Built to act.
For the better part of three decades, the data layer of institutional capital was built around one consumer: the human reading a report. The consumer is changing.
This brief reads the shift in two layers. The mental model you can hold in your head, three tiers with two operations between them. The structural lens beneath it, the layered architecture that codifies what the human used to perform invisibly. Same reorganization at different altitudes.
Three tiers, one question. Can you reconstruct every decision?
Every institution already carries a knowledge graph: entities and the relationships between them, scattered across systems and never named as such. Entity resolution is the act of applying that graph across three dimensions: who (legal entities: funds, LPs, GPs, SPVs, operating companies), how (structural exposure: private and public debt, equity, secondaries, joint ventures, club structures, co-investments), and when (vintages, hold periods, restructurings, secondary transactions). Once the knowledge graph is resolved across these three dimensions, it becomes a context graph: relationships are queryable as of any point in time, under any structural framing.
Layer three streams of metadata onto that context: how data moves, how meaning is governed, how decisions are made. The context graph becomes a decision graph. Queryable institutional memory of why every decision made sense at the moment it was made.
The Decision Graph
The Context Graph
The Knowledge Graph
What the architecture has to codify when the human stops being the consumer.
The reporting stack carried more weight than its diagrams ever showed. The human, reading the output, did the work the architecture never had to: reconciling the same entity across sources, interpreting what definitions meant, holding situational context, remembering why decisions were made, composing all of it into reasoning. The architecture did not install any of that. The human was the integrating layer. The semantic truth arbitrator. The final say on what the data meant.
The agentic stack does not have a human in that position. The cognitive functions persist. They have to, or no work gets done. The architecture has to perform them. The substrate stays. The transport changes shape. The cognitive work, which always existed, becomes explicit as a set of layers.
semantic truth arbitrator Mapping, understanding, interpretation, context, judgment, reasoning. All happening invisibly inside one person reading the output. The architecture never had to install any of it.
Persistent 09 Action surface +
Was "Consumption." Same layer; the consumer expands from human to human + agent.
Consumers are display-shaped. Architecture assumes a human reading a chart; agents need graph traversal, semantic queries, lineage retrieval.
·
Codified · reasoning 08 Reasoning surface +
Was implicit. The human composed across entity, relationship, definition, context, and memory to produce reasoning.
AI is bolted on. Models wrap around the dashboard. They see the outputs the human reads, not the substrate that produced them.
A reasoning surface agents can rely on. A stable interface that survives schema, vendor, and platform churn downstream.
Codified · judgment 07 Memory layer +
Was implicit. The human remembered what was decided last quarter and why it made sense.
No decision lineage. Approvals, overrides, sign-offs live in emails, chat, IC minutes. None of it structured or queryable.
A decision layer that is queryable. Capturing decisions as events with standard schemas is tractable. The reason it isn't solved is not technical. Operational and political.
Codified · context 06 Context graph +
Was implicit. The human held situational context: what was active when this happened, what definition was in force.
Context is reconstructed at query time. No native operation joins knowledge, events, and semantics; the work falls to whichever consumer asks the question.
A context graph that operates at population scale. Each firm's context graph stops at the firm's walls. The institutional memory worth having extends beyond any single firm.
Codified · interpretation 05 Semantic layer +
Was scattered. Definitions lived in modeling code, BI tools, analyst heads, IC memos.
Semantic logic is scattered and unversioned. Same metric, multiple implementations. A definition active a year ago is not retrievable today.
A semantic registry that survives versioning. Today's semantic layers do not survive across tools, let alone across firms.
Codified · understanding 04 Knowledge graph +
Was the human's mental picture of how entities related.
The data is not graph-shaped. Real estate is a graph (entities and relationships) stored as tables. Multi-hop reasoning becomes expensive joins.
·
Codified · mapping 03 Entity resolution +
Was implicit. The human reconciled "this building" across sources at query time.
No persistent canonical identity. Entity resolution happens at the BI or modeling layer, ad hoc, query by query. The same property has multiple IDs.
A canonical identity substrate every consumer trusts. Each firm builds its own, often poorly. No shared substrate exists at population scale.
Transformed 02 Ingestion + submission +
Was "Movement": ETL with semantic arbitration hidden inside it. The semantics extract to the codified layers above; the mechanical transport stays.
No event store. State changes are overwrites. The audit trail is "what was the value at month-end snapshot," not "what happened, when, why, in what order."
Schema-validated, append-only ingestion with row-level provenance. The layer as a contract surface, not a connector library.
Persistent 01 Systems of record +
Unchanged. Property management, fund accounting, valuation systems. The authoritative ledger for transactions in both eras.
·
·
Graph-native, not graph-bolted.
Storage shape matches data shape. Bolt-ons rarely perform; designed-in does.
Append-only, not overwrite.
Past state is reconstructible. Snapshot databases destroy what event-sourced ones preserve.
Identity is foundational.
Built first, the rest can be installed. Retrofitted after, every layer above absorbs the cost.
Meaning is versioned.
Definitions change. The architecture remembers which version was active when.
Agents are the design target.
Designed for the agent, the human is served. The reverse does not hold.
Same scale. Different wiring.
The Romans and the Mongols both built tiered hierarchies that scaled to hundreds of thousands. Both enforced discipline. Both conquered at continental reach. The difference was not capability, ambition, or scale. It was constituted at the base unit and compounded upward.
Authority never reaches the unit
Authority lives at the unit
Signal arrives at the soldiers. The Centurion must receive it, decide, and issue the order before anyone moves.
10 horsemen. No intermediary. The Arban carries its authority.
Each new AI capability requires the same approval chain. Each new domain routes through the same central team. For the Romans, orders took days or weeks. For the COE, every output waits for the review queue. The architecture of centralized approval compounds into the architecture of delay.
Each new pod carries the same governing principles into a new domain. No additional approval chain required. The architecture of distributed authority scales because the principle, not the approver, travels with the unit. The Arban acted in moments. The pod delivers in its domain without waiting.
The same logic applies to how firms deploy AI. Most AI functions are built the Center of Excellence way. A central team holds the capability. Domain teams route requests up. Outputs route back down for approval. The COE is the centurion. The practitioners are the soldiers. The governance structure is not the problem; it exists for good reason. What is built the Roman way is the delivery unit serving it.
Every AI output routes up for approval before it reaches the decision-maker. The governance structure waits for the COE to clear it.
Context stays local. Governing principles travel laterally. Every unit operates with the same framework and full latitude to act within it. Coordination is peer-to-peer. Authority does not need to flow to the top before a decision can flow back down.
Centers of Excellence were right for the reporting era. They centralized capability and enforced standards. Operating pods are right for the reasoning era. They distribute authority to the unit doing the work, governed by principles rather than approval chains. Better work reaches the decision-makers. Nothing about who decides has changed.