bbv KI Webinar - AI Agents in Practice

I spoke in a webinar together with my colleague Stefan Heberling on a topic that has fascinated me for a long time: generative AI and the so-called agentic way of working. Our goal was to show participants how these technologies can be deployed concretely in organisations and where the advantages over purely passive chatbots like ChatGPT lie. Stefan, head of the AI team at bbv, emphasised at the start of our meeting that we wanted to focus on the interplay between AI agents and operational processes. He invited me as a discussion partner because I myself work closely on such use cases in several AI projects.

"We are giving you an exclusive look into our BBV AI Hub today, starting from the challenges we encounter daily with our clients."

This quote from Stefan outlines the core of our event. Our aim was to show what generative AI can practically achieve in an enterprise context. Participants wanted to know how to introduce agentic systems in a way that makes them not merely a nice add-on but a genuine relief in day-to-day operations.

AI Agents as Autonomous Problem Solvers

In the webinar I first explained the term "AI agent". It refers to virtual units capable of handling their tasks independently, much like human employees. This means they do not just answer a question but receive a defined goal and independently work out how best to achieve it. Stefan put it this way in the conversation:

"Reacting autonomously to inputs, following a specific workflow and delivering the desired results on the basis of that workflow."

Behind this lies a sequence of steps similar to the human problem-solving process. The agent recognises a situation, interprets it and in the next step determines which actions make sense. It then carries out those actions and checks whether they lead to the desired result. If gaps or errors appear, it adjusts its approach. I emphasised that this goes far beyond simple chatbot logic. AI agents work according to an agentic workflow whose core elements are perception, planning, action and feedback loops.

To make this more tangible, I described an example of inventory management at a logistics company. An AI agent continuously checks how much material is still available, forecasts requirements based on historical data and, if necessary, independently initiates a new order. It even decides when which quantities make sense and automatically notifies the responsible department.

"An AI agent could monitor stock levels in real time, automatically trigger actions when levels are low and thereby save time and resources in the long run."

This is intended to illustrate that these programmes are capable of considerably more than pure chat functions.

The BBV AI Hub: Technology Platform and Opportunities

In a second step I presented the BBV AI Hub. This is a technology platform we developed to programme new AI agents as quickly as possible and embed them in existing IT landscapes. Anyone asking why such a thing is needed: I explained that classic software architectures are not always designed for the speed at which AI applications evolve.

The AI Hub brings clarity to the management of data and connections to various language models. This means one can opt for commercial services such as OpenAI or Google Cloud, or run open-source models in one's own data centre. For some organisations the latter is indispensable because sensitive data must not leave the company network. A further advantage lies in scalability. When an agent has to process requests very frequently, we can quickly determine whether a more powerful environment or a different model would be more appropriate.

Challenges in Client Projects

I am often asked where the biggest hurdles lie in developing these AI solutions in day-to-day practice. My answer is always that we as an AI team have the purely technical aspects well under control, but the factual correctness of results must be verified with the support of our clients. In the inventory management example it is clear: we do not know every detail of every sector. This means that an AI agent may work formally correctly but there is always a possibility that a sector-specific rule gets overlooked.

In the webinar we showed how we address this problem. We log the individual steps of our agents. As soon as an incorrect answer appears, one can trace in an agent logbook which texts or data sources were evaluated and whether the system perhaps used outdated information. The nice thing about this is that one learns a great deal from these analyses and continuously improves the agents based on client feedback.

"Particularly with critical decisions we want to ensure that the AI agent tells us which document it is relying on."

That is how I explained it in the webinar. When an automated action has far-reaching consequences, transparency is essential.

Collaboration Between Agents A topic that interested participants greatly afterwards was the question of whether multiple AI agents can communicate with one another. I confirmed that we have already developed solutions for this.

It is possible for a superordinate coordinator agent to recognise that a task requires several specialist agents. It then forwards requests to the relevant agents and collects their partial results until an overall solution is available. This principle follows the same logic as in real teams, where different experts work together.

In the webinar I showed, for example, how an agent responsible for monitoring inventory levels can cooperate with another agent that initiates orders or tenders. The result is an AI agent team capable of making independent decisions without a human having to control every intermediate step.

Roles, Rights and Data Access

In many client conversations the question of access control comes up quickly. Employees should only see information that they are permitted to access according to their role. In such cases I emphasise that we always define agents quite clearly. An agent receives exactly the data sources it needs for its task, no more and no less.

The AI Hub also provides connections to existing access and rights systems. Anyone not permitted to use the agent does not see it and therefore cannot read critical data. I pointed out that this is a key factor for AI systems to integrate securely into operational life in the long run.

Finding Knowledge Rather Than Searching for It

Many organisations have extensive document or knowledge repositories. According to studies, employees waste valuable time every day because they do not know exactly where a piece of information is stored. In the webinar Stefan mentioned a statistic showing that employees sometimes spend 30 minutes a day just searching for documents. I explained how AI agents in combination with so-called vector-based databases can provide a remedy.

When a question is asked, the agent searches the relevant sources and very quickly finds those text passages that are suited to answering it. These are far more complex search strategies than a simple full-text search. By evaluating context, the semantic proximity between terms, the agent finds the right passages even with synonyms or related topics. What is decisive is that we continuously improve this system and regularly check whether the agent is delivering correct results.

Live Demo: Transparency and Traceability

During the webinar I showed in a live demo how an AI agent works in a question-and-answer scenario. Participants could see that the agent goes through several intermediate steps with each question. It checks whether it is responsible for the question at hand, searches for relevant text passages, filters them and only generates an answer at the end.

We also demonstrated that all sub-steps can be viewed in the agent logbook, for example which documents or passages the agent used as a source for its answer. This transparency gives users confidence, because they can trace when needed why an answer was formulated in a particular way.

Outlook and Own Use Cases

At the end of the webinar Stefan asked me which AI agents we use internally ourselves. I told him about a so-called wiki agent we developed that can access our company-specific guidelines and our intranet. Employees looking for specific HR regulations or IT policies no longer have to manually search multiple platforms.

Beyond that, we support teams with the creation of project descriptions by having an agent automatically collect the relevant key data and insert it sensibly into a document. In addition, we have an agent that bundles internal AI knowledge. It answers colleagues' questions before those questions reach us as AI developers.

"We see in practice that many questions concentrate on a small core of company knowledge. That is exactly where it is particularly worthwhile to create a cross-departmental, central point of contact with AI agents."

I assured participants that we will continue to expand this principle in future, for example with agents that automate ordering processes or ease the creation of documentation entries in complex projects.

Stefan concluded by pointing to the Swiss AI Impact Report, which we would be publishing shortly. In it we evaluate a study we conducted on AI use cases in Swiss organisations. We want to share our findings in a future webinar and show which experiences Swiss companies have already had with generative AI.

Conclusion

I have the impression that the evening was an eye-opener for many participants. We were able to illustrate clearly that generative AI systems are already far more than simple chatbots. AI agents can address concrete operational challenges that reach deep into everyday business. Above all, I emphasised the importance of transparency. Only when it is clear how an agent arrives at its recommendations are organisations willing to hand over responsibility to such autonomous systems.

For me personally it remains clear that AI agents are only as smart as the knowledge made accessible to them. In collaboration with clients, however, we experience how quickly the full potential unfolds when the right information is provided and the right role and security concepts are defined. It is precisely this combination of technical innovation and thoughtful integration into existing operational processes that fascinates me.