Not every agent is an agent: What we learned from KI.Forum.NRW 2026

Recap from KI.Forum.NRW 2026: Why AI agents are understood very differently and what companies can derive from this for knowledge, governance, and learning.

By Martin Salwiczek · 2026-09-22

Shortly after the program began at the Confex in Cologne, half the room had their phones in hand. Moderator Alexa von Busse asked via Slido if those present already use AI agents. She immediately followed up: Copilot doesn't count. The result was close. "No, not yet" was in the lead, closely followed by "Yes, at least one agent in productive use."

That set the theme of the day before the first lecture began. Everyone is talking about agents. But they mean very different things, and they are at very different stages. The annual conference of KI.NRW was held under the heading "Entrepreneurship in the Cosmos of AI Agents". For us, that was precisely the exciting question: What insights into agentic AI will we get, from research to hardware retail?

An introduction that sticks

Before it was about technology, Michaela Benthaus spoke. The aerospace engineer from ESA has been paraplegic since a mountain bike accident and was the first wheelchair user to fly into space in December 2025. She spoke about parabolic flights where she tested whether she could buckle herself in without control of her legs in 20 seconds of weightlessness. She spoke about two weeks in a windowless habitat. And she explicitly described her mission as a team sport. It had nothing to do with AI. But a lot to do with the question of how to venture into unknown territory.

What exactly is an agent?

The range was enormous. At Carglass® Germany, according to Julia Piskurek, Senior Expert Innovation, the agent inventory extends to private lottery predictions. She finds that acceptable as long as it's an introduction to the topic.

At sipgate, on the other hand, so-called Agent Loops are used, explained co-founder Bastian Wilhelms. These are systems that constantly process new information and independently consider how they could surprise a team.

Sebastian Schoenen, Director Innovation & Technology at ControlExpert and Head of the Data & AI Center of Excellence at the Solvd Group, presented an entire agent factory for claims management. It processes cases independently and deliberately hands over personal injuries to case workers, for example.

Prof. Dr. Christian Hürter, CIO of DEUTZ AG, dryly calculated: If you have 15 agents writing code, you need five to check it.

It was striking how cautiously those who were more advanced handled terms. Wilhelms didn't want to talk about AI employees because the agents are integrated into work processes and don't replace a person. Kevin Jostmeyer-Zelles rejected the label "AI native" for his company. They are simply data-driven through and through.

Setting goals instead of work steps

Schoenen described his company's approach using the image of a rowing eight. The coxswain sets the direction and tempo but doesn't row himself. The agents are given goals and key performance indicators: costs, customer satisfaction, speed. The path to achieving them is left to them, within a clearly defined fence. Every decision is logged.

He was refreshingly sober about it. The models can handle increasingly complex expert tasks, but often only with a hit rate like a coin toss. If you demand more reliability, the task size shrinks significantly. His guiding principle: "Own the brain, rent the body." Models, computing power, and frameworks are interchangeable. Your own knowledge, data, and orchestration are not.

The unspectacular foundation: shared knowledge

At this point, an astonishing number of presentations converged. Wilhelms described a small example. His agents have long known that a supplier was renamed years ago. Then another team sends him a document, also created with AI, and it still contains the old name. Two teams, two knowledge bases.

Jostmeyer-Zelles sells special screws from the Würth Group environment with Get Special Fasteners. There is no catalog. It is crucial to immediately recognize which of the many companies in a large group is currently inquiring. He himself called these 'totally boring use cases' that many still fail at. The biggest open issue in his company is the common context of all information.

Schoenen drew a conclusion from this that many might not like: documenting and writing meeting notes are becoming more important, not less important.

Governance with a cool head

Dr. Maximilian Poretschkin, Head of the AI Assurance and Assessments Department at Fraunhofer IAIS, advocated for a systematic approach before tackling the AI Act. Which systems do I have? Where do I stand in the supply chain? Technical standards help to sharpen vague terms like "bias". Even the question of where AI begins in the sense of the law is open in borderline cases. New risks are added with agents. An agent can, for example, pretend to follow a rule and not do so. Or someone manipulates its memory.

Hürter added to the practice at DEUTZ. Production processes remain deterministic. Open models, including Chinese ones, run at DEUTZ only in isolated environments within their own data center. And as CIO, he made a remarkable statement: Technology alone will not save it. The top management and a trusting relationship with co-determination are crucial.

Carglass: Being able to read does not mean wanting to read

The presentation that occupied us the most came from Julia Piskurek. She compared AI competence to reading and writing. Basic training teaches the ABCs, but it doesn't spark enthusiasm. What's missing is her 'Harry Potter moment': the point at which someone realizes what it's all worth it for.

Carglass works with a step ladder from initial familiarization to the redesign of entire processes. Not everyone has to reach the top. But everyone should know where they stand and what their next meaningful step would be.

The training is tailored per team. In sales, it was about preparing the next customer meeting with a good prompt. The legal team, on the other hand, designed its own agent in design thinking workshops. Piskurek openly admitted that this tailoring is time-consuming, but still recommended it. In addition, there are AI Ambassadors in the departments.

Her most important point: Knowledge from training evaporates if it doesn't get space in everyday work. That's why there are teams with a fixed 'Friday with AI'. You can mandate basic training, but not motivation.

Was das für uns bedeutet

The range in the hall is the range we experience in companies. That's why when someone says 'agent', we first ask what they mean. This is not pedantry. Someone talking about lottery predictions and someone talking about an agent factory needs completely different support.

Piskurek's presentation confirmed for us how we understand human-centrically: the individual person is seen and enabled. This means starting from the team's current level, finding personal benefit, and freeing up time in everyday work. This is more complex than a standard course, that's true. But it's the difference between being able to and doing.

For many smaller businesses in the Ruhr area, the realistic first step is probably not an agent factory. What several speakers described as the foundation is more important: clean data and a common knowledge base.

Outlook

Asked about 2030, Prof. Michael Riesener, Managing Director of RWTH Innovation GmbH, opined that no one will program individual agents themselves anymore. Capabilities will be drawn from a global modular system. If that's true, what remains as a competitive advantage is what cannot be rented: knowing your own processes and having people who can assess results. That can be learned. Learn today what matters tomorrow.