Snowflake, Databricks and the Model Makers — Who Will Own the System of Intelligence in the Agentic Enterprise?
The emerging AI software stack that binds the agentic client to enterprise context, governance, and action.
The middle layer of the AI stack that harmonizes analytic data, converts siloed application logic into shared business rules, and captures institutional knowledge so that agents can observe business state, orient themselves, make decisions, and act with confidence. It is part specified and part learned — some intelligence is programmed through catalogs and semantic views; some is learned through usage, query history, skills, and human corrections.
The new system of engagement — Snowflake CoWork & CoCo, Databricks Genie, Microsoft Copilot, Google Gemini Enterprise, OpenAI Codex, Anthropic Claude Cowork — where business users, builders, and agents interact with data and get work done. The client becomes a teaching surface that feeds reasoning traces, corrections, and workflow patterns back into the SoI.
The key insight: The SoI and agentic client must be co-designed because each one teaches the other. This is Christensen's integrated innovation and Jensen Huang's extreme co-design applied to enterprise software.
Who is competing for the System of Intelligence control point, and where they start from.
Strongest in Layers 1-2 (mapping & rules) with Horizon Catalog, Polaris, Horizon Context, and semantic views. Moving into institutional memory via Cortex Sense. Agentic clients: CoWork (business users) and CoCo (builders). Key advantage: governed data gravity and the Observe acquisition for agent reasoning traces.
Pioneered the open data lakehouse with Delta Lake and Unity Catalog. Leaning into data intelligence with Unity Catalog, Genie for natural-language query, and expanding into governance. Strong in data engineering and ML workflows. Competes on open table formats (Iceberg/Delta) and multi-cloud neutrality.
Pushing Work IQ and Fabric IQ as the intelligence layer. Copilot is the agentic client across Office 365, Azure, and Fabric. Strongest ecosystem play — controls the canvas (Office) and the cloud (Azure). Build announcements focused on agent observability, evals, and CI/CD for agents.
Gemini Enterprise as the agentic client, with deep integration into Workspace and Google Cloud. Vertex AI for agent building. Strengths in frontier models, search infrastructure, and knowledge graph technology (a Semantic Web precursor). Leveraging massive web-scale context understanding.
Codex / ChatGPT as highest-volume agentic client. Frontier reasoning models (GPT, o-series). Strong in agent harness, coding, and general-purpose assistance. Missing piece: no enterprise back-end SoI to capture and harmonize interactions. Must move "downward" into data governance and business context.
Claude Cowork as agentic client with MCP (Model Context Protocol) evolving toward richer UI and back-end interaction. Strong safety and alignment focus. MCP direction supports the thesis that the client is becoming programmable and the backend intelligent. Also lacks enterprise SoI — must build or partner.
Clay Christensen's integrated innovation and Jensen Huang's extreme co-design, applied to the AI software stack.
The agentic client (SoE) captures user intent, corrections, skills, and reasoning traces. These flow into the SoI as signals. The SoI refines enterprise context — definitions, memory, relationships, rules — and feeds better context back to the client. Each side teaches the other.
"The system of intelligence is part specified and part learned." — Vellante & Gilbert
RAG-based chatbots failed because plugging an LLM into a vector database lacked deeper enterprise context. Agent orchestration alone cannot solve the trust problem. Agents need the SoI to understand business state and action boundaries. Without it, enterprises get useful point agents but no coherent operating model.
Query history, artifacts, skills, agent traces (what context pulled, tools called, outcomes), human corrections, popularity of definitions, lineage, shared workflows. The richer the interaction, the better the feedback. The better the feedback, the richer the context foundation.
The winning architecture connects user intent, business context, governed action, and continuous learning into a single reinforcing system. The company that best co-designs the intelligent client with the intelligent back-end will have the strongest claim on the next enterprise software control point.
How Snowflake is evolving Horizon from catalog governance toward active business context that feeds the System of Intelligence.
Pull context from databases, BI tools, ETL, schemas, dashboards, lineage, query logs, SaaS systems, and data lakes.
Add business meaning — lineage, popularity, semantic views, business glossary, descriptions, tags, ownership, certification — with human-in-the-loop validation.
Make context usable in CoCo, CoWork, Cortex Agents, and BI tools so agents produce better answers and more trustworthy actions.
Horizon Context is not purely automated. The operative phrase is "with human collaboration." Domain experts and builders must validate definitions, resolve conflicts, and encode how work actually gets done. Snowflake can infer patterns from query history and usage, but business meaning still comes from people who understand the domain.
"The future is the full ontology — modeling the verbs of the business, not just the nouns."
How the System of Intelligence matures from descriptive reporting to live operational intelligence.
Classic BI reporting — departmental cubes, metrics, dimensions. Semantic views and Unity Metrics push belong here. The enterprise is still in a "what happened" descriptive analytics world. Actions remain human-driven because the system lacks cross-domain context.
The enterprise harmonizes entities — one Customer, one Account, one Product across the estate. Master Data Management becomes foundational for AI because agents need a consistent view of entities they reason about. Cross-departmental aggregation becomes possible.
Events become first-class data. Streaming updates and real-time event flows update tables and context continuously. Customer interactions update product recommendations in real time. The system now understands sequences of events, not just snapshots.
The system classifies behavior: high-value shoppers, deal-driven replenishers, likely churners, suspected fraudsters. Patterns get named. The model interprets behavior rather than just recording transactions. Diagnostic analytics — asking "why" — becomes possible.
Predictions become part of the data model. Each entity carries continuously updated forecasts: "68% probability of repurchase within 30 days," "19% churn risk." Prescriptive analytics enables segment-level and eventually individualized recommendations with quantified confidence.
The model represents a web of related things informed by an Ontology — canonical customer established by deterministic reasoning and inference and associated with other canonical associated entities like orders, SKUs, promotions, inventory, fulfillment, carriers, payment status. Relationships are traversable (i.e., lookup friendly or dereferencable). Collectively, this expresses how metrics relate to business entities, in deterministic fashion. AI Agents and Skills can now answer richer "why" questions.
The Semantic Web also includes associated traces, lessons learned, and other context engineering issues associated with its construction.
The business starts modeling actions. A sale becomes a set of possible moves — apply a promotion, reserve inventory, authorize payment, split the shipment. Each action has preconditions and effects. Agents can reason about action spaces within governed boundaries.
The model becomes a live representation of the business. The state of every sale, customer, inventory, payment, and fulfillment path is represented in real time. Parts of operational application estates can be subordinated to the SoI as the live operating substrate.
The rules about how the business runs are managed as data — not buried in code. Changing a business process becomes a data update. Analytics can optimize the process itself. The enterprise can inspect, simulate, and improve its own operating logic.
Kingsley Idehen's alternative: hyperlinks as stable identifiers for loosely coupling identity, data, and intelligence.
While Snowflake, Databricks, and the hyperscalers compete for a centralized, tightly-coupled System of Intelligence within their proprietary platforms, the more open Semantic Web approach championed by Kingsley Uyi Idehen and OpenLink Software offers a radically different alternative: a decentralized, hyperlink-based architecture that uses HTTP URIs as stable, global identifiers to loosely couple identity, identification, authentication, authorization, and data spaces (databases, knowledge bases, filesystems, and APIs) — all without vendor lock-in, due to its use of existing open standards.
In this approach, hyperlinks act as stable standardized identifiers used to construct Knowledge Graphs that manifest a Semantic Web. Every entity — person, organization, product, data set, business rule, and agent — is denoted by a hyperlink that resolves to its structured description. The very same pattern applies to entity relationships and the ontologies defining the nature of entity and relationship types.
Every entity gets a unique, dereferenceable HTTP URI that serves as its global identity. Unlike Snowflake's Horizon-tied entity IDs, these URIs work across any system, any vendor, any graph. A customer, product, or business rule defined via HTTP URI is universally identifiable — no platform lock-in required.
Identity, identification, authentication, authorization, and data storage are all linked via hyperlinks rather than baked into a single platform. A WebID (HTTP URI identifying a person or agent) simultaneously serves as identity, authentication credential, and authorization key — loosely coupled through the Web's native linking mechanism.
RDF (Resource Description Framework) provides a universal data model that integrates databases, knowledge bases, filesystems, and APIs under one graph. SQL tables, NoSQL stores, REST endpoints, and SPARQL endpoints all become accessible as a unified hyperlinked data space. This replaces Snowflake's approach of centralizing data in a single governed warehouse.
WebID is a decentralized identity protocol that uses HTTP URIs to identify people, organizations, and agents. Unlike Snowflake's centralized user management, WebID is federated, cross-platform, and user-controlled. A single WebID can authenticate across Virtuoso, any SPARQL endpoint, and any WebID-compliant service.
Linked Data principles enable mesh integration across organizational boundaries without a central broker. Instead of Snowflake's Horizon Context collecting and enriching metadata into a proprietary context layer, Linked Data uses RDF links between independently maintained datasets. No vendor lock-in, no centralized governance bottleneck.
OpenLink Virtuoso is a hyperlink-backed platform that simultaneously serves as database, knowledge graph store, web application server, and linked data deployment engine. It provides the "System of Intelligence" without the centralized lock-in — SPARQL, SQL, GraphQL, REST, and WebDAV access to the same hyperlinked data space.
| Dimension | Snowflake's Centralized SoI | Semantic Web Hyperlink Approach |
|---|---|---|
| Identity | Platform-managed users and roles within Snowflake's governance boundary | HTTP URIs and WebID — decentralized, cross-platform, user-controlled |
| Data Model | Relational tables + semantic views within Snowflake's warehouse | RDF graph — integrates databases, KGs, filesystems, APIs via universal model |
| Entity Resolution | Horizon Catalog and Polaris — governed within Snowflake estate | Linked Data — entities resolve via HTTP URIs across organizational boundaries |
| Business Rules | Horizon Context semantic views + BI metrics — Snowflake-native | RDF ontologies (OWL, RDFS) — portable, reusable across any SPARQL endpoint |
| Context Layer | Horizon Context + Cortex Sense — proprietary, centralized | Linked Data context — federated, mesh-based, no central broker |
| Agent Traces | Observe acquisition — Snowflake-native observability | RDF provenance traces — portable, standards-based (PROV-O) |
| Vendor Lock-in | High — context and governance are Snowflake-native | None — standards-based (RDF, SPARQL, WebID, HTTP) |
| Integration | Collect → Enrich → Activate within Snowflake ecosystem | Native web — any HTTP(S) resource is a potential data source or target |
| Maturity Ceiling | Platform-dependent — limited by relational model for graph-like business representations | Ontology-driven — any level of expressiveness (RDFS, OWL, SHACL) |
| Governance | Horizon Polaris — centralized catalog governance | Web Access Control (WAC) + WebID — decentralized authorization |
The Snowflake SoI and the Semantic Web hyperlink approach are not mutually exclusive. Snowflake's data gravity in Layer 1 (mapping) and Layer 2 (rules) can be complemented by a Virtuoso-provided hyperlink-based context layer using RDF, WebID, and Linked Data. This layer sits above Snowflake's Semantic layer — using hyperlinks as stable identifiers to turn governed data into a globally addressable Semantic Web resource. Enterprises already invested in Snowflake that adopt this hybrid approach get Snowflake's analytic power + the web-like interoperability delivered natively by Virtuoso.
A seven-step pathway from fragmented data to agent-ready enterprise intelligence, grounded in open Semantic Web standards.
Identify all core enterprise entities (customers, products, orders, suppliers, employees, locations) and assign each a stable HTTP URI. Follow Linked Data principle 1: use URIs as names for things. Ensure URIs are dereferenceable and resolvable via standard web protocols. This creates the foundation where entities are identified independently of any single vendor platform, enabling cross-platform entity resolution at maturity level 2 of the nine-stage model.
Deploy WebID for every person, agent, and organization in the enterprise. Each WebID is an HTTP URI that serves as a decentralized identity for authentication and authorization. Unlike vendor-managed identity (e.g. Snowflake RBAC), WebID enables cross-platform authentication without a central identity provider. Use WebID-TLS or WebID-OIDC for secure, decentralized access control across databases, knowledge bases, filesystems, and APIs.
Represent all enterprise data spaces — databases, knowledge bases, filesystems, and APIs — as RDF graphs. Use schema.org vocabulary for interoperability. Each graph is identified by an HTTP URI and linked to other graphs through RDF triples. This replaces the Snowflake Horizon Context model of centralized metadata enrichment with a decentralized mesh of hyperlinked data spaces. Implement SPARQL endpoints for federated query across graphs, enabling the Federation of Intelligence vision.
Design an authorization model where access decisions are based on WebID relationships and linked data graph traversal rather than platform-specific roles and permissions. Use ACLs expressed in RDF (Web Access Control) where authorized agents are identified by their WebID URIs. This enables fine-grained, cross-platform authorization that naturally extends across organizational boundaries. Contrast this with Snowflake's approach where authorization is tightly coupled to the platform's RBAC system.
Define semantic views and business glossaries using RDF and SKOS, modeled on Snowflake's Horizon Context approach but published as Linked Data rather than locked in a proprietary catalog. Each business term (customer, revenue, churn, active user) gets an HTTP URI with rdfs:label, rdfs:comment, and skos:definition. Semantic views become SPARQL-queryable RDF graphs rather than platform-specific metadata. This enables the same business meaning enrichment as Horizon Context but in a cross-platform, standards-based format.
Implement the co-design feedback loop between the agentic client and the System of Intelligence using open standards. Agentic clients (CoWork-like interfaces) interact with enterprise data through SPARQL queries and WebID-authenticated access. Every question, correction, artifact, skill, and agent trace generates RDF annotations that enrich the enterprise knowledge graph. Use the Observe acquisition pattern: capture agent reasoning traces as RDF named graphs, evaluate them against rubrics expressed as SHACL shapes, and feed corrections back into the graph. This mirrors Snowflake's Cortex Sense approach but using open, cross-platform standards.
Establish governance policies for the hyperlinked enterprise knowledge graph using fine-grained, attribute-based access controls (ABAC) that apply equally to human users and AI agents. Define policies based on subject attributes (role, department, clearance, security group, agent capability profile), resource attributes (data classification, graph, entity type, predicate), and environmental conditions (time, location, authentication method, purpose of access). Layer these on top of the WebID-based identity and hyperlink-based authorization foundations — every access decision evaluates the agent's WebID attributes against policy rules expressed as RDF triples. Use SHACL for shape validation and OWL for ontology evolution. Add hyperlink-based cross-references between entity URIs across organizational boundaries. Promote reusable skills to governed ontology components. Continuously evolve the knowledge graph through the nine-stage maturity model, progressing from siloed snapshots through enterprise knowledge graph to process-as-data. The goal is a live Federation of Intelligence that combines governed data gravity with Semantic Web-style loosely-coupled cross-platform interoperability, where access policies are as portable and hyperlinked as the data itself.
Agent action is gated by the richness of the underlying data model and analytics built on top of it.
| Level | Analytics Maturity | What Agents Can Do | Human Role |
|---|---|---|---|
| L1-L2 | Descriptive (What happened?) | Answer questions, generate dashboards, conversational BI. "Talk to your data." No autonomous action. | Read and interpret reports |
| L3-L4 | Diagnostic (Why did it happen?) | Segment-level recommendations. Cohort-based heuristics. System identifies patterns but a human still pulls the trigger. | Review and approve segment campaigns |
| L5-L6 | Prescriptive (What should happen?) | Individualized, quantified recommendations. "Issue Maria free shipping — predicted to cut churn from 19% to 7%." Decision-grade guidance. | Oversee quantified recommendations |
| L7 | Action Modeling | Discover and compose actions within governance. Reason about preconditions: retain Maria by resolving complaint then issuing retention offer. | Set boundaries, handle exceptions |
| L8 | Live State Awareness | Act against live business state. If tent goes out of stock mid-flow, substitute or recommend alternative. | Govern boundaries, refine system |
| L9 | Process Optimization | Observe that resolving complaints before promotions retains more buyers. Recommend improvements to the process itself. | Set policy, teach the system |
The progression: Humans interpret static reports → Systems recommend segment actions → Quantified individual guidance → Agents plan, execute, and improve processes under governance. This is the path from reporting to action.
Stop treating AI readiness as a data modeling problem. Start building the governed intelligence layer that agents need to reason and act. That means:
"Data pros who master this shift — from moving data to modeling business meaning, context, actions, and learning loops — will become the architects of the agentic enterprise."
The middle layer of the AI stack that harmonizes analytic data, converts application logic into shared business rules, and captures institutional knowledge for agent reasoning and action.
The new system of engagement (CoWork, CoCo, Genie, Copilot, etc.) where users, builders, and agents interact with data and get work done. Also called the System of Engagement (SoE).
Snowflake's system for collecting, enriching, and activating metadata as active business context for AI, BI, applications, and agents.
Snowflake's layer for capturing tacit knowledge, ambient business context, user memory, skills, and artifacts — the beginning of institutional memory in the SoI.
Principles for using HTTP URIs and RDF to connect data across organizational boundaries without a central broker. Enables mesh-based, decentralized data integration.
A decentralized identity protocol using HTTP URIs to identify people, organizations, and agents. Enables federated, cross-platform authentication without a central identity provider.
The principle that the intelligent client and intelligent backend must be developed together because each teaches the other. Applied to enterprise software from Jensen Huang's hardware/software co-design philosophy.
The record of what an agent reasoned — context pulled, tools called, actions taken, outcomes — serving as raw material for observability, evals, and continual learning. The "new clickstream" of the agentic era.
Entities and relationships in the System of Intelligence and Semantic Web landscape.
Query the RDF knowledge graph via URIBurner's SPARQL endpoint. Select a recipe or write your own query.