Systeum Knowledge Hub: Agentic AI in Practice

For the fourth Systeum Knowledge Hub we invited three guests who look at agentic AI from very different positions. Kateřina Lesch runs AI strategy at an international pharmacy group, Peter Krejzl builds agentic platforms for enterprise operations, and Sara Polak studies technological change as an archaeologist and anthropologist. Each of them talked about something different, and yet their perspectives kept meeting at the same point: what decides the outcome is not the models, but people, rules and accountability.

So many questions came in from the audience that we had to hold some of them over for the closing panel. Here is what was said at this year's Systeum Knowledge Hub.

Kateřina Lesch: The success of AI projects is decided by people

Kateřina Lesch is Group AI Manager at Dr. Max, a pharmacy group that operates in six markets and employs 22,000 people.

Asked what gives companies the most trouble when adopting AI, she answered without hesitation: people and change management. This transformation is harder than the ones before it, she says, because it touches employees' professional identity. It shows up as early as process mapping, where people often cannot describe what they actually do in their jobs, and some of them suspect that describing it is exactly how they could lose it. That is why Dr. Max frames automation from the outset as replacing individual steps in a process, not the person. The hardest group, meanwhile, tends to be middle management: they make the decisions but do not work with the tools themselves, and out of fear that they will not understand the subject, they can stop a project.

Pilot projects, according to Kateřina, usually do not end because the idea was bad, but because no one carries them any further and no one wants to take responsibility for them. It pays off to put the unpopular basics in place first: a clear owner, order and classification in the data, and a secure gateway to language models. She takes an equally pragmatic view of costs. An invoice is never just the price of tokens, since cloud costs and exchange-rate differences get added on, so she handles AI rules and limits in contracts and tenders and allocates licenses by business case rather than by token consumption. Automation, moreover, does not always have to mean AI, and in some cases human work is still cheaper.

And asked how work will change in ten to twenty years, she replied that she fears the loss of meaning more than the loss of jobs.

You can find the full discussion with Kateřina Lesch on our YouTube channel.

Peter Krejzl: A rule in a prompt is not a rule, but a wish

Peter Krejzl is Senior Director of AI at Bloomreach, where he is behind the development of a production agentic platform built on an architecture of several cooperating agents. His thesis: building an agent is easy today, and what is far harder is trusting it enough to put it in front of a customer. He illustrated that with the case of a software company serving car rental firms, where an agent working on a simple task found an access token on a developer's disk and deleted the entire production database. It took nine seconds, and the developer had explicitly forbidden that kind of step in the agent's rules. Hence Peter's main message: one sentence in a prompt usually works, but not every time.

Security therefore has to be handled in the layer beneath the agent, not in its instructions: minimal permissions, an identity of its own, an explicit list of allowed tools and domains, no access to production or to a company card, and for deletions or payments a human in the loop and the option to roll the change back. Everything needs to be logged so the situation can be replayed later, and monitoring must never run on the agent, but on the platform. Behind every agent there should also be a named person and a kill switch that a colleague outside the development team can use. Peter also pointed out that whatever an agent says on your website, your company says, and that it pays to bring a lawyer into the project right at the start. Shadow AI, in his view, is not worth suppressing but worth making easier, and he closed by noting that feeling faster with agents does not mean we actually are faster.

You can find Peter Krejzl's full talk on our YouTube channel.

Sara Polak: Corporate resilience is built on memory and decision-making

Sara Polak is an archaeologist, anthropologist and technology strategist. She works at the University of Oxford and at CIIRC, the Czech Institute of Informatics, Robotics and Cybernetics at the Czech Technical University in Prague, where she founded the interdisciplinary Liminal center.

She offered a view of AI from the distance of 200,000 years of human adaptation to civilizational change, starting from the premise that technology is a cultural product: the mere fact that it exists does not mean it is useful. The Romans knew the steam engine, but for them it remained a party trick. Sara knows this from practice too: the same campaign technology was used at one point by seven candidates for the French presidency, and every one of them used it completely differently. That, she argues, is exactly why leading AI companies hire philosophers, linguists and archaeologists, because language and cultural context shape how models behave.

Her second big theme was memory. The internet today consists largely of AI-generated content, a significant share of human knowledge sits on unmaintained servers, and infrastructure is concentrating in the hands of a few players. She therefore advises companies to think about where their data is, who owns it and who could switch it off. What that looks like in practice is shown by a project with the Lobkowicz family, where a digitized library catalogue runs in a chatbot on their own hardware. What slows AI adoption most, in her experience, are human things: envy between colleagues, and shadow AI, which makes people talk to each other less, so that no one quite knows what decisions in the company are based on. As a counterweight she points to Bletchley Park and to the head of British intelligence: look for curious, committed people regardless of their field, and give them rules as simple as possible, and no simpler.

You can find Sara Polak's full talk on our YouTube channel.

Panel discussion

The closing panel opened with a topic that is close to us at Systeum: what the arrival of AI means for junior roles. According to our guests, many companies have stopped hiring juniors and are waiting to see what happens. Kateřina described interviews where a candidate turns up with dozens of generated slides but cannot say in five points what their strategy actually is. Peter said the best juniors he has ever hired were the ones who took a sheet of paper to the assignment and started thinking with common sense, and he advises today's students to come with a portfolio of things they have built themselves. Sara added that when choosing people it pays to look for curiosity and decency, because a team of ten egocentrics will fall apart however good their skills are. The conversation then turned to worries, with Sara most concerned about the concentration of power in a handful of technology companies and Kateřina about the flashes of an agenda of their own that she occasionally sees in agents, and to where AI brings the most: research and development in medicine, with cutting red tape in public administration added to it. And on the question of whether companies are giving up ownership of their source code came the thought that tied the whole morning together: the real value is a company's context and know-how, while responsibility for the outcome stays with a human being.

You can find the full recording of the panel discussion on our YouTube channel.

Thank you, and see you next time

Thank you to Kateřina Lesch, Peter Krejzl and Sara Polak for how openly they spoke, and to everyone who came for questions that were genuinely interesting.