Key points
- Core thesis: Palantir Ontology is not a data model, not a knowledge graph, and not a semantic layer. It is a decision operating system that fuses four dimensions at runtime: Data (nouns), Logic, Action (verbs), and Security. It models *how an enterprise decides and how decisions take effect*, not merely *what data the enterprise has*.
- Authoritative definition: Quoted from Palantir's Architecture Center: *"The Ontology is designed to represent the complex, interconnected decisions of an enterprise, not simply the data."*
- Three-stage evolution:
- Gotham (2003–2015): entity resolution and object-level security for defense/intel.
- Foundry (2016–2022): enterprise operational layer with Action Types, Functions, and dynamic security.
- AIP (2023–): orchestration of human-AI hybrid workforces; introduces decision lineage and Ontology MCP.
- Architecture primitives:
- Object Type = schema for real-world entities (analogous to a dataset's table).
- Link Type = bidirectional relationship between object types (no need for reverse-link definition).
- Action Type = governed write schema with parameters, rules, submission criteria, and side effects; this is the dividing line between Ontology and a pure data model.
- Decoupling and bi-directional binding: Ontology separates metadata (OMS) from instance data; supports virtual tables over Snowflake/Databricks/BigQuery/S3-Iceberg without copying. Actions expose two webhook types — writeback (pre-change, transactional-style) and side-effect (post-change). Palantir itself acknowledges this is not true two-phase commit.
- Key disambiguations:
- vs. Knowledge Graph: KG aims at discovery under open-world assumptions; Ontology uses closed-world managed schema with emphasis on operations.
- vs. Semantic Layer (dbt MetricFlow, Cube): semantic layers unify metric definitions and are read-only; Ontology unifies *how decisions are made* and can write back to ERPs.
- vs. Data Mesh: complementary, not competing; Data Mesh is organizational, Ontology is structural.
- vs. Graph databases: Ontology's "graph" is logical; the backend is OMS/OSv2/OSS microservices plus a lakehouse and search indexes.
- AIP era design: LLMs never access tools directly or hold tools. They emit intent; AIP Logic executes tool calls under the invoking user's runtime permissions. OSDK generates typed SDKs; Ontology MCP (OMCP) exposes ontology resources as MCP tools to external agents such as Claude Code and Copilot Studio, inheriting the same security policies.
- OAG vs. RAG: Palantir positions Ontology-Augmented Generation as superior for operational decisions involving live state, deterministic logic, and writeback. However, Palantir's own "Ontology-augmented generation" documentation page is largely classical RAG (chunking, embeddings, hybrid search, RRF) and states there is no singular best approach. OAG does not eliminate hallucination — it constrains blast radius via Action gates.
- Case studies with caveats:
- Ukraine: Palantir signed a Gotham agreement in 2022; Karp claims responsibility for "most target定位". Ukrainian officials note Delta excels at data collection while Palantir excels at visualization/semantic integration.
- Maven Smart System: sole-source vendor since 2023-05; ceiling over $1B.
- NHS Federated Data Platform: £330M / up to 7 years. Health Foundation analysis found trusts using the tool showed no significant improvement; the UK Statistics Authority intervened in 2026-07 and NHSE agreed to add caveats about causal claims.
- Vendor lock-in — hard evidence:
- NYPD (2017): Palantir refused to deliver analytical outputs in a portable format, claiming the organization/connections/visualization structure was its IP.
- UK MoD Hansard (2026-02-10): a minister acknowledged services can be run only by Palantir, with significant switching cost.
- Counter-examples include French DGSI switching to ChapsVision and Swiss Army rejecting Palantir.
- When NOT to use Ontology: weak data foundations, undefined decision patterns, pure reporting needs, rapidly shifting business models, and sovereignty-sensitive scenarios without viable exit paths.
- Common circulated metrics are unverified: claims such as "match-rate 35%→4%", "92% fault-prediction accuracy", and various inventory/throughput percentages have no primary source and appear to be marketing or AI-generated.
Bottom line
Ontology's true differentiator is governed write-back (Action Types + writeback webhooks), not its data-modeling surface. The same property that makes it enterprise-grade — runtime permission enforcement and decision lineage — is also the source of vendor lock-in. Treat published case-study percentages skeptically unless independently verified, and treat "OAG" as Palantir's marketing framing rather than an academic term.