Ontology is the new schema

The relational schema was the great contract of the software era. It answers structural questions with total precision: which tables exist, which columns they hold, which keys join them. It answers nothing else. A field reading status = 3 is valid in every database on earth and meaningful in none of them.

For thirty years that was acceptable, because humans supplied the meaning. The analyst knew that status 3 meant held for fraud review. The operations lead knew a shipment could not be cancelled after the carrier scan. The meaning of the business lived in people's heads, onboarding decks, and wiki pages nobody maintains. Software did not need to understand the business, because software never acted alone. A person was always in the loop, carrying the semantics.

A business is not its tables. It is entities — customers, orders, shipments, claims — bound by relationships and moved through state machines by actions. Cancel the order is not a row update. It is an operation with preconditions, side effects, a refund policy, and an owner. None of that appears in the schema.

This is why new hires take months to become useful even when they can read SQL on day one. What they are learning is the ontology: what things mean, how they connect, and what they are allowed to do. Every enterprise already has an ontology. Almost none have written it down.

An agent wired to raw database access has intelligence and no world model. To act safely it needs three things a schema never supplies. First, semantics: what the entities are and how they relate, so cancel the order refers to a real operation and not a token pattern. Second, a typed action interface: a catalogue of operations it may invoke, each with preconditions, effects, cost, and reversibility, so capability is granted rather than improvised. Third, permissions with provenance: whom the agent may act as, on what, within which limits — with a decision trail for the day audit asks why.

This is the core lesson Palantir spent two decades proving in the field: integrate meaning, not just data. Model the business objects and actions once, and every later application — human-facing or agentic — inherits them. Agents do not raise the bar for data integration. They raise the bar for meaning integration.

The strongest counterargument: we tried this. Enterprise ontologies, master data management, the semantic web — the 2000s are littered with meaning projects that collapsed under their own ceremony. Fair, and mostly true.

But those projects failed for a specific reason: they built meaning nobody consumed. No software existed that could read an ontology and do useful work with it, so the artifact rotted the day the consultants left. That constraint is gone. Agents are the consuming software, and they consume ontology hungrily — every entity you define is a capability they gain. The second failure mode, totality, still applies: an ontology of everything ships nothing. The fix is scope. One workflow, one thin slice, consumed by a working agent within weeks — then widen.