be patched out, or why a permission prompt gives you a false sense of control. That takes explanation rather than a warning.
Practical
| Length | Two and a half hours, no hands-on |
| Language | English or Dutch |
| For | Executives, managers and decision makers with some IT background |
| Prerequisites | None. Some experience with ChatGPT helps. No coding, no mathematics |
| Format | Talk with questions throughout |
| Delivered by | Frank Dekervel or Rein Lemmens |
| Where | Public editions in Brussels, or in house at your offices |
Over 100 people have been through our agentic AI trainings across the developer and executive editions.
Book a seat
Public editions run at Pepibru Business Factory in Brussels. Dates are on the events page, and a seat is one form away.
Want it in house for your management team instead? Book a slot and tell us the audience.