Context graphs: what capturing the reasoning behind decisions unlocks for finance

Here’s an experience every enterprise AI user is familiar with: giving a model a task and rejecting the output – not because the model reasoned poorly, but because it was working with incomplete information.

"When a human doesn’t approve of a model’s output, it might be because they have some context the model doesn't," said Jaya Gupta, Partner at Foundation Capital.

This context isn't in one document that the model can easily retrieve. It often lives across Slack threads, escalation calls, deal desk conversations, and in people's heads. This forms what Gupta calls “the accumulated judgment of an organization.”

This insight led Gupta and her co-author Ashu Garg to write what became their viral essay, AI's Trillion Dollar Opportunity: Context Graphs.

The thesis behind context graphs: To act truly autonomously, models need access to something most enterprises don’t fully capture yet: the reasoning behind why a decision was made.

Gupta joined Ramp for a webinar to unpack context graphs and explore what they mean for finance teams.

Agents need more than rules

We all feed our agents rules, like “procurement requests over $X need approval.”

But Gupta explains that these rules capture what happened, not why something happened in a specific case. The missing layer, she argues, that captures the why: “decision traces,” a structured record of who decided what, with what information, under what constraints. Think: “procurement requests over $X need approval because the CFO flagged this spend category last quarter, and the last few similar requests were manually overridden because they were mislabeled as renewals.”

Decision traces, over time, form what Gupta calls a context graph, an agent’s source of truth for autonomy that captures searchable precedent.

Which companies are best positioned to capitalize on the context graph? According to Gupta, they are startups who sit in the execution path, where decisions are made. This architecture gives these startups a structural advantage: they are able to capture inputs and edge cases in real-time, rather than reconstruct them after the fact.

"If you're building a system that captures the decision state as decisions happen, you get a compounding asset. If you're trying to infer context from audit logs after the fact, you're fighting entropy," she says.

The green flags Gupta looks for in companies that are building toward or owning the context graph: product plasticity and comfort with uncertainty.

“Models are moving so quickly that you have to rewrite your product over and over again. Cognition reshaped their product, and Ramp, whose engineers I follow on X, does a phenomenal job at shipping quickly,” Gupta says.

“You have to absorb all the change happening on the ground underneath without breaking,” she concludes. “That also means you need a founder who's comfortable with a lot of chaos and uncertainty. The best leaders right now are managing not just the employee uncertainty, but the investor uncertainty as well.”

Finance is the perfect proving ground for the context graph

Gupta believes most functions have a cold-start problem on context graphs: they lack audit trails, structured schemas, and outcome signals. But finance doesn't, and this makes it the ideal proving ground for context graphs.

“Finance has regulations: there's trade logs, approval claims, compliance sign offs – which gives the function a head start on context graphs,” she says.

Finance also has an advantage since it is cross-functional. This allows it to capture rich decision traces across other functions like sales, legal, and procurement. To illustrate, Gupta points to Palantir: she estimates that roughly 75% of the company’s use cases run on top of SAP, suggesting that valuable context for AI is concentrated in finance and ERP systems.

"This happens because finance is at the heart of decision traces, exceptions, precedents, escalations – it sits at the intersection of different workflows," she notes. It’s harder to track down decision traces in another function, like sales for example, where reasoning might be less codified and more intuition-based. "GTM context graphs will come," Gupta says, "but the structural foundation isn't as strong."

The finance use cases Gupta is bullish on: variance analysis and contract-to-cash reconciliation.

Today, an agent might flag that a number is off, but it has no way to explain whether that variance is an anomaly. With a context graph, the agent can contextualize the flag against prior decisions. Similarly, instead of a finance team manually cross-referencing a contract against payment records, a context graph can give an agent full contract history, payment records, and exception logs simultaneously, enabling it to flag discrepancies.

The unresolved question in finance is knowing what precedents to remove and when. An assumption baked into a model one week may be wrong the next, Gupta says.

She notes that this is a remaining technical challenge, and Foundation Capital is actively looking for companies working on it.

B2B companies will move from retrieval to prediction

Gupta predicts that the context graph will become to B2B what behavioral signals are to B2C.

She explains that consumer internet companies like TikTok and Amazon became some of the most powerful companies by instrumenting behavior at a granular level. They record when users click, hover, pause, and abandon sessions, and feed those signals back into systems that learn and improve.

Enterprise software never had this level of context, because decisions are harder to observe. Gupta says that’s changing.

As context graphs accumulate enough decision history over time, B2B companies will move from retrieval to prediction. Gupta says the question will change from "how did we do this in the past?" to "if we do this in the future, what's likely to happen?"

She illustrated this shift with an example from one of Foundation Capital's portfolio companies. An incident response agent had built a context graph of a codebase, its tickets, and its observability data. The agent knew that changes made by a specific engineer often coincided with downstream bugs. With enough decision traces, the agent could move beyond retrieving past incidents to simulating new cases: “if Joe ships a change next time, what’s likely to happen downstream?”

“We're going to start seeing this kind of predictive ability in finance,” Gupta says.

The ten-trillion dollar question is permission

Context is necessary for agentic autonomy, but where does a company draw the line on data access? Gupta says this is the ten-trillion dollar question in AI.

“You are basically asking a CFO or a general counsel to let you into a company's decision-making processes and systems,” she notes. This begs the question: which model does a company trust to let into its operations?

This question gives rise to what Gupta sees as the next frontier in AI: permissions, or which model is allowed to touch what data, under what conditions.

According to Gupta, this is part of a broader framing shift. For the last two years, every enterprise has wondered if models are capable enough. Increasingly, the answer is yes. "If OpenAI and Anthropic said they weren't going to make the models better anymore, they'd still be massively successful companies,” Gupta says. “The labs are no longer just supplying intelligence, they're moving toward permissions, competing to become the control plane within a company."

She points to Mythos and Anthropic’s managed settings as signs of where AI is headed: the labs are trying to be the strongest model in simulating attacker capabilities (for testing) and at defending against misuse, attempting to earn their right to operate in high-trust, high-risk environments.

As permissions gain importance, Gupta sees the distance between application companies, infrastructure companies, and AI labs starting to collapse.

“Everyone is essentially selling one thing, and that's trust.”

Companies are changing how they work to better capture the ‘why’

Gupta says companies are already starting to build differently to capitalize on decision traces.

"We’re moving toward a culture of recording calls and doing work on instrumentable surfaces like Slack to make decisions more visible across a company,” she says.

Her advice to finance leaders who want to start acting on decision traces: find exception-heavy workflows (like accounts payable or bank reconciliation), and identify the mechanisms that capture data and decisions within that workflow.

Gupta gives the example of a company that started having people manually log their rationale for decisions in Monday standups, after learning about context graphs.

“It might not be the long-term solution,” she says. “But it is a start."

Jaya Gupta is a partner at Foundation Capital. You can follow her on X and LinkedIn.

Watch our full webinar “From Rules to Reasoning: How Context Graphs Are Powering Smarter AI” on demand.