Corporate digital transformation has hit a wall: gigabytes of digitized data are stubbornly refusing to evolve into operational intelligence. In a recent position paper, ETH Zurich researchers Anne Marks and Mennatalla El-Assady state that accumulated corporate knowledge remains fragmented across software, manuals, and implicit employee expertise. The problem is that these assets were created exclusively for human consumption. When companies attempt to grant AI agents decision-making power, they discover that this "legacy" is a toxic asset. According to the authors, the bottleneck for AI adoption has shifted from raw computing power to knowledge grounding. Without it, a system’s behavior in non-standard situations remains unpredictable due to the semantic vacuum of the underlying data.
The Mixed Agency Model
To bridge this gap, the ETH Zurich team proposes the concept of "mixed agency," moving away from a binary view of automation. The framework links three levels—user, mixed, and system—with a taxonomy of corporate knowledge. In this setup, an orchestrator (human or AI) dynamically allocates authority based on how deeply the algorithm is grounded in context. By decoupling task complexity from autonomy levels, the researchers demonstrate that even basic operations require deep contextual understanding; otherwise, the cost of error in the physical or economic world becomes prohibitive. The degree of machine freedom must fluctuate based on the real-time "Company World State."
Moving Beyond RAG and Document Dumps
The study emphasizes that simply "feeding" files to an AI via Retrieval-Augmented Generation (RAG) is a dead end for serious decision-making. Routine visual quality inspection is one thing; choosing a strategic site for a new factory is quite another. While routine tasks can drift toward full system autonomy, strategy remains a human domain. The reason is simple: key knowledge often stays tacit or locked in siloed systems. Even a digitized document loses context when fragmented, becoming "inedible" for an agent that lacks a holistic view of the organization.
To move from primitive chatbots to autonomous operating systems, ETH Zurich calls for a complete overhaul of knowledge storage architecture. Chaotic PDFs must be replaced by structured representations, such as Knowledge Graphs. This is the only way to ensure AI doesn't just predict the next token but actually relies on the operational reality of the business.
The current strategy of "piling on more data" is fundamentally flawed. As long as data remains a heap of documents rather than a semantic map, your agents are doomed to languish in low-risk pilot projects.
The primary challenge for tech leads today is not selecting the "most powerful" model, but redesigning the IT landscape to achieve true grounding. The goal is to find the tipping point where the cost of structuring data is finally justified by the efficiency of autonomous decisions.