Qurated: Context Graphs for Proactive Enterprise Agents
The Oracle Problem: Why Your AI Agent Waits to Be Asked
Most enterprise AI is a brilliant employee who only speaks when spoken to. It answers perfectly, then goes silent — even while watching a contract clause expire, a service ticket spiral, or a sales deal go cold. This paper's core insight: proactivity isn't a feature bolt-on, it's an architectural choice, and it hinges on giving your agent a memory of relationships, not just a memory of facts.
The Reactive Trap
RAG systems are libraries with excellent card catalogs. You ask, they fetch. But a library never taps you on the shoulder to say "that book you checked out is now overdue and your professor just changed the syllabus." Reactive agents inherit this blindness — they have no concept of state changing over time, only content sitting still.
Mental model: Reactive AI = a dictionary lookup. Proactive AI = a nervous system, sensing change and routing signal to attention before you feel the pain.
The Context Graph: Enterprise as Living Organism
Instead of documents, model your enterprise as a graph: entities (contracts, tickets, deals) as nodes, relationships as edges, and — critically — state transitions as first-class citizens. A contract isn't a static PDF; it's a node with a trajectory: drafted → negotiated → signed → approaching renewal. The graph doesn't just know what things are, it knows what they're becoming.
This reframes AI's job: not "answer the question" but "watch the delta."
Four Components, One Pipeline
-
Delta Detection Engine — continuously diffs graph state, like a seismograph for organizational tremors. A missed check-in, a stalled deal stage, an unassigned incident: each is a detected delta, not a query result.
-
Proactivity Scorer — the triage nurse. Not every delta deserves an interruption. The paper formalizes a unified score combining urgency (how fast this decays), relevance (does it touch this person's actual work), and persona-fit (is this the right messenger for this news). This is the antidote to notification fatigue — precision over volume.
-
Surfacing Layer — an LLM that doesn't just alert, it explains, grounding each nudge in the graph's actual relationships so the human trusts the interruption instead of dismissing it.
-
The full loop runs continuously, not on-demand — proactivity requires always-on sensing, not query-triggered computation.
The Numbers That Matter
- Precision@5: 0.83 — when the system surfaces five things, four-plus are worth your attention. Trust is earned through hit rate, not volume.
- False positive rate: 0.11 — low enough that "cry wolf" fatigue doesn't set in.
- Mean time to surface: 47 minutes → under 30 seconds. This is the number that changes org design. A 94x speedup in signal-to-attention isn't incremental — it's the difference between "we caught it" and "we found out in the postmortem."
The Actionable Shift
Stop asking "what should our agent answer?" Start asking "what relationships and state transitions matter enough to model explicitly?" Build the graph before you build the chatbot. The graph is the nervous system; the LLM is just the voice reading out what the nerves already felt.
If your AI strategy is entirely retrieval-shaped, you've built a librarian. Enterprises need a lookout.