Nearly every enterprise agent we use today shares a structural flaw: they wait to be queried. RAG, tool calling, ReAct, and all current patterns start from a human prompt. The paper introducing Context Graphs flips this script: what if AI doesn’t answer, but anticipates the question? And, crucially, how do you do that without drowning workers in useless notifications?

The authors’ answer is a three-stage engine built on a live, relational data structure — the Context Graph. Instead of vectorizing documents and matching on demand, the graph models enterprise entities — contracts, incident tickets, sales opportunities — together with their relationships and, critically, state transitions over time. A Delta Detection Engine continuously tracks changes; a Proactivity Scorer ranks urgency, relevance, and persona-fit; finally, an LLM packages and delivers notifications with grounded explanations tied to the data.

The results are not just theoretical. Tests on three generic case studies — contract lifecycle management, engineering incident response, and sales pipeline hygiene — show Precision@5 of 0.83, a false positive rate of 0.11, and mean time to surface cut from 47 minutes (reactive baseline) to under 30 seconds. All implemented with NetworkX and Anthropic Claude APIs.

That’s where the tension arises for anyone evaluating on-premise deployment. The reference implementation uses an external LLM via cloud API, Claude. But the Context Graph, by definition, ingests the living state of the enterprise: sales pipelines, contracts under negotiation, security incident details. Handing those data to a cloud endpoint means accepting a sovereignty risk that many regulated industries — finance, healthcare, defense — cannot tolerate. This is not a technical footnote; it’s the ridge that separates a proof-of-concept from an adoptable product.

The good news is that the architecture is modular: the LLM engine is swappable. Replacing Claude with a self-hosted model — a quantized Llama 3, a domain-tuned Mistral — is a near-natural step. And it is precisely here that the paper, although silent on the matter, sends a second-order signal: proactivity shifts computation from “I answer when called” to “I monitor and reason in the background,” multiplying inference volume. Processing a handful of queries a day is one thing; evaluating hundreds or thousands of state events per hour, generating scores, and calling the LLM to produce an explanation when needed is another.

This mechanism reshapes infrastructure incentives. Cloud LLM services that bill per token lose, because the inference multiplier pushes organizations to internalize the model to contain TCO and protect data. Winners are on-prem inference hardware suppliers and orchestration platforms that manage continuous evaluation pipelines, not just reactive chatbots. Enterprise software itself may pivot: classic ERPs and CRMs that today emit static alerts could embed Context Graph–driven proactivity engines, making the passive dashboard obsolete.

A third-order effect concerns data quality. For a proactivity scorer to work with low false positives, you need a well-modeled, near-real-time graph — not something solvable with a nightly dump. Companies are forced to invest in robust integrations and data engineering, or face noisy notifications that destroy user trust. In this sense, the framework doesn’t promise easy automation but a pact: proactivity in exchange for data discipline.

In the short term, the proposal remains confined to teams willing to experiment with NetworkX and build a custom scoring service. But the direction is clear: the digital assistant that waits in silence is becoming an antiquated idea. And anyone who truly wants an agent that speaks first will inevitably have to bring inference in-house.