Skip to content
Ontology Isn't a Semantic Layer: The Action Layer AI Agents Need

Ontology Isn't a Semantic Layer: The Action Layer AI Agents Need

Palantir argues ontology differs from a semantic layer by modeling both data meaning and actionable operations — actions, functions, and security — which becomes critical when AI agents must write data, not just read it; the article contrasts vendor approaches and gives four design principles for agent-ready ontologies.

AI Large-Model Wave and Transformation Guide
AI Large-Model Wave and Transformation Guide
AI Large-Model Wave and Transformation Guide

The Semantic Layer Looks Familiar

Palantir's ontology overview splits ontology into "semantic elements (object types, properties, link types)" and "action elements (action types, functions, dynamic security)." The semantic elements map closely to what any analytics engineer already knows:

Object types are "schema definitions of real-world entities or events," e.g., Employee or Flight. In a semantic layer, that's an entity or logical table.

Properties are characteristics of an object type, e.g., a flight's departure time. That's a column or a dimension.

Link types are "schema definitions of relationships between two object types." That's a declared join path.

Compared with a semantic layer, this part lacks a focus on metrics. A semantic layer organizes around measures — ARR, churn, net revenue — compiled into the same SQL for every tool. An ontology organizes around the objects that people and software operate on. You can compute on top of it, but the center of gravity is the object, not the aggregate.

The Action Layer Is the Real Differentiator

An action type is "a definition of a set of changes or edits to objects, property values, and links." Functions let developers write business logic in TypeScript or Python that reads properties, traverses links, and makes edits. Dynamic security coordinates permissions at the moment a person or thing interacts with an object.

Palantir's architecture page states this ambition explicitly: the ontology "is designed to represent the complex, interconnected decisions of an enterprise, not just its data," and models those decisions through "the quadruple integration of data, logic, action, and security." Actions range from simple transactions to multi-step updates and must "write back to operational and edge systems in real time." The logic behind an action can be a business rule, a machine-learning model, or an "LLM-driven function."

That is why Palantir says this job cannot be done by "a thin semantic layer." A model that only describes nouns leaves every decision to happen elsewhere. The ontology adds verbs.

Agents Make the Action Layer Important

For dashboards, the action layer is irrelevant — dashboards are read-only. For an analytical agent that answers "What was last quarter's ARR?", it is also largely irrelevant. That agent needs the semantic layer plus governed metric definitions — exactly what a semantic layer built for AI provides.

Once an agent is expected to do something — apply credit, reassign accounts, pause a shipment — the question changes. It is no longer "Does the agent know what the data means?" but "What can it change, through which path, with whose permissions?" Palantir's answer is to route changes through the ontology's actions, which it claims lets "humans and AI agents" collaborate on the same operational workflow. It also states that as teams build AI agents, their security scope should inherit from human users or follow the permissions of defined projects.

There is independent evidence that write paths matter even for agents that should be read-only. A 2026 ablation study on the DABStep benchmark compared governed and ungoverned agents on the same tasks. The governed agents emitted zero data-modification statements. The ungoverned agents emitted 166.

Author's view: The action layer is less a Palantir feature than a design rule. Agents should read through governed definitions and write through named, permissioned operations — never through free-form SQL.

Most 2025–2026 Ontology Waves Are Only the Semantic Part

"Ontology" has suddenly appeared in every data platform's vocabulary. But reading the documentation closely, most new entrants describe something closer to the semantic layer, packaged as agent context:

Microsoft Fabric IQ ontology items "provide a common context layer." They hold entity types, relationships, natural-language rules, and metrics inherited from Power BI DAX measures. Microsoft explicitly says it "is not a general-purpose data query engine": agents use it as context and execute against bound data sources. It is still in preview.

Databricks titles its explainer "The Context Layer Your AI Agents Are Missing." Genie Ontology, released June 2026 with Genie One, is "a new self-improving context layer" that extracts business knowledge from tools like Slack, Jira, and Confluence. Databricks reports that Genie with Ontology answered 84.5% of 28 enterprise questions correctly on the first try, versus 52.4% for the strongest general-purpose coding agent they tested, calling this an internal benchmark.

Snowflake avoids the term entirely. Horizon Context is marketed as a governed context layer, with Semantic Views as its semantic component.

None of this is criticism. For analytical problems, the semantic layer is the part that lifts accuracy. But it does mean "we have an ontology now" is not the same claim Palantir makes. When a vendor says ontology, one question sorts them quickly: Can an agent take a verified, permissioned action through it, or can it only read?

What to Take Away If You're Not on Foundry

You don't need Palantir to apply this idea. You need to separate the two parts and build them in order.

Build the semantic layer first, and keep it small. Describe entities in business language, declare keys as relationships, give canonical definitions to the metrics people argue about, and define allowed values for status-type columns. This operational minimum is covered in more depth in our data ontology guide, and it does not require OWL.

Keep the numbers compiled. Object models are weak on "what is the number?" Board-level metrics belong in a semantic or metric layer that compiles to the same SQL every time.

Treat writes as a separate product. If agents will act, define every action explicitly: its inputs, validation, who can run it, and where it writes. Do not hand an agent write permission and a prompt.

Enforce permissions outside the model. Whatever you build, an agent should inherit a person's or project's permissions at the moment of action, not decide permissions for itself.

Palantir is right that its ontology is not a semantic layer. The more useful reading is the reverse: a semantic layer was never the whole story for agents. It is the read path. If your agents are going to act, you need to design the write path with the same deliberation.

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

AI agentsData ModelingSemantic LayerontologyVendor ComparisonAgent SecurityAction LayerPalantir Foundry
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.