The term gained momentum from a December 2025 essay by Foundation Capital, which argued that the next trillion-dollar enterprise software layer won’t come from better AI models — it will come from capturing how organizations actually make decisions. Two months later, every metadata vendor has rebranded around it. Here’s what the architecture actually requires, what exists today, and what’s still narrative.
01 What Is a Context Graph?
The term gained momentum from a December 2025 essay by Foundation Capital (Jaya Gupta & Ashu Garg), which argued that the next trillion-dollar enterprise software layer won’t come from better AI models — it will come from capturing how organizations actually make decisions.
The core insight is simple: your CRM records that a 25% discount was approved. Your ERP records the final price. But neither system records why — the customer was a strategic reference account, the VP invoked a partnership exception, and a similar precedent was set last quarter. That reasoning is the context. Making it queryable is the graph.
This matters now for two reasons. First, AI agents need this context to act autonomously. Without it, an agent is an extremely capable intern on day one — it can follow written rules but gets tripped up by every unwritten exception. Second, humans need it too — a new hire who could query how pricing exceptions actually work, a compliance team that could audit decision patterns across business units, a manager who could see whether the “one-time” exception from last quarter has quietly become policy. The context graph isn’t just an agent-enablement layer. It’s organizational intelligence made queryable, for both human and machine consumers.
02 The Debate That Started It All
The current discourse traces back to a two-part exchange between a SaaS analyst and a VC firm in December 2025. Understanding both positions is essential for evaluating vendor claims.
Thesis A — “Systems of Record Endure”
Agents don’t replace systems of record — they raise the bar for what a good one looks like. The more you automate, the more critical it is to know where the truth lives. Warehouses become “truth registries,” CRMs become “state machines with APIs.” It’s evolution, not replacement.
Thesis B — “The Missing Decision Layer”
Ball is right about truth — but wrong about what’s missing. The data agents need doesn’t exist anywhere yet. Decision traces, exception logic, and cross-system reasoning were never treated as data. The reasoning connecting data to action was never captured. That’s the trillion-dollar gap.
Both are correct. Systems of record will survive and remain canonical sources of truth. But a new layer is emerging above them — variously called “systems of action”, “decision record layers,” or “context graphs” — that captures how truth gets applied, bent, and overridden in practice. The disagreement is about who builds this layer: incumbents like Salesforce adding intelligence to existing data, or new entrants sitting in the agent execution path and capturing decision traces natively.
03 The Full Context Graph Architecture
No single vendor provides a complete context graph today. What follows is the full stack as described across the primary sources — with an honest assessment of what exists, what’s emerging, and what’s still theoretical.
The Context Graph Stack — Maturity Map
Consumption
MCP ServersAgent-facing interfaces
Search & DiscoveryHuman-facing portals
Agent FrameworksLangChain, CrewAI, etc.
Activation
Governance PoliciesAccess, classification
Automation RulesAlerts, workflows
Decision OrchestrationCross-system routing
Tooling exists today ▲
Context Store
Metadata LakehouseIceberg, open formats
Business GlossarySemantic definitions
Decision Trace StoreStructured precedent
Graph Model
Data Asset GraphTables, columns, dashboards
Lineage GraphColumn-level data flow
Process / Decision GraphWorkflow traces, precedent
Analytical Context
Schema & TechnicalAuto-crawled metadata
Quality MetricsMonitors, scores
Usage & PopularityQuery log analysis
Semantic Layerdbt, Cube, AtScale
Largely unsolved ▼
Operational Context
CRM / ERP WorkflowsSalesforce, SAP, etc.
Support PatternsZendesk, Intercom
Decision TracesApprovals, exceptions
Institutional KnowledgeTribal knowledge
Source Systems
WarehousesSnowflake, Databricks, BQ
BI ToolsLooker, Tableau, PBI
Pipelinesdbt, Airflow, Dagster
Operational AppsCRM, ERP, Ticketing
Production-ready tooling available
Existing systems (data sources / apps)
No production tooling exists yet
04 What Each Layer Means — In Plain Language
Source Systems
The plumbing you already own. Warehouses (Snowflake, Databricks), BI tools (Tableau, Looker), pipelines (dbt, Airflow), and operational applications (Salesforce, SAP, ServiceNow). These store
what happened. Every context graph sits on top of these — none replaces them.
Ball’s thesis is that these remain canonical, and he’s right.
Operational Context
The messy reality of how work actually happens. CRM workflows, support escalation patterns, approval chains, Slack threads where exceptions get negotiated. This is where
Foundation Capital argues the trillion-dollar gap lives. Decision traces — the reasoning behind exceptions, overrides, and precedent — currently exist in people’s heads and ephemeral communications.
No vendor captures this today. Critics question whether it’s even reliably capturable, since humans rationalize decisions post-facto.
Analytical Context
Structured metadata about your data assets. Auto-crawled schemas, column descriptions, data quality scores, usage patterns, freshness signals. This is what
Gartner’s 2025 MQ for Metadata Management evaluates. Active metadata platforms (Atlan, Alation, Collibra, Informatica) provide this layer well. It answers “what data do we have, where did it come from, and is it any good?” — but not “how should an agent
use this data to make a business decision?”
Graph Model
How entities relate to each other. Data asset graphs (this table feeds that dashboard) and lineage graphs (column-level data flow) exist and are well-established. The missing piece is the
process/decision graph: modeling workflows, decision points, approvals, and reasoning chains as first-class graph elements.
Analyst Sanjeev Mohan’s prediction: “Every company will claim to have a context graph, but the industry won’t be successful” — because context is too slippery to model operationally.
Context Store
Where context is persisted and made queryable. Metadata lakehouses on open formats (Apache Iceberg) exist and are production-ready. Business glossaries with semantic definitions are standard. What’s missing is a
decision trace store — a structured repository of
why decisions were made, who approved them, what precedent was invoked.
Diginomica observes that Foundation Capital’s architects “offer few pointers as to how to actually build” this, with little precedent and no defined architecture.
Activation
Making context drive action. Governance policies (who can access what, under what classification) and automation rules (alerts, triggers, workflows) are mature. Decision orchestration — routing real-time decisions across systems based on context graph queries — is the missing link.
Deloitte’s 2025 research found only 11% of enterprises have agentic solutions in production. The bottleneck isn’t model capability — it’s data architecture.
Consumption
How humans and AI agents access context. MCP (Model Context Protocol) — donated to the Linux Foundation’s
Agentic AI Foundation in Dec 2025 — has become the de facto standard for agent-to-tool communication. OpenAI, Google DeepMind, and Microsoft have adopted it. Metadata platforms like Atlan and Alation ship MCP servers. Human-facing search/discovery portals are mature. This is the layer with the most production tooling.
05 Hype vs. Reality — Claim by Claim
Context graphs are a real architectural concept addressing a genuine enterprise need. But the gap between thesis and production is significant. Here’s how the core claims stack up.
Context Graph Claim Maturity Assessment
Active metadata management
Real
Column-level lineage, auto-crawled
Real
MCP-based agent interoperability
Real
Metadata lakehouse on open formats
Real
Knowledge graph for data assets
Partial
AI context layer for agentic systems
Partial
Cross-system decision orchestration
Aspirational
Queryable decision traces & precedent
Aspirational
Institutional knowledge capture
Aspirational
■ Real — Production tooling ships today
■ Partial — Early tooling, not complete
■ Aspirational — No production solution exists
The pattern: the analytical context half of the context graph vision (metadata, lineage, quality, governance, MCP interfaces) is real and production-grade. Multiple vendors compete here. The operational context half (decision traces, institutional knowledge, cross-system orchestration) is the part that actually makes context graphs transformational — and it’s the part nobody has built yet.
Only 27% of AI-adopting enterprises had knowledge graphs in production by late 2025 — barely an uptick from 26% eighteen months earlier. The jump from knowledge graph (entities and relationships) to context graph (decision traces and precedent) is an order of magnitude harder. As one independent analyst put it: most decisions are implicit, never justified or explained, and deeply ingrained in how people think. Creating a formal decision trail so agents can learn from them requires solving a human problem, not a technology problem.
11%
of enterprises have agentic solutions in production (Deloitte, 2025). The bottleneck isn’t model capability — it’s data architecture. Only 27% of AI-adopting enterprises had knowledge graphs in production by late 2025. The jump to context graphs is an order of magnitude harder.
06 Who’s Playing Where
No single vendor covers the full context graph stack. Here’s where the major categories of players sit — and their structural advantages and limitations.
Context Graph Vendor Landscape — By Stack Layer
Operational Incumbents
Salesforce (Agentforce), ServiceNow (Now Assist), Workday — Own operational data but are
architecturally limited to current-state storage. A discount approval in Salesforce records the final price, not the context that justified it. Adding agents to these systems inherits their parent’s blind spots.
Warehouse / Lakehouse
Snowflake, Databricks, Google BigQuery —
Ball positions these as “truth registries” for the agentic era. Strong on data storage and compute. Weak on capturing decision context because they sit in the read path, not the write path.
VentureBeat identifies Snowflake’s agentic document analytics as an emerging bridge.
Agentic Startups
Vertical agent companies across sales, support, finance, legal —
Foundation Capital’s bet: these sit in the execution path, see the full context at decision time, and can persist decision traces natively. Structural advantage for capturing operational context. Structural disadvantage: each only sees one workflow.
Atlan CEO’s counter: enterprises have dozens of agents from dozens of vendors, each building their own context silo.
Graph / Identity
Neo4j, IndyKite, TigerGraph — Provide graph infrastructure.
IndyKite is explicitly building operational context graphs as a “decision fabric” connecting entities, events, and governing conditions.
GraphRAG (graph-powered retrieval for agents) is gaining traction as the technical backbone.
07 Evaluation Framework — Questions to Ask Any Vendor
When a vendor claims “context graph” capabilities, these eight questions separate shipped product from roadmap narrative.
01
Analytical vs. Operational
“Show me where your product captures decision traces — not metadata about data assets, but the reasoning behind business decisions.”
Most “context graphs” are really metadata graphs rebranded. The operational context layer is what’s new — and what’s hardest.
02
Lineage Depth & Autonomy
“Is your lineage auto-crawled at column level, or does it rely on manual curation or third-party tools?”
Native, automated column-level lineage is table stakes. If a vendor depends on Manta, IBM, or manual input for lineage, the “active” metadata claim weakens considerably.
03
Agent Interoperability
“Do you ship an MCP server? Can Claude, ChatGPT, or Cursor query your metadata natively?”
MCP has become the de facto agent integration standard. A platform without MCP support in 2026 is like a database without SQL in 2006.
04
Open Architecture
“Can I query your metadata store with standard SQL? Is it on open formats like Iceberg, or is it proprietary?”
Context graphs are multi-vendor by nature. Proprietary storage creates a context silo — the opposite of what context graphs are supposed to solve.
05
Connector Coverage
“Show me your connector list for my specific stack — not the marketing page, the actual maturity and freshness of each connector.”
Analytical context quality depends on connector coverage. A platform with 100+ connectors that are shallow is worse than 30 deep ones matching your stack.
06
Time to Value
“How long from contract to production value? Give me reference customers in my industry with their deployment timeline.”
Legacy platforms take 6–12+ months. Modern platforms deliver in 4–8 weeks. If the vendor can’t prove fast time to value with references, the “active” in active metadata is theoretical.
07
Total Cost of Ownership
“What’s the actual price with governance, quality, AI, and all modules I’ll need — not the base catalog price?”
Add-on pricing is common. The advertised price for a data catalog often doubles or triples when governance, quality monitoring, and AI features are included.
08
Shipped vs. Roadmap
“For every feature you just described — is it GA, beta, or roadmap? Can I see it in a live environment today?”
In a category this early, marketing routinely runs 12–18 months ahead of product. The gap between demo and production deployment is where most context graph claims fall apart.
The Bottom Line for C-Suite
The Concept Is Real. The Category Is Not.
Context graphs address a genuine architectural gap — AI agents need organizational decision context to act autonomously, and that context doesn’t live in any system today. The analytical metadata half (schema, lineage, quality, governance) is mature and shippable. The operational context half (decision traces, institutional knowledge, cross-system precedent) is 12–36 months from production at minimum.
What to do now: Buy metadata platforms for what they are — strong active metadata capabilities that make your data discoverable, governed, and AI-ready. These are necessary investments regardless of where the context graph vision lands. But don’t pay a premium for “context graph” branding on capabilities that are metadata catalogs with better marketing.
What to watch: Where decision traces actually get captured. Today, most organizational decisions are made by humans — in email threads, Slack messages, approval workflows, and meeting side-channels. Those traces are unstructured, scattered, and often never written down. As AI agents enter real workflows (approvals, escalations, pricing), they’ll generate decision traces as a byproduct of execution — and those traces will be natively structured and programmatically loggable. But agents are one capture path, not the only one. The connectors to reach human communication channels already exist, and LLMs are increasingly capable of extracting structured decision context from unstructured messages — identifying the decision, the reasoning, the precedent, and the stakeholders from a messy email thread. The context graph needs to absorb decision context from both human and machine sources. The vendors that solve this dual-capture problem — using LLMs as the extraction engine for human context, persisting both alongside clean agent traces, and making the combined graph queryable — will define the category.
The “trillion-dollar” framing is VC narrative positioning. The real opportunity is large but will emerge incrementally — workflow by workflow, decision by decision — not as a single platform purchase. As one analyst observed: this isn’t a technology problem anymore. It’s a data maturity problem. And data maturity takes time.
Practitioner Takeaway
The analytical half of the context graph — metadata, lineage, quality, governance, MCP interfaces — is real, production-grade, and worth buying today. The operational half — decision traces, institutional knowledge, cross-system precedent — is where the transformational value lives, and nobody has built it yet. Invest in the foundation now. Watch for the vendors who crack the dual-capture problem: machine-generated agent traces alongside LLM-extracted human decision context, unified in a single queryable graph. That’s where the category actually gets defined.
Context Graphs — C-Suite Briefing — March 2026
This briefing synthesizes 40+ sources from August 2025 through February 2026, including Foundation Capital, Clouded Judgement, Gartner, Deloitte, and independent analyst commentary. All references below.
