thinking sense · Concepts
The 7 Layers of Enterprise Intelligence
Like the OSI model for networking, enterprise AI has a layered architecture — from raw classification at the base to a governed AI Control Plane at the top. Most of the industry has mapped five layers. thinking sense auto-discovers the ontology at Layer 2 — resolving entity relationships across your systems, flagging conflicts for human approval — then builds the two layers the field guide doesn’t cover: the Semantic Evolution Harness that executes and generates clean training signal, and the AI Control Plane that fine-tunes task-specific models and governs everything above it.
Extends the field guide: Ontology vs. Taxonomy vs. Knowledge Graph vs. Semantic Layer vs. Context Graph — dataengine(), Aug 12, 2026
Layer by layer
Taxonomy
Output: Classifies
A hierarchical classification scheme — things organized into is-a and part-of trees. Think product catalogues, org charts, content tagging, library systems. The starting point for any enterprise knowledge effort. Without consistent labels, nothing above it can be built reliably.
Ontology
Output: Defines
“Semantic layer + knowledge graph + vector search. Auto-discovered, federated, data in place. That is the Governed Ontology Layer.”
The formal vocabulary of enterprise data — but a real enterprise ontology is three things the industry currently sells separately. A semantic layer resolves what data elements mean in business terms: "revenue" in SAP is not the same as "revenue" in Salesforce. A knowledge graph maps how entities relate across systems: this customer account, this SAP debtor, and this contract counterparty are the same real-world entity. Vector search retrieves semantically similar unstructured context: the policy document, the contract clause, the email thread relevant to this question. Palantir calls their version an ontology too — but it requires moving all your data into their platform first, built by Forward Deployed Engineers over months. thinking sense auto-discovers all three layers from your existing systems — schemas, primary/foreign keys, data overlap, query history, and uploaded documents — with data staying in place. Where entity mappings are unambiguous, it resolves silently. Where they aren't, it surfaces the conflict for human approval.
Knowledge Graph + Semantic Layer
Output: Facts + Meaning
Two parallel layers. The Knowledge Graph populates the ontology with real instance data — actual entities and typed relationships queryable at scale, enabling multi-hop reasoning across sources. The Semantic Layer runs alongside it, mapping raw warehouse columns (fct_orders.amt_usd) to governed business terms (Revenue) so every tool — Tableau, Python, or an AI agent — uses the same definition.
Context Graph
Output: Retrieves
The AI-era retrieval and assembly layer. At query time, it selects the relevant subset of knowledge — entities, documents, relationships, session state — and hands it to the model as context. What makes RAG more than text search. But it stops at the model's context window: it retrieves and assembles, then the model synthesizes. This is where the established field guide stops.
Semantic Evolution Harness
Output: Executes
“Context Graph retrieves. Semantic Evolution Harness executes — and generates the clean signal that makes fine-tuning possible.”
Where thinking sense begins. The Semantic Evolution Harness doesn't assemble context for a model to reason over — it compiles intent against the enterprise semantic graph and executes a typed, governed plan. Every execution path is semantically grounded and policy-enforced. That means two things: every answer is verified and auditable, and every execution trace becomes clean training signal for task-specific models. General models train on the internet. thinking sense trains on your enterprise's actual meaning, relationships, and policies. That's a moat no foundation model provider can replicate.
AI Control Plane
Output: Governs
“The models change. Enterprise context — and the models trained on it — stay.”
The operating system above semantic execution. Routes queries to any model — open or frontier — with governance policy compiled at plan time, not at answer time. Governs Execution Distillation: every execution is verified against identity, policy, provenance, and business rules — those verified traces distill into task-specific models (Meaning Model, Policy Model, Execution Model). Stable skills internalize into model weights; live facts and permissions stay in the active context graph, governed on every execution. The graph remains the live source of truth. The model learns how to operate against it. Models change. Governed enterprise context compounds.
The missing layers
Knowledge at rest.
Activating context.
The bottom five layers describe knowledge at rest — classified, defined, populated, translated, assembled. The Semantic Evolution Harness activates that knowledge into a governed, verifiable answer and produces verified traces. The AI Control Plane distills those traces into enterprise-specific models through Governed Execution Distillation — teaching them your enterprise's language, decision patterns, and workflows. Models change. Governed enterprise context compounds.
No lab builds these layers for you. OpenAI wants your prompts. Your hyperscaler wants to run the inference. Your catalog vendor wants to own your semantic layer. The AI Control Plane has to be vendor-neutral by design — because none of them can hold it neutrally.
The complete picture
How it all comes together with thinking sense
Business questions enter at the top. The Governed Ontology Engine — spanning semantic understanding, knowledge and ontology, execution and orchestration, model routing, and verification — converts them into governed, auditable answers. The RLBUF flywheel feeds every verified outcome back into the system. Intelligence compounds.
The Enterprise AI Harness for Real Outcomes · thinking sense Governed Ontology Engine
Industry convergence
On August 16, 2026, Salesforce President & Chief Platform and Engineering Officer Rohan Kumar independently named the AI Control Plane as the foundational layer of the agentic enterprise — with trusted governance, context, and action sitting on top.
When the President of Salesforce arrives at the same architectural conclusion independently, the category is no longer a thesis. It is a consensus.
Read Rohan Kumar’s article →