Anyone wishing to understand just how quickly AI is making its way into everyday practice should look to software development. There, the developments of recent years have been clearly evident: first intelligent autocomplete, then chat-based coding, and now agent-based collaboration. AI no longer just writes individual lines of code, but uses tools such as APIs or databases, modifies files, tests results and works iteratively towards a goal. This is precisely why software development is a good benchmark for the maturity of Agentic AI.

This development is also of strategic importance for Europe. Agentic AI is transforming the way work is organised in the digital sphere. However, the crucial question is not only whether Europe can keep up in the race to develop the largest LLMs. What matters most is whether we build the most powerful AI systems for the real economy: for processes, data, decision-making and accountability.

Many things already work remarkably well today, particularly when it comes to clearly defined tasks with seamless tool integration and verifiable success criteria. In such cases, Agentic systems are already delivering productivity gains. However, strategic tasks are not yet handled reliably. A sober assessment is needed here: language models can plan, formulate and often draw convincing conclusions. But they are not designed to reliably predict real-world business processes. Will an order arrive on time? Will a customer cancel? Will an invoice be paid? The answers to these questions cannot be derived from the texts on which the language models used by agents are based. Here, spreadsheets and database entries must be used to identify patterns from existing records for relevant and reliable predictions.

Large relational models, such as SAP RPT-1, therefore represent a significant step forward. They are trained on precisely this kind of structured data and thus differ fundamentally from language models. Whilst LLMs are optimised for linguistic continuity, relational models must identify relationships between columns, rows, time series and numerical patterns. They do not speak better – they predict operationally relevant outcomes with greater precision.

The real quantum leap occurs when these two worlds come together. Language models are powerful generalists: they understand queries, conduct dialogues, plan actions and coordinate tools. Relational models are specialists in making predictions based on structured data. Knowledge graphs add a third crucial layer to this architecture: they make meanings, relationships and responsibilities explicit, thereby helping agents to identify and use the right tools confidently and reliably. In agent-based systems, these components can therefore work together perfectly. The language model uses knowledge graphs to control the interaction, whilst the relational model provides the robust forecast. This creates a new user experience. A user can then ask not only: “Where is my order?”, but also: “When will it arrive, how reliable is that statement, and which alternative would be faster?” Such questions require not only language understanding, but predictive intelligence.

This is precisely where I see Europe’s chances. We should not define ourselves solely by catching up in the field of generic LLMs. Europe’s strength lies particularly in building large relational models for economically relevant data and process environments. Knowledge graphs are central to this, as they semantically link fragmented data sets and make relationships, responsibilities and rules explicit. In doing so, they provide the models not only with more input, but also with a more robust structure. This makes the models more context-aware, more consistent and better suited for real-world business processes. For businesses, this represents tangible added value: AI becomes not only more powerful, but also more transparent, controllable and reliable. This is precisely what is required for agent-based systems to be deployed safely and effectively in the business world.

To achieve this, however, we need to do more than simply build individual applications. We need shared foundational models for economically relevant data and process environments – not as a central data repository, but as a trustworthy infrastructure for collaborative learning. Europe should promote foundational models that learn generic capabilities for the economy from broad, high-quality datasets. Companies can then adapt these for their specific applications. A shared foundation, individual differentiation: that would be the European way.

The future of Agentic AI therefore does not lie in a single all-knowing model. It lies in an architecture of composite intelligence. This involves dialogue-capable generalists, predictive specialists and knowledge graphs. These enable agents to access data, tools and internal corporate knowledge in a precise, traceable and scalable manner. Those who master this architecture will build the most productive AI systems.

This is good news for Europe. Our strength lies not in copying the platform logic of the consumer internet, but in the combination of domain knowledge and process understanding. If we build on this, we can take a leading role in global competition. To this end, we should invest specifically in base models for economically relevant data and process environments, in knowledge graphs and in trustworthy infrastructure for collaborative learning. Europe’s opportunity lies not in AI that merely sounds good, but in AI that delivers reliably in the economy.

March 2026