Context Graphs Are Coming to BI

Ashu Garg at Foundation Capital recently published a thesis that caught our attention: AI’s Trillion-Dollar Opportunity: Context Graphs. His argument — that the next great enterprise software category will capture decision traces, the reasoning and context behind how organizations actually operate — resonates deeply with what we’ve been building at Spotonix.

We believe context graphs are coming to business intelligence. And the implications are profound.

Garg’s Thesis: Decision Traces Are the Missing Layer

The core insight: systems of record (Salesforce, Workday, SAP) capture current state — what happened. But they don’t capture why it happened — the reasoning, exceptions, precedents, and context that justified the decision.

This “decision context” currently lives in Slack threads, Confluence pages, email chains, and people’s heads. When someone leaves the organization, the context leaves too. The decisions remain but the reasoning evaporates.

Garg argues that startups positioned in the execution path — where decisions actually happen — can capture these traces systematically. Over time, these traces form a context graph: a queryable record of how organizational decisions were actually made.

What This Means for Business Intelligence

BI has the exact same problem, amplified.

An analyst answers “Which stores are losing habitual buying customers?” They produce a number. But the number is only half the value. The other half is the context:

• How was “habitual” defined? (Frequency-based? Value-based? Both?)
• What time period was used? (QoQ? YoY? Rolling 12 months?)
• Which customer segment was included? (All customers? Only active? Only with $500+/quarter spend?)
• What did “losing” mean? (Absolute decline? Rate of decline? Compared to benchmark?)

In traditional BI, this context lives in the analyst’s head, maybe in a Slack thread, possibly in a notebook comment nobody reads. When the analyst leaves, the number persists in a dashboard but the context — the reasoning — vanishes.

Text-to-SQL makes this worse. It generates SQL that produces a number, but the reasoning is completely opaque. There’s no trace of how “habitual” was interpreted, what assumptions were made, or how to reproduce the logic next time. Each query is an island of context that disappears the moment it executes.

The Analytical Context Graph

This is what Spotonix builds: a context graph for business analytics.

Every time Spotonix answers a question, it doesn’t just produce a number. It captures the full decision trace:

• The interpretation: “Habitual buyers” was resolved to customers with $500+/quarter spend, based on the existing Habitual Buyers segmentation your team validated last quarter.
• The composition: The answer was composed from three building blocks — the Habitual Buyers segmentation, a QoQ Customer Count Change calculation, and a Store x Quarter dimension.
• The validation: The user confirmed the interpretation before execution. The ambiguity between “frequency-based” and “value-based” was surfaced and resolved.
• The trace: Every building block traces to its definition. Every composition step is auditable.

This trace persists permanently in the context graph. The next time anyone asks about habitual buyers, customer behavior, or store performance, the system doesn’t guess — it discovers the existing context.

Why This Is Different from a Semantic Layer

Semantic layers (LookML, dbt metrics, Power BI’s DAX) define terms. They’re dictionaries. Important, but static.

A context graph captures how those terms are used in practice — the reasoning traces, the resolved ambiguities, the composition patterns. It’s the difference between a dictionary and a conversation history. The dictionary tells you what “revenue” means. The context graph tells you how “revenue” was actually used, combined, and interpreted across 500 questions by 30 different people.

Garg calls this the “institutional memory” that organizations currently lose with every departure. In analytics, the loss is quantifiable: we’ve talked to 70+ analytics leaders, and the consistent finding is that 40%+ of analyst time goes to re-answering questions that were already answered somewhere — because the context was never captured.

The Compounding Effect

Here’s where it gets interesting. Garg describes a “flywheel” where accumulated decision traces make future decisions faster and more reliable. The same dynamic applies to analytical context:

• Question 1: “Show me top stores by revenue.” The system builds 2 new building blocks. Took 45 seconds.
• Question 5: “Which stores are losing habitual buying customers?” The system discovers 3 existing building blocks, composes 1 new one. Took 12 seconds.
• Question 50: “Compare habitual buyer retention across channels for Q3 vs Q4.” The system discovers 6 existing building blocks, composes from all of them. Took 3 seconds.

As reviewed answers accumulate, they can become reusable traces. Where that context is persisted and retrieved, the context graph doesn’t just remember — it accelerates.

What We’re Building

Spotonix is the context graph for business analytics. We:

1. Discover existing analytical context before building anything new
2. Compose answers from proven building blocks — segmentations, calculations, and business analyses
3. Validate interpretations in business terms before executing — surfacing ambiguity instead of hiding it
4. Learn permanently — every validated answer enriches the context graph

The moat isn’t the LLM. It’s the ability to turn ambiguous business language into reusable semantic objects that survive clarifications and follow-ups. That’s not a prompt engineering problem. It’s a context graph problem.

Garg is right: context graphs are the next trillion-dollar opportunity. We’re bringing them to BI.

---

Read our whitepaper: The Interpretation of Business Questions

Demo access: try Spotonix