What Palantir Got Right With Ontologies
Palantir put the ontology — not the LLM — at the centre of enterprise AI, and that architectural bet is being proven right. But AI itself has now collapsed the cost of building an ontology. You no longer need an army of forward-deployed engineers, and you certainly don't need to hand your most sensitive data to a US vendor to get one.
We're not shy about our differences with Palantir. We've built a sovereign alternative precisely because we believe European and regulated organisations deserve better than proprietary formats, premium lock-in, and dependency on a US vendor.
But credit where credit is due: Palantir made one architectural bet, years before it was fashionable, that the rest of the industry is only now catching up to. They put the ontology at the centre — and the model at the edge.
In an excellent deep-dive, Palantir's 12-layer agentic architecture, Anand dissects how Palantir's AIP actually works, and lands on the insight that matters most:
"The Ontology is the agent's brain, not the LLM."
While most vendors bolt an LLM onto an existing data stack and call it enterprise AI, Palantir inverted the architecture. The language model is a replaceable reasoning component. The durable intelligence — what the business is, how it works, what happened last time — lives in the ontology: a semantic model of your real business entities, their relationships, and the rules that govern them.
That inversion is exactly right. And it's the same foundation our entire Cognitive Enterprise is built on.
Why the Ontology Is the Brain
Every impressive agent demo eventually collides with the same wall: the model doesn't know your business. It doesn't know that "the Antwerp contract" means account 4711 with a renegotiated liability clause, that invoices above €50k need a second approval, or that the last three times this alert fired it was a sensor fault.
The common workaround is to stuff ever more context into the prompt window. As Anand puts it:
"Every agentic AI system needs memory. Most platforms solve this by stuffing context into an LLM's prompt window — a fundamentally limited approach."
Prompt-stuffing (and naive RAG on top of a document dump) treats business knowledge as text to be retrieved. An ontology treats it as structure to be reasoned over: entities, relationships, state, permissions, and history. The difference shows up immediately in accuracy, explainability, and hallucination rates — we've written before about why an ontology beats traditional RAG as grounding for agents.
There's a deeper point here. LLMs are becoming commodities — interchangeable, rapidly improving, and largely undifferentiated within a capability tier. Your ontology is the opposite: it encodes how your organisation actually operates, and it compounds in value with every process that runs through it. Betting your AI strategy on the model is betting on the part everyone has. Betting it on the ontology is betting on the part only you have.
Palantir's Own CEO Is Now Making This Argument — Loudly
In early July 2026, Palantir CEO Alex Karp put the same point rather more bluntly. Enterprises buying metered access to frontier models, he argued, are quietly handing over the very thing that makes them valuable:
"I am paying for tokens that create no value. These people are stealing the weights and alpha of my business."
His prescription is that companies "want to know they own the means of production" rather than "deploying tokens that transfers the alpha to a third party." Strip away the theatrics and it is the ontology-first thesis stated as a warning: if your organisation's intelligence lives inside a rented model, your edge leaks to whoever controls the weights. If it lives in your ontology, it stays yours — governed, auditable, and compounding on your side of the line.
We don't agree with Palantir on much, but on this we do. Where we part ways is what "owning the means of production" actually requires — which brings us to what has changed since Palantir built its business.
Four Kinds of Memory, One Living Model
Anand highlights how Palantir maps the ontology onto the four memory types cognitive science says any intelligent actor needs:
- Working memory — the immediate task context: which order, which customer, which step of the process the agent is in right now.
- Episodic memory — what happened before: previous runs of this process, past decisions and their outcomes, with temporal markers.
- Semantic memory — what things mean: the entities, relationships, and business concepts that turn raw data into knowledge.
- Procedural memory — how things are done: the encoded rules, workflows, and logic that don't need to be re-derived by a model on every run.
"By using the Ontology as a unified memory system, AIP ensures a consistent surface area across all four memory types for both human and agentic users."
This is precisely how our living ontology works, and why we treat agent memory as an ontology concern rather than a vector-database bolt-on. Every interaction — by a system, an agent, or a human — is written back as structured knowledge linked to real business entities. Working context is assembled from the graph, episodes accumulate as history on the entities they touched, semantics are the graph itself, and procedures are executable, versioned, and governed.
The result: knowledge accumulates instead of evaporating after each session. The thousandth run of a process is better informed than the first.
From Ontology to Decision Intelligence
Here is where the ontology stops being a data-management story and becomes a business capability.
Every meaningful decision in your organisation — approve this claim, reroute this shipment, escalate this incident, extend this credit line — depends on context that is scattered across systems and heads. An ontology pulls that context into one governed, queryable model. That is the raw material of decision intelligence: the discipline of making decisions explicit, connected to the data and rules they depend on, and improvable over time.
With decisions anchored to the ontology, three things change:
- Decisions become explainable. Every recommendation traces back through the graph to the entities, rules, and history that produced it — for the regulator, the auditor, and your own operators alike.
- Decisions become consistent. The same rules and the same context apply whether the actor is a human on Tuesday or an agent at 3 a.m. on Sunday.
- Decisions become data. Each outcome is written back into the ontology, so the organisation learns from every decision it makes. Judgment stops being trapped in individual heads and starts compounding as an institutional asset.
This is why we say the ontology is the basis of decision intelligence: without a shared semantic model, "decision intelligence" collapses into dashboards and gut feel. With one, it becomes an operating capability.
And Decision Intelligence Needs an Operating System
An ontology that only describes the business is a very good map. The value multiplies when it starts to drive the business — when decisions flow into actions on real systems, executed by agents, humans, and existing software working together.
That requires an orchestration layer: something that routes each step to the right actor, supplies the right slice of ontology context, enforces policy, and writes every outcome back. That layer is our AI OS. Our Agentic AI engine grounds purpose-built agents in the ontology with identity, scoped permissions, and full audit trails, while the AI OS composes them — alongside your people and your existing systems — into governed, continuously improving processes.
The AI OS for Humans & AI Agents
Ontology & Semantic Layer, one connected model for your data, knowledge & processes
Sovereign Foundations
Ontology → decision intelligence → AI OS. Each layer makes the next one possible. Palantir understood this stack early, and deserves the credit for proving it works in production, at scale, in some of the most demanding environments on earth.
What's Changed: You No Longer Need the Army
So why not just buy Palantir? Two things.
First, AI collapsed the cost of building an ontology. Palantir's model was forged in a pre-LLM world: multi-year engagements where forward-deployed engineers hand-modelled your business, entity by entity, at a price only governments and Fortune 500s could justify. That was the price of ontologies then. It isn't anymore.
Modern AI is remarkably good at exactly the work that used to require the army: profiling source systems, proposing entity models, extracting relationships and rules from documents and process traces, and keeping the model current as the business changes. Our platform is AI-native in precisely this sense — AI helps build the ontology, agents run on it, and every execution enriches it. What took a deployed team eighteen months is now a matter of weeks, with your own domain experts validating the model instead of narrating their jobs to consultants. The moat was never the ability to model a business; it was the cost of doing so. That moat has drained.
Second, sovereignty. An ontology is, by construction, the most concentrated representation of your organisation that will ever exist — every entity, every relationship, every decision, every rule, in one place. Now ask the uncomfortable question: do you really want to hand that to a US vendor, on their terms, in their proprietary formats, under their jurisdiction?
Palantir clearly senses this too: its June 2026 "Sovereign AI" reference architecture with Nvidia — open models running in air-gapped environments on top of AIP — is a direct answer to exactly the anxiety Karp is voicing. But notice how far his own logic runs. If the point is to own the means of production and not "outsource the battlefield of this country to the consensus view in Silicon Valley," then for a European government or bank, a sovereign environment sold and operated by a US vendor, on proprietary formats, under US jurisdiction, is only half the argument answered. Owning the means of production means owning them.
For European governments, critical infrastructure, defence, and regulated industries, that is the real bar. Our platform delivers the same ontology-first architecture on open table formats such as Apache Iceberg, running entirely inside your own perimeter — from a fully air-gapped on-premises cluster to hyperscale cloud — with your keys and your roadmap. You keep the idea Palantir got right, you take Karp's own advice to its conclusion, and you keep control of everything your ontology knows.
The Takeaway
Palantir got the architecture right: the ontology is the brain, the model is a component, and decision intelligence emerges from grounding every actor — human or agent — in a shared, living model of the business.
What they got right about architecture, though, no longer implies what it used to about delivery. AI has lowered the barrier to a decent ontology from an army of forward-deployed engineers to a platform capability. And the more valuable your ontology becomes, the less acceptable it is for it to live in someone else's perimeter.
The right idea, sovereign, and finally affordable. That's the Cognitive Enterprise.
Email us
Prefer to write? Email hello [at] scrydon.com and we will get back to you.