AI Publications

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

The last era of enterprise software produced giants by becoming systems of record: the place where data and workflows “officially live.” Think CRM, ERP, HRIS—tools that define what happened inside a business.

AI agents won’t replace those systems. They’ll sit on top as the new interface: you ask, they act, they coordinate. But for agents to work reliably, they need more than clean APIs and better governance. They need the missing layer that actually runs most companies day to day: decision traces.

The missing layer: decision traces

In real operations, the important moments rarely look like a tidy workflow. They look like:

• an exception approved “just this once”

• a policy override after a call with a key client

• a discount that breaks the rule but matches precedent

• a risk review that happens in chat, not in the tool

And that’s the problem: the system of record usually stores the outcome (“20% discount”), but not the rationale (why it was allowed, what evidence supported it, who signed off, what precedent existed).

So when we say “agents in real workflows,” the biggest blocker isn’t missing data. It’s missing reasoning artifacts—the context humans rely on every day but don’t consistently capture as structured data.

What a context graph is

A context graph is the accumulated structure of those decision traces over time. It’s not a model’s chain-of-thought. It’s something more practical: a persistent, queryable record of why decisions happened and why they were allowed. If systems of record are the ledger of state, context graphs become the ledger of judgment.

How it gets built (and why it compounds)

The context graph emerges when the agent orchestration layer is instrumented so that every run emits a trace, for example:

• what inputs were pulled (tickets, contract terms, usage, prior cases)

• which policies were evaluated

• what exceptions were invoked

• who approved (or what rule authorized)

• what was written back to each system

Imagine an agent proposing a discount above the policy limit. It gathers evidence across systems (support incidents, escalation history, renewal risk, prior approvals), routes the exception to the right approver, then records the full trace. The CRM may only keep “20% discount,” but the context graph keeps the story.

That story is valuable immediately (auditability, debugging), and it becomes more valuable over time because repeated exceptions turn into searchable precedent. Even with human-in-the-loop, the trace library compounds into a durable asset.

Why incumbents struggle to build it

Incumbent platforms are structurally disadvantaged for two reasons:

1. They optimize for current state, not decision context. They’re designed to store “what is,” not “why we did it.”

2. They often don’t sit in the execution path across systems. A CRM “agent add-on” inherits the CRM’s limits; it can record the final field updates, but not the full decision-time world state and rationale across tools.

   
   

Meanwhile, data platforms typically sit in the read path (ETL after the fact). They can tell you what changed, but not the context that disappeared at commit time. Capturing decision traces requires being present at the moment decisions are made and executed, across systems. That’s where “systems of agents” have an advantage.

Three paths for startups

If you’re building in this space, there are three clean strategic routes:

1. Replace the system of record with an AI-native one (hard, but possible during major platform shifts).

2. Replace a module where exceptions and approvals concentrate, syncing final state back to incumbents.

3. Start as orchestration, evolve into a new system of record—where the durable product is the decision trace itself, and the business comes to you to answer: “why did we do that?”

Where to look: signals that context graphs matter

Two signals reliably point to opportunity:

• High headcount doing manual coordination/reconciliation (a sign the logic is too complex for traditional automation)

• Exception-heavy decisions where precedent matters and “it depends” is the real rule.

A third signal is especially powerful: glue functions that exist at the intersection of systems—RevOps, DevOps, SecOps. These roles exist precisely because no single tool captures cross-functional context. Automate the workflow and you can also capture the decision traces those teams implicitly generate every day.

The takeaway

Systems of record aren’t going away. But the next major platform opportunity may be the system that captures what they don’t: how decisions actually get made.

A context graph turns invisible judgment into durable data—making autonomy safer, workflows more auditable, and edge cases learnable. The trillion-dollar category is a new “system of record” for decisions—where the why is stored, searched, and reused.