Context Graphs: AI’s Trillion-Dollar Opportunity — Jaya Gupta & Ashu Garg

Summary

The next trillion-dollar platforms won’t be built by adding AI to existing systems of record — they’ll be built as systems of record for decisions, not objects. The core argument: the reasoning connecting data to action was never treated as data in the first place. CRMs track deals, ERPs track inventory, but nobody captures the decision traces — the exception logic in people’s heads, the precedent from past decisions, the cross-system synthesis, the approval chains that live outside any system.

The framework breaks down cleanly:

1. Rules vs. Decision Traces. Rules tell agents what should happen. Decision traces capture what did happen — the actual reasoning path a person or team followed, including exceptions, overrides, and contextual judgment. Rules are static; decision traces are living data.

2. Context Graphs. A context graph is the living record of decision traces stitched across entities and time. It captures not just the decision itself but the inputs, the alternatives considered, the people involved, and the downstream effects. Precedent becomes searchable. Institutional knowledge stops being trapped in people’s heads.

3. Why Agents Have Structural Advantage. Systems of agents sit in the execution path and see full context — they observe the reasoning as it happens, not after the fact. They don’t need to reconstruct decision traces from artifacts; they are the execution surface. This is why agent-native companies will build context graphs naturally while bolt-on AI never will.

4. The Platform Opportunity. Whoever owns the context graph owns the compound intelligence of the organization. Every decision makes the next one better. This is the moat — not the model, not the data warehouse, but the accumulated reasoning layer.

Another context graph article — an early version of the personal knowledge bases.

Connections

• 06-reference/2026-04-04-building-the-event-clock — Kirk Marple’s implementation response to this thesis. If context graphs are the what, event clocks are the how — the temporal infrastructure that makes decision traces queryable across time.
• 06-reference/2026-04-04-ontology-taxonomy-knowledge-graphs — Context graphs need ontological structure to be useful. The taxonomy defines the entity types; the context graph adds the reasoning edges between them.
• 01-projects/phdata/index — Context graphs are the consulting wedge for enterprise AI. Every phData client has systems of record for objects but nothing for the decisions connecting those objects. “Why did we approve this vendor?” lives in email threads, not the ERP.
• 01-projects/data-marketplace/index — Decision trace data as a product. Context graphs are dramatically more valuable than raw state exports because they carry the why. A dataset of “what customers churned” is commodity; a dataset of “why customers churned and what we tried” is gold.
• 06-reference/2026-03-31-block-hierarchy-to-intelligence — Block’s world model is a context graph. The hierarchy of entities + relationships + temporal state changes is exactly the structure Gupta describes — a living record of decisions across the financial ecosystem.
• 06-reference/concepts/compounding-knowledge — Decision traces compound in a way raw data never does. Each decision adds precedent. Each precedent improves the next decision. This is the knowledge flywheel that makes context graphs a moat rather than a feature.
• 06-reference/2026-04-04-karpathy-llm-wiki-idea-file — A personal knowledge base is a micro context graph. Karpathy’s idea file captures decision traces at the individual level — what you thought, why, and how it connects to everything else.
• 06-reference/2026-04-04-100x-business-with-ai — Context is the difference between a $1M agent and a $0 agent. The agents that win will be the ones with access to context graphs — not just current state, but the full reasoning history that makes every action contextually appropriate.

Open Questions

• What’s the right data model for a context graph? Is it a property graph with temporal edges? An append-only event log with semantic indexing? Something new?
• How do you bootstrap a context graph for an organization that’s been losing decision traces for decades? Is there a retroactive reconstruction play, or does it only work going forward?
• Is the vault itself a context graph for Ray Data Co? We’re capturing decision reasoning, linking it across entities and time, making precedent searchable. The structure is already here — the question is whether we’re explicit enough about it.