It gets difficult when knowledge is allowed to act

That is why TYPO3 was never just a CMS to me. It was a production environment for information, with its own permissions, roles, workflows, and human detours. You quickly learn that systems rarely fail because of code alone. They fail because of assumptions, unclear responsibilities, side channels, and knowledge that supposedly everyone has but nobody has documented properly.

What used to sound like dry systems work suddenly becomes the core problem with agentic AI, because AI does not stop at writing. As long as AI suggests text, polishes emails, sorts ideas, or produces summaries, much of it can still be treated as a productivity question. Then the discussion is about quality, corrections, hallucinations, and control.

That matters, but it is not the tipping point yet.

The tipping point begins when agentic systems stop merely preparing tasks and start intervening in workflows. They answer requests, shift priorities, trigger bookings, pass information along, initiate workflows, and act on behalf of an organization. At that point, a suggestion becomes a commitment, a draft becomes a decision, and context becomes action.

That brings us back to the old questions, only faster. It has to be clear which source is authoritative, which role is allowed to act, which context is still current, when a statement becomes binding, which decisions require approval, when a case must be escalated, and who is responsible when things go wrong.

These are not side issues for admins, lawyers, or process people. They are the questions that determine whether AI becomes productive or merely accelerates uncertainty.

Better models alone will not solve this problem. Systems will of course become more capable, context windows will grow, answers will become more precise, and tools more impressive. But the real gap lies somewhere else. It lies in the connection between the model, organizational context, permissions, roles, and responsibility.

Organizations do not consist of knowledge alone. They consist of rules, responsibilities, official maps, and unofficial city maps. They include people who know things, people who are allowed to do things, and people who ultimately have to answer for the consequences. An AI system that is supposed to act in such an environment cannot simply be given more documents and then told to “just do it.” Between organizational knowledge and binding action, there needs to be a layer that determines what is allowed to happen, and under which conditions.

That is not a nice-to-have feature.

And this is where my winding path comes full circle. I experienced information as raw material long before we called it “content.” I know production under real-world conditions and systems where permissions and workflows were not just words on slides. In open source, I saw how much depends on informal structures, and in politics and governance I learned that official maps rarely show the whole city.

I have founded, built, rebuilt, lost, reorganized, and learned that limits are not a weakness as long as you take them seriously. That does not produce a smooth resume. Good. A smooth resume would probably have left me watching the next tool demo. This path leads to a different question: how do you turn organizational knowledge into action that is controllable, traceable, and accountable?

This is where looking back ends and the present begins. I have little interest in the next grand debate about whether AI changes everything. Of course it changes a lot. What is more interesting is where old problems reappear because the new speed makes them more visible.

The difficult part begins where output starts to have consequences: when information becomes a commitment, context becomes a decision, knowledge becomes action, and afterward it still has to be possible to explain why something was allowed to happen at all. That is where the thread running through this series becomes a direction for my work. Not as a neat founder story, but as a concrete field: how can organizations use AI without surrendering accountability to machine rhetoric, blind faith in tools, or informal instructions shouted from the sidelines?

That is less spectacular than most AI promises, but much closer to the actual problem. In the end, the decisive question will not be who used the best text generator. It will be who still knows what is authoritative, who is allowed to act, which basis applies, and why a decision was made. At that boundary, my winding path becomes useful, as a toolbox, not as a nice story.

Anyone who sees the same problem from technology, operations, sales, or capital will probably recognize that this is not a topic for someday. This is where a common thread turns into work.