Why Context Graphs Could Unlock the Next Trillion‑Dollar AI Opportunity

The piece introduces the concept of context graphs —a real‑time, searchable representation of the "why" behind AI agent actions—as a potential trillion‑dollar opportunity in the AI era.

Three concurrent shifts

ChatGPT has generated a strong enterprise demand for AI that truly understands a company’s own business, not just public web data.

The Model Context Protocol (MCP) standardizes how agents expose and consume context, enabling any agent (e.g., Cursor, Claude, custom agents) to share the same contextual view.

While many firms are experimenting with agents, they lack a persistent context layer, causing agents to hit governance walls because they cannot access operational or decision context.

Therefore, building a dedicated context infrastructure becomes essential.

Decision traces, not raw data, drive the next AI wave

Current systems such as CRM or ERP store only the current state, omitting the rationale—"why" a decision was made. Autonomous agents need access to historical decision traces (exceptions, approvals, precedents) that currently reside in Slack, email, or human memory.

What a context graph is

A context graph captures a complete snapshot of each execution: input → rule → exception → approval → output. Over time, this cross‑system, time‑stamped graph becomes a valuable asset, serving as the enterprise’s new "truth source".

Why existing giants can’t fill the gap

Operational leaders like Salesforce store only outcomes, not processes.

Data‑warehouse providers (e.g., Snowflake) read data after the fact and cannot record decision‑time traces, missing the commit point needed for reconstruction.

Startup pathways to build the layer

Re‑engineer the entire record system as an AI‑native CRM/ERP.

Target a high‑frequency exception module (e.g., reconciliation, deal desk) as a decision system and sync it back.

Create a cross‑system orchestration layer that persists decision traces from day one, eventually becoming a full decision‑record system.

Rules vs. decision traces

Rule : tells an agent what should normally happen (e.g., "report using official ARR").

Decision trace : records the specific context of an exception (e.g., "under policy v3.2 we used definition X, approved by VP, based on precedent Z").

Context graphs as the persistent layer

When a startup builds an orchestration layer that logs every decision event, it creates a searchable, auditable record of why actions occurred. This enables debugging, compliance, and the transformation of one‑off exceptions into reusable precedents.

Concrete example

An renewal agent proposes a 20% discount, but policy caps discounts at 10% unless an exception is approved. The agent pulls three SEV‑1 incidents from PagerDuty, an open "unless‑fixed‑then‑cancel" ticket from Zendesk, and a prior renewal thread approved by a VP. It routes the exception to finance, receives approval, and the CRM records the fact: "20% discount approved."

Feedback loop

Each automated decision adds a new trace to the graph, continuously expanding a searchable precedent library. Initially, humans remain in the loop, but as the graph grows, more paths can be fully automated.

https://foundationcapital.com/context-graphs-ais-trillion-dollar-opportunity/