Colrows
Book a demo

Notes from the semantic execution layer

Whitepapers, technical deep-dives, and field reports from teams shipping enterprise AI to production.

First-Party Research

Semantic Layer Accuracy: 14.5% Raw vs 98.2% Compiled

Our own numbers: raw text-to-SQL scored 14.5% execution accuracy against 98.2% compiled through the semantic layer, plus a survey of 50 data leaders and the Cipla and SSP deployments.

Read the research
Thesis / Whitepaper

From Data Gravity to Semantic Gravity

Colrows' thesis on why semantic platforms become the control plane of the AI-native enterprise. As data gravity defined the cloud era, semantic gravity will define the AI era.

Read the thesis
Whitepaper Jan 2026

Semantic Layer Architecture for Enterprise AI

The full technical paper: system architecture, the seven-phase compilation pipeline, the Consensus Semantic Graph, join-path proof, drift detection, multi-layered semantic search, and compile-time governance. PDF editions linked inside.

Read the whitepaper
Comparative Analysis

Semantic Layers Were Built for Dashboards. AI Agents Need Something Else.

Colrows, ThoughtSpot, Timbr.ai, and Databricks Genie compared across four architectural generations, and why the AI agent era demands continuous semantic intelligence.

Read the analysis

Selected writing

The research pieces and architecture arguments worth your time. Pick a topic to see everything we have written on it, including the comparisons, pricing breakdowns and tool guides.

Selected 35 Company Brain 15 Semantic Layer & AI Agents 28 Model Context Protocol 12 Data Architecture & Modeling 9 Governance, Security & Compliance 13 Analytics & Search 16 Enterprise Strategy 11 Field Notes 1 Comparisons & Evaluations 36
A scattered cloud of grey and orange dots labelled a choice and a probability funnels into a structured grid where one clean orange route is compiled to a single governed output.
Semantic Layer & AI Agents

Jev AI vs LLMs: A Model That Cannot Hallucinate Still Needs a Semantic Layer

A former OpenAI researcher raised $40M for a model that never writes a sentence. Jev returns a choice and a probability. It still cannot tell you what "net revenue" means.

Read more
Left panel labelled Hands, no map shows grey agent nodes firing action arrows in every direction with orange error crosses; an orange arrow; right panel labelled Governed map shows the same agents routed through one central governed map tile that emits a single clean orange line to one result.
Semantic Layer & AI Agents

Agentic AI Bots Compared: They All Have Hands Now. None of Them Have a Map.

A plain-English field guide to 2026's agentic bots — OpenAI dots, Claude, Gemini, Grok, Meta Muse and more. They can all act now. Why none can act safely on your data without a governed map.

Read more
A tangle of grey nodes labelled Raw text-to-SQL on the left; orange lines converge into a central orange compiler ring; a single clean orange line runs to one node on the right, labelled One governed query.
Semantic Layer & AI Agents

Gemini 4 Argon Can Reason. It Still Cannot Compile Your Revenue Number.

Google shipped the best model yet. Point it at your warehouse and it still writes confident, wrong SQL. Why a deterministic semantic layer is the fix.

Read more
A tangle of grey AI agent nodes with crossing lines and orange collision marks, labelled Uncoordinated swarm, 18 of 30 branch collisions, on the left; an orange arrow; and orange agent nodes joined to one central writer, labelled Governed semantic graph, single writer, zero drift, on the right.
Semantic Layer & AI Agents

Multiplayer AI Fails Because You Treat It Like a Prompt Problem

Anthropic put 30 agents on a shared codebase; 18 collided on the same branch name. Multi-agent AI fails on coordination, not reasoning. The fix is the single-writer principle.

Read more
Scattered grey dots labelled Raw Text-to-SQL at 21 percent accuracy on the left, an orange arrow, and a clean orange connected graph labelled Semantic Execution at 100 percent accuracy on the right.
Semantic Layer & AI Agents

OpenAI Dots Have a Data Problem. Here Is the Fix.

OpenAI shipped always-on autonomous agents. The moment they query enterprise data, accuracy drops below 21%. The fix is a deterministic semantic execution layer.

Read more
Three AI agents answer 'What was Q3 revenue?': an ungoverned swarm returns three conflicting numbers; governed by Colrows, all three resolve against one contract and return the same number.
Semantic Layer & AI Agents

How to Build Multi-Agent Systems That Don't Fail: The Coordination Contract Enterprises Are Missing

Agents don't disagree because they can't talk. They disagree because they never agreed on what your metrics mean. The 2026 failure research, and the fix.

Read more
Two ways an AI agent answers a revenue question: a vector store path that free-hands SQL into a confident wrong number, and a Colrows governed semantic layer path that compiles one governed answer or a safe refusal.
Semantic Layer & AI Agents

AI Agent Memory Fails on Enterprise Data: Why a Governed Semantic Layer Beats a Vector Store

Vector stores let agents invent your numbers. Why a governed semantic layer is the semantic memory enterprise agents need, with 2026 benchmark evidence.

Read more
A week-one deployment flow: databases, catalog systems, BI tools, and query history crawled into one canonical form, analyst-certified knowledge as the source of truth, the first scoped subgraph, and a governed query shipped by Friday.
Field Notes

Field Note: Week One of a Colrows Deployment

No slideware. An engineer reads your schema, builds the first scoped subgraph, and runs the first governed query against real data by Friday.

Read more
A tenant B agent asking for tenant A data, restricted to a tenant B subgraph that contains only B fields, with the reach into tenant A fields marked uncompilable.
Architecture Updated

Multi-Tenant Semantic Isolation: Enforcing Tenant Boundaries at Compile Time

Cross-tenant leaks (ChatGPT Redis, DeepSeek) keep happening because runtime filters can be skipped. How compile-time semantic isolation makes a cross-tenant query uncompilable.

Read more
Three control points for an AI agent query, a jailbreakable prompt, a bypassable post-query filter, and the compilation boundary where RBAC and ABAC predicates are injected into the SQL before execution, the only point the agent does not control.
Governance & Security

How to Secure AI Agent Database Access: Why Post-Hoc Guardrails Get Bypassed

Guardrails are bypassed 65 to 84 percent of the time, and text-to-SQL models violate access rules up to 76 percent even when handed the rules. The fix is compile-time enforcement.

Read more
Two stacked timelines, transaction time for what the system knew and valid time for what was actually true, crossed by an as-of query cursor that reproduces the answer, the data, and the permissions in force at a past moment.
Governance & Security

Point-in-Time Query Reproducibility: The Audit Gap Costing Banks Billions

SEC recordkeeping penalties passed $2 billion and MiFID II gives 72 hours to reconstruct a trade. Why point-in-time query reproducibility is the missing piece.

Read more
A typed semantic graph at the centre of an enterprise, surrounded by INFER, VALIDATE, and GOVERN agents - the company brain in motion.
Enterprise Strategy Updated

Company Brain for Enterprise AI: Why the Data Layer Decides Everything

A company brain turns fragmented knowledge into a governed layer AI can act on. Why the data-semantics pillar decides if your agents are trustworthy.

Read more
Comparison diagram: left side shows unstructured company brain (Slack, emails, vector search, agent hallucinations); right side shows deterministic brain (AI agent, Colrows semantic layer, data warehouse, auditable SQL)
Enterprise Strategy

YC's Company Brain RFS: What Hyper, GBrain, and the Competition Got Right (and Wrong)

Hyper, GBrain, and Savant are racing to build the Company Brain. But they're solving 40% of the problem. The other 60% is metric consistency and governance—where the real value lives.

Read more
Five AI knowledge management approaches ranked by whether they only find knowledge or also govern the answer, with governed execution alone clearing every gate.
Company Brain

AI Knowledge Management in 2026: Why Finding Knowledge Is Not the Same as Governing the Answer

Every tool makes knowledge findable. Almost none make the answer trustworthy. Five approaches scored on the axis the AI knowledge management listicles skip: governance and correctness.

Read more
Four enterprise AI agents - analytics, action, assistant, and governance - each reasoning through a single shared semantic graph at the centre.
Semantic Layer & AI Agents Updated

Semantics for Enterprise AI Agents: The Deterministic Foundation for Reliable Autonomous Work

Why AI agents hallucinate, how errors compound across steps, and how a deterministic semantic layer makes enterprise agents reliable and auditable.

Read more
A vertical OS-style stack with applications on top, semantic OS kernel in the middle, and data sources at the bottom.
Semantic Layer & AI Agents Updated

The Semantic Operating System Inside the Enterprise

Why the semantic layer is becoming the enterprise operating system: semantic graph as kernel, MCP as syscalls, compile-then-execute as security.

Read more
A sealed static data-product crate transforming into a living, connected semantic graph.
Strategy Updated

Data Products Are Dead: The Era of Semantic Products

The data-mesh era is closing. The semantic-product era is opening.

Read more
Two cliffs separated by a chasm: the left cliff has a dense semantic graph and rising velocity, the right cliff is sparse and fading.
Strategy Updated

The Semantic Divide: Why Deterministic Infrastructure is the New Competitive Moat

Why future-ready enterprises will outpace the rest - and what's at stake for laggards.

Read more
A central semantic core surrounded by self-orbiting feedback loops - observe, learn, evolve - that re-feed the core continuously.
Architecture Updated

The Rise of Autonomous Semantic Systems

A new category of infrastructure that learns the enterprise - and updates itself.

Read more
A flat list of metric definitions transforming into a living concept graph with lineage, context, behaviour, and policy orbits.
Architecture Updated

Metric Stores to Knowledge Machines: The Evolution of Semantic AI

Why static metric definitions can't scale to AI - and what replaces them.

Read more
Three legacy modeling paradigms - Dimensional, Vault, and Metric Store - pressing against a cracked wall that opens onto a flowing semantic graph.
Architecture Updated

Breaking the 20-Year Deadlock in Data Modeling: From Tables to Meaning

Why dimensional, vault, and metric-store paradigms all hit the same wall - and what comes next.

Read more
Five metadata tools - data catalog, business glossary, lineage, observability, and dictionary - decaying around the perimeter of an orbital ring with arrows fading into a central glowing semantic-layer core.
Architecture Updated

The Decline of Metadata Tools: Why You Need a Semantic Compiler

Why standalone data catalogs failed: market evidence, vendor trajectories, and the structural reason semantic layers won.

Read more
Bar chart of the text-to-SQL accuracy cliff: 86-91% on the academic Spider 1.0 benchmark collapsing to 10-21% on real enterprise data.
Semantic Layer & AI Agents

The Text-to-SQL Accuracy Cliff: Why Deterministic Compilers Beat LLM Guessing

What Spider 2.0, BEAVER and BIRD actually measure, the three gaps that create the cliff, and what provably closes it.

Read more
Enterprise text-to-SQL accuracy benchmark: the 91% to 21% cliff
Analytics & Search Updated

The Enterprise Text-to-SQL Accuracy Benchmark: Every Major Study in One Place

The same model scores 91% on a textbook benchmark and about 21% on real enterprise data. Spider 1.0/2.0, BIRD, and BEAVER in one cited table and chart. Free to reference.

Read more
Two cards side by side: the semantic layer as the artifact holding metrics, entities, and policies, and the semantic compiler as the four-stage runtime that enforces it.
Semantic Layer & AI Agents Updated

What Is a Semantic Compiler? Deterministic SQL for AI

A semantic compiler resolves business metrics into deterministic, governed SQL. Definition, architecture, and a 5-point buyer test.

Read more
Three intent sources flow downward into a central glowing band labelled SQL Intermediate Representation, which then flows into three execution engines at the bottom - illustrating SQL as a compile target between intent and execution.
Architecture Updated

SQL as a Compiler Target: The Future of Governed Enterprise AI

Why deterministic semantic SQL compilation beats text-to-SQL: Spider 2.0 accuracy, compiler phases, regulation-ready auditability.

Read more
A natural-language question traced through three semantic hops - filter, join, aggregate - to produce a grounded answer.
Business Intelligence Updated

Multi-Hop Query Understanding: The Deterministic Compiler Approach

Multi-hop queries are where LLMs and traditional BI both fail silently. See why joins, cardinality, and ambiguous paths break text-to-SQL, and how a semantic execution layer makes multi-hop deterministic.

Read more
A typed semantic graph at the centre - entities, metrics, and events connected by relationships - orbited by INFER, VALIDATE, and GOVERN agents that drive continuous semantic consensus.
Architecture Updated

The Enterprise Memory Graph: Why AI-Native Companies Need a Memory They Can Trust

A technical deep dive into the six-layer architecture of semantic consensus.

Read more
Code-based access rules on the left and semantic-graph-bound policies on the right - illustrating governance bound to meaning, not files.
Governance & Architecture

The Semantic Control Plane: Deterministic Governance for AI

A semantic control plane declares, observes, and enforces what your data means at compile time, before any query runs. Why it is the next infra layer.

Read more
A clean lattice of concepts on the left gradually drifting and decaying toward the right; a side panel reports detected delta counts.
Technology Updated

Stop Semantic Decay: Why AI Needs an Autonomous Compiler

The new technical debt of AI systems - and how autonomous maintenance keeps the graph honest.

Read more
An iceberg metaphor: a small visible data-access layer above water, a stack of hidden costs below - semantics, dialect, RBAC, audit, drift.
Engineering Updated

The Hidden Cost of Building Your Own Data Access Layer

Roll your own semantic + governance + dialect handling - here's the bill.

Read more
A side-by-side comparison of semantic layer and knowledge graph for enterprise AI - metrics, definitions, relationships, governance versus entities, connections, context, reasoning.
Semantic Layer & AI Agents Updated

Semantic Layer vs. Knowledge Graph: Choosing Your AI Data Foundation

dbt 2026 benchmark: semantic layers hit 98-100% accuracy on covered queries. CypherBench: best LLM reaches 61.58% on knowledge graphs. Why deterministic execution wins on metric governance.

Read more
Four enterprise AI agents reasoning through a single shared semantic graph - the substrate that complements but does not replace document retrieval.
Semantic Layer & AI Agents Updated

RAG vs. Semantic Layer: Why AI Needs Deterministic Governance

RAG is retrieval-first; a semantic layer is compilation-first. Architecture, failure modes, cost, and when enterprises need both.

Read more
A question routed three ways: semantic search returns ranked passages, RAG returns a generated summary, and a semantic execution layer returns a computed governed number.
Comparisons & Evaluations

RAG vs Semantic Search: What Each One Can Actually Answer, and the Question Neither Can

Semantic search ranks passages. RAG summarises them. Neither computes a governed number. A decision guide for routing questions to the right architecture.

Read more
Why data catalogs (Alation, Atlan, Collibra) cannot execute AI agents deterministically.
Comparisons & Evaluations Updated

Why Data Catalogs (Alation, Atlan, Collibra) Can't Execute AI Agents

Catalogs document, ground, and govern metadata - they do not compile and execute governed SQL. The case for catalog + semantic execution layer.

Read more
Eight governed steps from an agent's question to a safe response, plus six supporting governance pillars.
AI Governance Updated

Governing AI Agents: Why Compile-Time Security is Mandatory

Governance must move from documentation to execution. A practical seven-layer model: identity, semantic resolution, policy enforcement, query validation, response guards, and audit trails.

Read more
Two side-by-side panels showing code-based access rules on the left and semantic-graph-bound policies on the right.
Governance, Security & Compliance Updated

Governance as Code to Governance as Semantics

Manual tagging decays, policy-as-code stays runtime, semantic governance compiles policy into meaning for provable, audit-ready compliance.

Read more
An MCP connection reaching data but ending in ambiguity, contrasted with the same connection passing through a governed semantic layer to a single precise answer.
Model Context Protocol

MCP Is Not Enough: Why Enterprise AI Agents Need a Governed Semantic Layer

MCP won the transport war. The dbt 2026 benchmark, the Gartner projection, and why meaning, not connectivity, is where production agents fail.

Read more
A USB-C plug labelled MCP connects through a governance gate that checks identity, RBAC, semantics and audit before reaching enterprise data.
Model Context Protocol

MCP Is the USB-C Port for AI Agents. Connectivity Is Not Governance.

The USB-C metaphor is correct. A port standardizes the plug, not identity, permission, or whether the answer is right. What a governed semantic layer adds behind every MCP call.

Read more
Open Semantic Interchange definitions flowing as portable YAML into a governed semantic execution layer that compiles dialect-perfect SQL across many warehouses.
Semantic Layer & AI Agents

Open Semantic Interchange and Apache Ossie: Portable Definitions Are Not a Governed Execution Layer

OSI makes semantic definitions portable across tools. An interchange format standardizes how meaning is written down. It does not compile, prove, or govern the query. Where a semantic execution layer still fills the gap.

Read more

Follow Colrows on LinkedIn

Product updates, technical deep-dives, and field notes from teams shipping AI to production.

Follow on LinkedIn Book a demo