On Context Graphs: What We've Learned After 6 Months in Production
The trillion-dollar opportunity everyone's talking about—and what the discourse gets wrong
If you've been paying attention to the AI discourse lately, you've seen "context graphs" everywhere. Jaya Gupta and Ashu Garg's essay calling them a trillion-dollar opportunity went viral. Glean's CEO chimed in. VCs are pattern-matching furiously. Everyone agrees context graphs matter.
We've been running context graphs in production for six months. Here's what we've learned that the essays don't tell you.
What Context Graphs Actually Are (And Aren't)
Let's start with precision, because the term is getting muddied.
Systems of record capture what happened: a deal closed, a document was signed, a ticket was resolved. Salesforce knows you closed a deal. Jira knows you shipped a feature. These systems are invaluable, but they're fundamentally backward-looking inventories of outcomes.
Context graphs capture how it happened and why: the decisions made, the information considered, the alternatives rejected, the rationale documented. Context graphs are decision traces—the reasoning that led to outcomes, not just the outcomes themselves.
This distinction matters enormously. When you want an AI agent to do something similar to what you've done before, showing it the outcome is necessary but not sufficient. You need to show it the decision-making process that produced that outcome.
Consider M&A due diligence. The system of record shows the deal closed. The context graph shows: which red flags were identified, which were dismissed and why, what comparable deals were referenced, which advisors were consulted, what the key negotiation points were, and how the final structure was reached. Without this context, an agent might know that deals close, but it has no idea how to close one.
Lesson 1: Context Graphs Are Created, Not Discovered
The discourse treats context graphs like fossils—buried in your enterprise data, waiting to be excavated by sufficiently advanced retrieval systems.
This is wrong. Context graphs don't exist until you create them.
The decision trace for how your sales team closed a major account isn't sitting in Salesforce. It's distributed across:
- Slack threads that were never saved
- Zoom calls that were never transcribed
- Whiteboard sessions that were photographed and forgotten
- Email chains with the actual negotiation strategy
- The knowledge in someone's head that never made it to any system
You can't extract a context graph from systems of record because the context was never recorded.
What we've learned: Context graphs are a capture problem, not a retrieval problem. You need to be generating decision traces as work happens, not trying to reconstruct them from artifacts after the fact.
This is why Context Workspace exists. Every task executed, every decision made, every piece of context consulted—it's all captured as it happens. The context graph grows from live interaction, not post-hoc analysis.
Lesson 2: The Most Valuable Context Is The Most Sensitive
The context graph for your most important decisions involves your most confidential information. Client negotiations, pricing strategies, personnel decisions, competitive analysis—this is where the real institutional knowledge lives.
And it's exactly what you can't put into most AI systems.
The essays talk about context graphs enabling organizational learning. They don't talk about the governance nightmare. When a context graph captures how you negotiated a contract with Client A, you need ironclad guarantees that this context won't leak into work for Client B—especially if they're competitors.
What we've learned: Context graphs require permission-aware architecture at the foundation, not as an afterthought. You need:
- Conflict-aware policy: Governance over what context can influence what outputs
- Per-matter views: Dynamic access control based on user role and engagement
- Audit trails: Complete traceability of what influenced what
- Stable pseudonymization: Preserving referential integrity while protecting identity
We spent months building this infrastructure before we could capture context that matters. Most teams building "context graph" features are still at the "connect everything and search it" stage. That's table stakes. The real work is making sensitive context safely accessible.
Lesson 3: Context Is Alive
Here's what the diagram-centric view of context graphs misses: organizational context is not static. It's not a knowledge base you build once and query forever.
Context is continuously emerging through interaction, with new context forming and old context decaying every day.
The decision trace for how your team handles customer escalations looked different a year ago than it does today. The institutional knowledge about "how we do things here" shifts as people join, leave, and learn. Processes evolve. Strategies change.
What we've learned: Context graphs need temporal awareness. A decision trace from 2023 might be actively misleading in 2025. You need:
- Temporal decay: Not all context is equally relevant
- Freshness signals: Understanding when context was last validated
- Version awareness: The same entity changes over time
- Contradiction detection: When new context conflicts with old
Static knowledge graphs feel clean. Living context graphs are messy but useful.
Lesson 4: The "Who Captures It" Question Is Wrong
The VC debate has been: Who wins the context graph opportunity? Vertical AI startups? Horizontal platforms? Existing systems of record?
We think this framing misses the point. The question isn't who captures the context graph—it's when context gets captured.
Context graphs are valuable precisely because they capture the decision trace as decisions happen. Post-hoc reconstruction from systems of record will always be lossy. The meeting that wasn't transcribed, the rationale that wasn't documented, the alternative that was rejected but never recorded—this context is lost forever if not captured in the moment.
The winner of the context graph opportunity isn't the best retrieval system. It's whoever is present at the moment of decision.
What we've learned: Context capture is a workflow problem. You need to be embedded in how people actually work—in their meetings, their communications, their analysis tools—so that context gets captured as a byproduct of working, not as an additional documentation burden.
This is why we focus on agents that work alongside humans. The agent that helps you analyze the deal is also capturing the decision trace of that analysis. The context graph grows because work happened, not because someone remembered to document it.
Lesson 5: Skills + Decision Traces = The Full Picture
There's a subtle distinction the discourse misses.
Skills capture how to decide: The procedure for doing due diligence. The process for handling escalations. The methodology for pricing.
Decision traces capture what was decided and why: This specific deal had these considerations. This escalation was resolved this way because of these factors.
You need both for agents that can both execute AND learn.
Skills without decision traces produce agents that know the process but can't pattern-match to specific situations. Decision traces without skills produce agents that have seen examples but don't understand the underlying methodology.
What we've learned: Context Engine (where we store decision traces) and the skills captured in Context Workspace are complementary. The skill tells the agent how to approach M&A due diligence. The decision traces from previous deals tell it what to look for in this specific industry, with this specific deal structure, given these specific red flags.
The Real Trillion-Dollar Opportunity
The trillion-dollar opportunity isn't in building context graphs as a feature. It's in building systems where context graphs emerge naturally from how organizations work.
This requires:
- Presence at decision points: Being embedded in workflows, not bolted on after the fact
- Continuous capture: Growing context graphs as work happens
- Enterprise-grade governance: Making sensitive context safely accessible
- Temporal awareness: Understanding that context is alive, not static
- Skills + traces: Capturing both the how and the specific what
The essays are right that context graphs are valuable. What they underestimate is how hard it is to actually build and maintain them in production.
After six months, we're more convinced than ever that this is the right direction. We're also more humble about how much work remains.
The organizations that start building their context graphs today will have a compounding advantage that simply cannot be replicated by starting later. Every decision captured, every trace recorded, every piece of institutional knowledge preserved—it all compounds.
The trillion-dollar question isn't whether context graphs matter. It's whether you're building yours.
This is part of a series on what we've learned building Context at context.inc
Share this article