Introduction

Salesforce, Workday, and SAP built trillion-dollar businesses by recording what happened. The deal closed. The role was filled. The ticket was resolved. But as Foundation Capital's Jaya Gupta and Ashu Garg recently argued, none of these systems were designed to capture why — and in an era of AI agents, that missing "why" is becoming the most expensive gap in the enterprise stack.

These systems are "state-centric": they record the final output of human activity, but fail entirely to capture the reasoning, the precedents, and the messy, cross-system negotiations that led there. As enterprises transition to automating key workflows with AI agents, that reasoning becomes essential - not just as a reference, but as the guardrails for autonomy. Without it, agents risk making decisions that are technically correct according to the record but strategically catastrophic compared to the real decision.

This is precisely the gap that has given rise to the concept of the context graph: a representation of the reasoning that connects data to action.

The Balance vs the Ledger

Think about the difference between a bank balance and a ledger. The balance indicates there's $47,000 in the account currently. The ledger tells you why — every deposit, withdrawal, fee, and reversal that got you there.

The balance is useful. The ledger is intelligent.

Enterprise software has spent decades perfecting the balance. Salesforce tells you the deal closed at a 20% discount. Workday tells you the role was filled. ServiceNow tells you the ticket is resolved. What none of them capture is the reasoning trail: the competitor offer that forced the discount, the VP of Finance who approved it over Slack to land a lighthouse account, the precedent it set for every deal that followed.

Animesh Koratana frames this as the "Two Clocks Problem" in his recent post on context graphs:

• The state clock records what is true right now: the current discount, the current headcount, the current contract terms.
• The event clock records the sequence of reasoning that made it true: the negotiations, the exceptions, the trade-offs, and the approvals that produced that state.

The state clock is what gets stored. The event clock is what (currently) gets lost.

This gap is manageable when humans are making the decisions, because the people involved carry the context in their heads. But when an AI agent steps in, that institutional memory vanishes. The agent sees a 2x liability limit in a contract and treats it as "standard policy" - not as the hard-fought, one-time exception it actually was, negotiated at 11pm to save a relationship with a key client. Without the event clock, every precedent looks like a rule, and every exception looks like a baseline.

So the question becomes: how do you build the event clock at enterprise scale?

The industry is currently divided on how to build this. Many enterprise software incumbents are attempting to "bolt on" memory of these decisions to their existing systems, but they face two major hurdles:

1. The Silent Data Gap: Most actual decisions don't happen in a CRM or a hiring app; they happen in the "unstructured ether" of email, Slack, and Zoom.
2. The Join Problem: A context graph requires joining data across five different "coordinate systems" - events, timelines, semantics, attribution, and outcomes. These systems don't share keys, and it's hard to match a "Slack handle" to a "Salesforce ID" and a "recorded transcript" using a traditional SQL join.

Graph database companies are jumping on the bandwagon, but I would argue that the word “graph” in “context graph” is actually causing confusion.

This isn't about a specific technology solution or a new data schema: the context graph is a learned representation of organizational reasoning that emerges over time.

The challenge is not where we record it, but how we capture it in the first place.

Here is the good news: it turns out that AI agents don’t just need context graphs to perform better; they are the perfect vehicle to build them.

AI agents create the graph by doing work. As they traverse the organization to solve problems, their "trajectories" create the map. By following how an AI agent resolves an escalation, the system can learn the implicit hierarchy and logic of the organization. The ontology isn't specified upfront; it is discovered through use.

Enterprise Use Cases

To better understand the practical utility of context graphs let's look at a few example use-cases where judgment under uncertainty is the standard operating procedure.

Customer Experience and Support Orchestration

In standard customer support environments, an AI agent might know that a customer filed a complaint and know the company's refund policy, but it often fails to connect these facts to understand why an exception should be granted. A context graph enables the agent to traverse relationships between customer history, product configurations, and documented resolutions. By accessing past "VP overrides" related to similar product defects, the AI agent can autonomously propose a resolution that aligns with the firm's strategic priorities, reducing ticket resolution time and research overhead.

Strategic Capital Allocation & Budgetary Intent

In many organizations, the "state" of the budget is simply a line item in an ERP. But the "event" is the strategic rationale behind why a certain department was allowed to overspend while another was capped. A context graph allows a CFO’s AI Agent to move beyond simple variance reporting. Instead of flagging a 15% budget overrun as a "red" state to be corrected, the AI agent identifies the "event" trace: a series of executive Slack approvals and a revised project charter that prioritized speed-to-market over cost-efficiency for a specific Tier-1 initiative.

Insurance and Legal Precedent

In the insurance industry, a sector characterized by manual, document-heavy workflows, context graphs serve as the connective tissue that captures how brokerages actually operate. Similarly, in the legal space, context graphs link statutes to relevant case law, opposing counsel history, and internal firm precedents. This ensures that critical information is never overlooked during litigation strategy or compliance risk assessments.

Conclusion: The New Infrastructure of Intelligence

The transition from systems of record to systems of understanding might be the defining challenge of this decade. The companies that solve it won't necessarily be the ones with the best models, they'll be the ones that figure out how to capture reasoning at the speed of work.

We're still early. The tooling is immature, the representations are contested, and most organizations haven't even begun to inventory the decision logic they're losing every day in Slack threads and Zoom calls that no one will ever revisit.

But the promise is compelling: with AI Agents we need to transition from building a better system of record to building a true system of understanding. One that captures not just what happened, but why, and uses that knowledge to make AI agents that automate enterprise workflows smarter and more accurate.

At Vectara, we’re building the trusted agentic infrastructure that is already helping our customers turn scattered decision traces into durable organizational intelligence.

If you’re also moving in this direction, we’d love to share what we’ve learned alongside our customers and hear about your journey