The next major platform shift in enterprise software will not come from building a better system of record. It will come from capturing how decisions are made
At the end of 2025, Jaya Gupta and Ashu Garg from Foundational Capital shared a perspective that the gap to unleashing AI in enterprises is not in owning more data or compute, but in capturing the decision traces that systems of record systematically discard. This gap becomes existential as software transitions from passive workflow tooling to autonomous, agentic systems operating at scale.
The Fishing Merchant’s Lost Wisdom
Before software, decision-making was inherently contextual. The village economy revolved around a small group of salmon merchants selling the daily catch from the dock. There were no pricing systems, inventory tools, or written policies.
Each merchant carried an internal model of the village. They knew which buyers were price-sensitive, who required flexible payment terms, and which relationships mattered more than margin. When a customer received a discount, it reflected situational context: an incoming storm, freezer capacity nearing its limit, a past favour, or the risk of spoilage. All of this logic lived in people’s minds.
As the village grew and merchants were operating on a larger scale, decisions became inconsistent. Two customers buying the same salmon on the same day could receive different prices, with no shared explanation. What followed are conflicts (e.g., disputes from customers, attribution from employees, and conflicts on different versions of truths). As time goes on, the merchants spend more time debating past decisions than making new ones.
In the 2010s, the advent of software stabilized the surface area of the system. Digital ledgers and inventory systems introduced a shared record of customer records, transactions, stock levels, and revenue. For the first time, the merchant had a canonical view of what happened: how many salmon were sold, at what price, and to whom.
But something subtle was lost in the transition. The software captured outcomes, not reasoning. It could show that a crate of salmon was sold at a discount, but not why. It couldn’t see the approaching storm, the constrained freezer space, or the competitive pricing pressure from a nearby port. That context still existed, but it lived outside the system, scattered across email replies, WhatsApp threads, and offline conversations.
This is the core limitation of systems of record. They preserve outcomes, not decision traces.
AI Needs Context to Work
Today’s enterprise applications (e.g., Salesforce, Workday, SAP) have become the bedrock of business by locking down canonical data and owning workflows around it. But they remain static vaults, each optimized for a narrow domain: an opportunity, an employee, an invoice. They are excellent at describing what the world looks like right now, and blind to how it got there.
They are incredibly good at telling you what the world looks like right now, but they are blind to the "why" behind it. These legacy platforms own the data, but they completely lose the decision traces, the actual context, tradeoffs, exceptions, and tribal knowledge often reside outside of the platform, across other point solutions, email threads and offline interactions.
When it comes to AI applications, adopting AI vendors on top of these records inherits the same blindness. When they are only trained on outcomes, agents lack access to operating intent and precedents. They either over-apply brittle rules, hallucinate reasoning that was never there, or require constant human intervention. The lack of organizational context and verifiability has become the core hurdle to value realization in AI across the enterprise.
What is a Context Graph
Ashu Garg defines context graph as “institutional memory for how an organization makes decisions: not how the process doc says it should, but how it actually works in practice.”
If we view an organization's know-how as a network, where customers, product usage, and billing records represent the nodes. Building a context graph is the act of establishing causal links between them, the relationships that explain how those nodes interact over time.
In most modern enterprises, this context is fragmented, implicit, and inaccessible because systems are not designed to communicate with each other. The construct of a context graph sits above all systems and captures not just events and states, but the causal chain behind decisions
For examples:
Cross-System Synthesis: A salmon merchant decides to discount a shipment after checking freezer capacity, reviewing unsold inventory from the prior day, and hearing from the dockmaster that a storm may delay tomorrow’s catch. The ledger records only the discounted price, losing the synthesis of operational signals that drove the decision.
Exception Logic: The merchant consistently gives a preferred rate to a regional fishmonger who buys in bulk and pays late, knowing their distribution network absorbs volatility during slow weeks. This exception is understood among senior traders but never written down, so new apprentices default to list pricing and trigger disputes.
Approval Chains: A junior seller offers a price concession after getting a nod from the head merchant during a shouted conversation on the dock. The transaction record reflects the final price, but not who approved the deviation or the reasoning behind it at the moment of sale.
Key Elements to Context Graph
Context graphs do not replace systems of record. They emerge as an orchestration layer, capturing where decisions are evaluated, executed, and recorded in time. Three properties matter.
A. Opinionated, domain-specific data models (Quality of Context Nodes)
Legacy systems rely on generic abstractions like “Account” or “Opportunity” to serve broad markets. These abstractions force humans to translate real-world decisions into software. Context graphs begin with narrow ontologies that reflect operational intent, embedding reasoning directly into the data model
B. Inferential density (Links Between Nodes)
As more nodes are added across the product and service lifecycle, the graph doesn’t just grow wider, it grows denser. With enough relational richness, the “why” behind an outcome becomes reconstructible and verifiable.
C. Verifiability (Strength of the Links)
Verifiability is the key to training AI agents for an enterprise-grade. A verifiable system can be interrogated and audited at the level of individual decisions. It preserves the full state of the world at decision time, including the signals observed, the policies in force, the precedents invoked, and the tradeoffs evaluated.
Who’s Best Suited to Capture the Context Graph?
Matt Brown’s article “Vertical software already won the context graph” depicted that vertical platforms that are designed for a specific vertical have already provided the foundation to tackle the context problem:
First, vertical software begins with domain-specific ontologies. Its data models are built to reflect the operational reality of a single industry (e.g., Toast for restaurants, Veeva for healthcare) rather than generic abstractions designed to span many. Horizontal platforms model entities like “Account” or “Opportunity,” which require humans to supply the missing context. Vertical platforms start with objects that already encode intent and causality. In Toast’s case, entities like Order, Menu_Item, Customer, and Prep_Station are inherently linked: orders tie to customers, customers to inventory, and inventory to fulfillment.
Second, vertical platforms consolidate workflows by default. Point solutions such as POS, CRM, scheduling, staffing, procurement, and billing live inside a single system. This consolidation increases inferential density without explicit effort. Each additional workflow adds new nodes and strengthens the links between them. As platforms expand, they capture more of the decision surface.
Third, inferential density compounds into product advantage. With dense context as a core asset, vertical platforms are best positioned to expand into judgment-intensive, exception-heavy workflows like marketing, procurement, and bookkeeping. These are not platform feature extensions, but a strategic shift toward service-like products. Because decisions are already grounded in rich, verifiable context, agentic execution becomes a natural progression rather than a leap of faith.
Read More