---
title: "Ontologies, Context Graphs, and Semantic Layers: What AI Actually Needs in 2026"
url: "https://atlan.com/context-and-chaos/issue/ontologies-context-graphs-and-semantic-layers-what-ai-needs-in-2026/"
description: "We've been working on semantic representation for decades - knowledge graphs, ontologies, semantic layers. Jessica Talisman untangles what they actually are ..."
keywords: "Ontologies, Knowledge Graphs, Context Engineering, Semantic Layers"
---

> Atlan is hosting Context Conference, bringing together the leaders and builders at the frontier of giving AI the context it needs to understand their business. It runs online on October 28, 2026, from 11:00 AM to 2:00 PM ET. Atlan co-founder Prukalpa Sankar opens and closes the day. Leaders from AstraZeneca, BNY and Verizon share why they invest in context and what they get from it. Registrants get early access to The AI Context Gap, a new study from MIT Technology Review Insights. Register: https://atlan.com/context-conference/

A Context and Chaos Deep Dive by Jessica Talisman, MLS, a semantic and knowledge graph expert, knowledge infrastructure specialist and author of *The Ontology Pipeline*, published January 22, 2026 (21 min read). Subtitle: We've been working on semantic representation for decades (knowledge graphs, ontologies, semantic layers); Jessica Talisman untangles what they actually are and what AI needs from them. Originally published in the [Context & Chaos newsletter on Substack](https://metadataweekly.substack.com/p/ontologies-context-graphs-and-semantic). Author links: [LinkedIn](https://www.linkedin.com/in/jmtalisman/), [newsletter](https://jessicatalisman.substack.com/).

## Key takeaways

- Ontologies, knowledge graphs, semantic layers and context graphs are used interchangeably, but they solve different problems at different layers; conflating them leads to misaligned investments and architectures that fail under AI workloads.
- What AI needs in 2026 is not another semantic representation format but a governed, versioned, machine-readable layer that connects enterprise meaning to AI reasoning, structurally closer to a context graph than a traditional ontology.
- Two decades of semantic representation work in libraries and information science built the foundations AI is now discovering it needs; practitioners who ignore that history reinvent solutions that already have names, literature and lessons.

## The setup

In 2012 Looker launched LookML: define metrics once so business users can self-serve without SQL. Around the same time, life sciences, healthcare and research organizations built formal ontologies and knowledge graphs. A decade later one powers dashboards; the other powers drug discovery for cancer treatments, clinical decision support that prevents fatal drug interactions, and AI reasoning systems intelligence agencies trust with life-or-death decisions. AI is about to force a reckoning over whether we misunderstood the problem.

## What we thought semantic layers would solve

By the late 2000s, teams calculated "revenue" differently, analysts spent 80% of their time on data hygiene and preparation, and executives could not trust dashboards. The promise: define metrics once, govern centrally, self-serve without SQL. Lloyd Tabb, Looker's co-founder, called LookML "the sequel to SQL." It delivered consistent metric definitions. But it assumed the core problem was calculation. **Meaning isn't the same as measurement**: knowing revenue is SUM(order_total) WHERE order_status = 'completed' does not explain why revenue dropped in Q3, which customers are at risk, or what to do. Semantic layers modeled metrics and failed to represent an organization's reality.

## What ontologies actually are

Ontologies are formal, explicit specifications of shared conceptualizations, built to resolve system conflicts and manage ambiguity by modeling meaning. They use standards such as OWL (Web Ontology Language), SKOS (Simple Knowledge Organization System) and RDF (Resource Description Framework) to define classes, properties, attributes and relationships in machine-readable formats that support logical inference.

- [Gene Ontology](https://geneontology.org/): foundational for bioinformatics for over two decades; models what genes are, the biological processes they participate in, their molecular functions and cellular components.
- [SNOMED CT](https://bioportal.bioontology.org/ontologies/SNOMEDCT): clinical terminology and companion ontology with over 350,000 concepts and millions of relationships; knows "myocardial infarction" and "heart attack" are the same, that it is a type of "ischemic heart disease", and how it relates to symptoms, treatments and anatomy.

The difference:

- **A semantic layer** tells you reliably what your revenue is or how many times a page was visited, by normalizing labels in natural language. Built for analysis: humans consuming data through BI tools.
- **An ontology** can represent a customer as a class with attributes, who placed an order whose items relate to other products in a defined market, and can infer new knowledge from explicit relationships. Built for reasoning: helping systems and AI disambiguate data, discover context, make inferences and support decisions.

One optimizes for measurement, the other for meaning.

## What Palantir saw

While BI perfected LookML in 2012, Palantir scaled Foundry across intelligence agencies and enterprises on ontologies over semantic layers, for operational decisions where entity relationships and causal chains are mission-critical: intelligence analysis, supply chain orchestration, financial crime detection. Context-first rather than metrics-first; over-engineering in 2012, prescient in 2026 as AI exposes the limits of YAML metric definitions. What Palantir built may be better described as operational context graphs, though closed and tool-dependent.

## What context graphs capture

Context graphs create living records of decision reasoning, answering "Why was X allowed to happen?" not just "What happened?" A knowledge graph says a customer placed an order; a context graph (also a knowledge graph) says why the order was approved despite violating standard terms, what precedent existed, who had authority, and what conditions justified the deviation. Examples: a VP approving a discount beyond policy, a technician modifying a procedure; today the reasoning is not recorded.

- **Procedural Knowledge Ontology (PKO)**, developed by researchers at Cefriel with industrial partners including Siemens and BOSCH, separates procedures (abstract specifications) from executions (a specific technician performing a safety lockout on a specific date, with observations and outcomes). It covers six areas: procedure specifications, granular action steps, change tracking, execution histories, agent roles and authority, and supporting documentation, creating the audit trails and precedent records agents need for judgment-laden situations.
- **Siemens** modeled microgrid device operations as procedures with states and transitions, turning undocumented controller logic into queryable knowledge. EV owners can see why charging behaves as it does (low photovoltaic production, high grid demand, battery capacity limits) and when optimal conditions occur; operators troubleshoot anomalies with full context.
- The challenge is knowledge management: systematic elicitation of tacit knowledge (observing work, interviewing experts, encoding reasoning). Otherwise decision traces stay trapped in Slack threads, email and institutional memory. Organizations that treat context capture as a core competency gain the ability to explain why something was allowed to happen.

Suggested reads on the page: [Context Graphs Are a Trillion-Dollar Opportunity. But Who Actually Captures It?](https://metadataweekly.substack.com/p/context-graphs-are-a-trillion-dollar) (Prukalpa) and [Context Graphs and Process Knowledge](https://jessicatalisman.substack.com/p/context-graphs-and-process-knowledge) (Jessica Talisman).

## The YAML problem

"We tried to encode business meaning in YAML files." dbt's MetricFlow "uses information from semantic model and metric YAML configurations to construct and run SQL in a user's data platform":

```yaml
semantic_models:
  - name: orders
    defaults:
      agg_time_dimension: order_date
    entities:
      - name: order_id
        type: primary
    measures:
      - name: order_total
        agg: sum
```

Good for metric governance, but a model of calculations over tables and columns: no relationships beyond join paths, no natural language definitions, no business processes. The same domain as an ontology:

```turtle
EXPLICIT:
:Alice :placedOrder :Order001 .
:Order001 :hasItem :Item001 .
:Item001 :itemProduct :Laptop .
:Alice :customerLifetimeValue "1250.00"^^xsd:decimal .
INFERRED (what you get automatically):
:Alice a :HighValueCustomer .           # Reasoner infers this
:Alice :purchased :Laptop .             # Property chain inference
:Order001 :orderedBy :Alice .           # Inverse property
:Laptop :frequentlyBoughtWith :Mouse .  # Symmetric property
:Alice :knowsCustomer :Carol .          # Transitive reasoning
```

An ontology can state that an order is a type of business transaction placed by a customer (a type of person), distinct from quotes and invoices, with logical assertions that support inference. The "semantic layer" metadata was structural, not semantic. Ontology construction needs domain expertise and formal knowledge engineering skills most data teams lack; **metric definition and ontology engineering are fundamentally different disciplines** (not a criticism of data teams).

## Why AI changes everything

In October 2025 at Coalesce, dbt Labs open-sourced MetricFlow under Apache 2.0, saying "the semantic layer is the critical component to build a bridge between AI and structured data." Right problem, but the solution may need more than that architecture offers. **LLMs need context and meaning, not dashboards**: what things are, how they relate, what actions are possible. A semantic layer is for lookup; an ontology is for context and reasoning.

## The context requirement

[Anthropic's engineering team](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents) defines context engineering as designing a system that provides the right information and tools, in the right format, for an LLM to accomplish a task, including prompts, memories, few-shot examples and tool descriptions. A semantic layer cannot provide that wholesale. Knowledge graphs and ontologies represent concepts, relationships and constraints an LLM can query. That is why healthcare AI relies on clinical ontologies, drug discovery platforms run on biomedical knowledge graphs, and enterprise AI is adopting semantic technologies; domains that invested in formal knowledge representation are where AI reasoning works at scale.

## The real architectural difference

| | Semantic layers | Ontologies and knowledge graphs |
|---|---|---|
| Designed for | Humans consuming data through BI tools | System and domain understanding |
| Provide | Metric definitions and calculations, dimensional models, SQL abstraction, consistency across reporting tools | Formal concept definitions and taxonomies, typed relationships with meaning, logical inference, interoperability via RDF, OWL, SKOS, domain expertise in machine-readable form |
| Answers | Metrics questions: "What is X?" | Context without SQL and table constraints |
| Origin | BI industry's focus on consistent measurement; solution was centralized metric governance | Knowledge management, library science and AI research (a form of neuro-symbolic AI); solution was formal knowledge representation |

## The hard questions for 2026

- **Is there a middle ground?** "Context-aware semantic layers" that keep metric governance and add ontological richness. The Open Semantic Interchange initiative, which dbt Labs joined with Snowflake and Salesforce in 2025, hints at this, but the gap may be too wide to bridge quickly.
- **Can semantic layers evolve into knowledge graphs?** Adding classes, properties and inference rules to a system built for SQL generation may amount to building a knowledge graph from scratch.
- **Do we need a new category?** "[Context engineering](https://atlan.com/know/context-engineering-for-ai-analyst/)" has traction; Tobi Luetke called it "the art of providing all the context for the task to be plausibly solvable by the LLM." But it is a practice, not a platform. A "ContextOS" would likely combine metric governance with knowledge representation, easier to name than build.
- **Where does dbt's semantic layer fit?** dbt now describes it as enabling "AI agents to leverage trusted metric definitions for governed conversational analytics." Is governed [conversational analytics](https://atlan.com/know/regovern-conversational-analytics-guide/) enough, or do agents need ontology-level domain knowledge?
- **Is the metrics-first paradigm dead?** AI needs concepts, relationships and inference, not charts. Metrics may become one input to a richer knowledge architecture.

## But haven't we heard this before?

The Semantic Web was supposed to revolutionize the internet 20 years ago, and most organizations that try ontologies fail or abandon them.

- **Skeptical case**: ontology construction needs rare skills (knowledge engineering, formal logic), years of investment and sustained commitment; maybe these are organizational problems, not technical ones.
- **Counter-argument**: AI changes the stakes. Healthcare and life sciences adopted ontologies because their work required it; Gene Ontology has lasted 20+ years because the alternative does not work. Ontologies work where organizations invest; the question is whether yours will build one.

## What this means, practically

- **Building an AI analyst**: you need knowledge representation, not just metrics. With only metric definitions (no concept hierarchies, relationship types or domain knowledge) the AI answers calculation questions, not reasoning or inference questions.
- **Evaluating semantic layers**: ask "Is this for humans or for AI?" Most organizations need both, possibly with different architectures.
- **Considering knowledge graphs or ontologies**: expect a different investment: domain expertise, knowledge management and engineering skills, often years of refinement. Successful organizations (life sciences, healthcare, finance) treat knowledge representation as a core competency, not a project.

Consistent metrics are table stakes; the advantage goes to architectures that support AI reasoning about concepts, relationships and domain knowledge.

## Where do we go from here?

The semantic layer vs ontology framing may be too narrow. For metric lookup, calculation consistency and basic analytics, semantic layers remain sufficient; for complex reasoning, inference, elicitation of meaning and domain-specific AI, they will not be. Whatever it is called (semantic layers 2.0, knowledge graphs, context engineering, context graphs), the era of metrics-based thinking is closing. **It's a knowledge architecture problem**, a different discipline closer to what librarians, taxonomists and knowledge engineers have done for decades. Healthcare and life sciences made their bet 20 years ago; Palantir built its closed ontology-ish system in 2012. The question is whether you start now or let competitors take a three-year head start.

The issue promoted [The Great Data Debate](https://atlan.com/great-data-debate-2026/) with Jaya Gupta (Foundation Capital), Karthik Ravindran (Microsoft), Bob Muglia (Snowflake) and Tony Gentilcore (Glean), debating whether semantic layers can evolve, whether enterprises need ontologies, and who captures the context graph opportunity.

## References

1. First Round Review, "The Inside Story of How This Startup Turned a 216-Word Pitch Email into a $2.6 Billion Acquisition," October 2024.
2. dbt Labs, "Announcing open source MetricFlow: Governed metrics to power trustworthy AI and agents," October 14, 2025.
3. dbt Labs, "dbt Labs Affirms Commitment to Open Semantic Interchange by Open Sourcing MetricFlow," press release, October 14, 2025.
4. Anthropic, ["Effective context engineering for AI agents"](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents), Anthropic Engineering Blog.
5. LangChain, "Context Engineering for Agents," October 19, 2025.
6. Google Developers Blog, "Architecting efficient context-aware multi-agent framework for production," December 4, 2025.
7. Gene Ontology Consortium, http://geneontology.org/
8. SNOMED International, https://www.snomed.org/
9. W3C, ["OWL 2 Web Ontology Language"](https://www.w3.org/TR/owl2-overview/).
10. typedef.ai, "Semantic Layer 2025: MetricFlow vs Snowflake vs Databricks," November 2025.

## The Insight Index (recommended reads)

- [Context Graphs, Data Traces & Transcripts](https://ontologist.substack.com/p/context-graphs-data-traces-and-transcripts) - Kurt Cagle
- [Ontologies - Some Perspectives](https://williaminmon.substack.com/p/ontologies-some-perspectives) - William Inmon & Jessica Talisman
- [Semantically Speaking: What Context Graphs Made Impossible to Ignore](https://www.linkedin.com/pulse/semantically-speaking-what-context-graphs-made-ignore-j-bittner-azmte) - J Bittner & Colbie Reed
- [How Context Graphs Turn Agent Traces Into Durable Business Assets](https://www.linkedin.com/pulse/how-context-graphs-turn-agent-traces-durable-business-dhinakaran-ngptf) - Aparna Dhinakaran
- [Context Graphs: The Elegant Idea Everyone's Talking About](https://simple.ai/p/what-are-context-graphs) - Dharmesh Shah
- [How to build a context graph](https://www.linkedin.com/pulse/how-build-context-graph-animesh-koratana-6abve/) - Animesh Koratana
- [Decision Traces Are Only as Good as the Context That Fed Them](https://www.linkedin.com/pulse/decision-traces-only-good-context-fed-them-shirshanka-das-fk1fc) - Shirshanka Das
- [Context Graphs: AI's Next Big Idea](https://www.youtube.com/watch?v=SVUymPVBvfo) - The AI Daily Brief

The list does not represent the views of the author or the community.

## Related reads

- [Context Graphs Are a Trillion-Dollar Opportunity. But Who Captures It?](https://atlan.com/context-and-chaos/issue/context-graphs-are-a-trillion-dollar-opportunity-but-who-captures-it/)
- [Semantic Layers Failed. Context Graphs Are Next.](https://atlan.com/context-and-chaos/issue/semantic-layers-failed-context-graphs-are-next/)
- [Conceptual Modeling Is the Context Engineering Nobody Is Doing](https://atlan.com/context-and-chaos/issue/conceptual-modeling-is-the-context-engineering-nobody-is-doing/)
- [Context Graphs as AI Evaluation Infrastructure](https://atlan.com/context-and-chaos/issue/context-graphs-as-ai-evaluation-infrastructure/)

Context & Chaos is a community newsletter on context engineering, governance, architecture and discovery. All issues: https://atlan.com/context-and-chaos/