Researchers have published a significant critique of current enterprise AI architectures, arguing in arXiv paper 2607.07721 that widely-adopted Retrieval-Augmented Generation (RAG) systems and agentic frameworks suffer from a fundamental limitation: they operate reactively, only initiating action after receiving explicit human queries. The paper proposes addressing this constraint through context graphs—structured data representations that maintain dynamic maps of enterprise state across multiple dimensions. Rather than abstract representations, the system would encode specific operational relationships: for instance, a context graph might track the interconnected statuses of supplier relationships, inventory levels, regulatory compliance requirements, and personnel availability as a unified knowledge structure. The authors argue this architecture enables agents to continuously monitor enterprise conditions and identify problems before they reach critical thresholds, such as detecting supply chain vulnerabilities before procurement teams file formal requests for intervention.
The practical implications are substantial for enterprise operations. Current RAG-based systems excel at answering targeted questions—a user asks about quarterly sales figures, the system retrieves and synthesizes relevant data—but cannot independently flag emerging issues like cascading supply disruptions or anomalous spending patterns. Context graphs would theoretically allow agents to evaluate these interconnected conditions continuously, triggering alerts or preparatory actions without waiting for human prompting. However, significant technical and organizational challenges remain unaddressed. The paper does not thoroughly examine the computational overhead of maintaining real-time context graphs across complex enterprises, nor does it tackle the governance question of when autonomous agent action is appropriate versus when human approval remains necessary. Existing enterprise systems like SAP's analytics platforms and Salesforce's Einstein already attempt some predictive monitoring, but the paper argues they lack the flexible, semantic reasoning capacity that LLM-based agents could provide.
The core limitation the authors identify is the gap between current enterprise needs and existing AI capabilities. While the context graph framework presents a conceptual advance, the paper demonstrates the approach only in theoretical scenarios rather than against production enterprise datasets. The researchers acknowledge that building comprehensive context graphs requires standardizing how different business systems represent operational state—a significant practical hurdle in organizations running legacy infrastructure. Additionally, the paper notes that autonomous enterprise agents making decisions based on inferred context could amplify errors systematically across operations if context graphs become corrupted or if reasoning chains produce confident but incorrect conclusions. The work represents genuine progress in moving beyond purely reactive AI, but implementation would require solving substantial engineering and trust-establishment problems before enterprises adopt such systems for mission-critical operations.