«Simply dumping 20 years of documents into a vector database is not a strategy»
From the business department up to the CIO, everyone is currently feeling the same pressure: we have to do something with AI now. Proof-of-concepts are being launched, pilot projects announced – yet key questions remain open and uncertainties persist: can we trust the AI? How do we explain its decisions? And who really keeps control in the end?
This is exactly where our webinar comes in: together with Prof. Dr. Ute Schmid from the University of Bamberg and our CEO Thomas Gomell, we look behind the buzzwords and discuss how to use LLMs and AI in such a way that explainability, trust and controllability are not lost along the way
Watch the recording now:
What the webinar video is about
Artificial intelligence in a reality check: why data curation matters more than the next model
AI is everywhere: in presentations, in strategy papers – and, it seems, in every second piece of software. But while the hype grows, many fundamental questions remain open:
- Can I trust the answers of my AI?
- How do I keep my employees professionally competent when “the machine” takes over more and more?
- And: how do I prepare corporate knowledge so that AI can deliver meaningful results at all?
That is exactly what the expert webinar was about – with a clear focus on knowledge management, data structuring and knowledge graphs as the basis for explainable AI.
1. Between hype and reality: where AI really stands today
At the beginning the question was on the table: is the current AI hype justified – or are we sitting on a bubble?
Prof. Dr. Ute Schmid, who has been researching artificial intelligence since the 1990s, put it into perspective: the bubble will not burst completely, “but in some corners it will”.While the “battle of the large models” and generative AI currently dominate the headlines, the actual innovation is happening more quietly – for example with agentic models and in the area of Retrieval-Augmented Generation (RAG).
What matters here is less how large a model is, but rather how well the knowledge behind it is structured.
2. The real problem: information overload, data junk and hallucinations
A recurring theme of the webinar: trust in information.
Hallucinations are often a symptom – not the cause
Hallucinations – that is, false or entirely invented content – are not an exotic fringe issue but everyday reality in many enterprise scenarios. What is notable: frequently they are not (only) a model problem but a context problem.
If the input is vague, contradictory or imprecise, even the best model delivers poor answers. Or, as it was put pointedly:
„If you ask bad questions, you get bad answers.“
Big data without curation: more does not mean better
From the big data era, one hope still lingers: We simply store everything – AI will sort it out. That is precisely what causes massive friction today:
- Over decades, companies archive almost everything instead of deciding what really matters.
- For humans and AI alike, this blurs the line between relevant knowledge and „data noise“.
- The AI has to fish the few high-quality pieces of information out of an unstructured, overfilled data pool – a game played with a bad hand.
The result: models struggle to deliver reliable answers because the input is neither curated nor structured.
„More content“ is not a strategy
Around 40 % of companies already use AI actively, primarily for marketing, content creation and chatbots. But if AI is only used to produce even more content even faster, we continue to „clutter“ our digital landscape – without solving the underlying structural problems.
3. Humans and AI as a team: human-in-the-loop instead of autopilot
A central theme of the webinar was the question: How do we prevent employees from becoming „rubber stampers of AI output“?
Skill skipping: when competencies erode unnoticed
In large companies it can already be observed that after intensive use of ChatGPT & Co., employees hardly feel able to write a text „from scratch“ any more. Instead, they only prompt, correct and approve. This carries risks:
- Professional and linguistic competencies are gradually eroded.
- Employees increasingly rely on systems that they neither control nor really understand.
- In the extreme case, a „WALL-E scenario“ emerges: people hand over responsibility and reduce themselves to commentators on AI results.
High performers use AI as a sparring partner
But the webinar also made clear: this is not about demonizing AI – on the contrary. Anyone who uses AI as a sparring partner benefits enormously:
- Better ideas through rapid variation and iteration
- Higher quality through structured cross-checking
- More output with the same or even greater depth of content
The clear forecast: Employees and companies that learn to use AI reflectively and to deliberately develop their own competencies will overtake those who close themselves off from the technology – und those who surrender to it uncritically.
4. Knowledge graphs & neuro-symbolic AI: from a heap of data to enterprise knowledge
The most interesting question in the second part of the webinar was: how do we create the basis for AI in companies to become reliable, explainable and controllable?The answer: knowledge graphs und neuro-symbolic AI.
„Simply dumping 20 years of documents into a vector database is not a strategy.“
What is needed instead is deliberate curation of data:
- Which documents are professionally relevant and up to date?
- Where do the binding truths lie (policies, contracts, architecture baselines, approvals)?
- Which information may enter AI contexts at all (compliance, data protection)?
The recommended approach: Cherry Picking – Only selected, high-quality information is transferred into the knowledge graph. This lowers costs, reduces environmental impact (less computing effort) and – above all – raises the quality of the answers.
What makes a knowledge graph so powerful?
A knowledge graph breaks unstructured information – e.g. documents, emails or tickets – down into entities (objects, people, systems, processes) and relationships (who is connected with whom or what, and how). Instead of merely storing files, a network emerges consisting of:
- technical terms
- systems and interfaces
- responsibilities
- rules, guidelines and exceptions
This linked information creates a „connected energy“:The AI no longer accesses disconnected text fragments but an enterprise model with context and structure.
Neuro-symbolic AI: the best of two worlds
Knowledge graphs are not a new idea – they come from a time when AI was still strongly knowledge-based. Today they are combined with modern language models:
- Neural components (LLMs) provide language understanding, generation and interaction.
- Symbolic components (knowledge graphs) supply structure, rules, relationships and verifiable facts.
This neuro-symbolic approach makes AI:
- more explainable (answers can be traced back to concrete nodes and edges in the graph),
- more controllable (rules and governance can be represented),
- more sustainable (knowledge survives generational changes of tools and models).
The webinar showed by way of example how a knowledge graph framework such as aikux.Brain or the underlying graph technology can serve as the foundation for solutions such as migRaven.MAX – intelligently connecting enterprise knowledge, IT data and AI. (Migraven)
5. Technological sovereignty: why Europe needs its own answers
The closing part addressed a strategic perspective: Do we want to make our corporate intelligence permanently dependent on a few US or Chinese providers? The risks are obvious:
- Strategic dependence on individual platforms and clouds
- Lack of clarity about data flows and usage rights
- A difficult starting position for regulation, auditability and compliance
At the same time: if employees reduce their own thinking and outsource more and more decisions to (non-explainable) systems, an „extreme danger for the company“ arises – both professionally and organizationally.
The appeal of the webinar:Europe must investieren – in its own platforms, in explainable AI, in data and knowledge architectures that are not only efficient but also sovereign and auditable.
Conclusion: no responsible AI without data curation
The webinar made it clear:
- The problem is not AI – it is the way we handle knowledge and data.
- Anyone who does not curate their information landscape receives unreliable answers that are hard to explain.
- Anyone who does not enable employees to work with AI in a considered way risks losing important skills.
- Anyone who structures their corporate knowledge – for example in a Knowledge Graph – creates the basis for explainable, secure and sovereign AI applications.
In concrete terms, this means for companies:
- Clean up and curate data instead of „simply storing everything“.
- Establish human-in-the-loop structures, so that expertise and AI reinforce each other.
- Evaluate Knowledge Graph approaches, in order to make corporate knowledge modellable, linkable and auditable in the long term.
- Secure your own sovereignty – technically, organizationally and strategically.
