Context Graphs: A Concept for Proactive AI Agents in Enterprises

Modern corporate AI agents based on Retrieval-Augmented Generation (RAG) remain reactive: they act only in response to a human request. The authors of this paper propose the concept of proactive agents, which provide useful information on their own before an employee even asks for it.

Context Graph Architecture

At the heart of the solution lies the Context Graph—a dynamic relational data structure that models enterprise entities, their relationships, and state changes over time. Three components are built on top of the graph:

• Delta Detection Engine—continuously monitors state changes in the graph.
• Proactivity Scorer—ranks potential notifications by urgency, relevance, and alignment with the user’s role.
• Surfacing Layer — an LLM-based layer that delivers sorted notifications with reasoned explanations.

The authors formalize each component and derive a unified proactivity evaluation function (Proactivity Score). A full-scale implementation in Python was carried out using the NetworkX library and the Anthropic Claude API.

Performance Evaluation

The approach was tested across three scenarios: contract lifecycle management, responding to engineering incidents, and sales funnel monitoring. Results:

• Precision@5 was 0.83
• The false positive rate was 0.11
• The average time to provide information was reduced from 47 minutes (with a reactive approach) to less than 30 seconds

Source: cs.AI updates on arXiv.org