The Institutional Model

General intelligence is becoming a commodity. Specific intelligence: how one company thinks, decides, and evolves, has never been held by any system. Our research is about modeling it.

The reading problem

Enterprise AI fails for a reason nobody wants to say out loud: the model has read the internet, but it has never read the company. It does not know why the pricing decision went the way it did, which of the three “official” processes people actually follow, or what the leadership team has quietly stopped believing since last quarter. It sees the artifacts of work, documents, tickets, dashboards, and none of the thinking that produced them.

The industry’s answer has been retrieval: connect the model to the documents and let it look things up. Our position is that this cannot work, even in principle. A company is not a corpus. It is a system with intent, expertise distributed across thousands of people, and a direction that shifts as the market moves. You cannot retrieve your way to understanding a thing like that. You have to model it.

Four problems

The four problems that stand between today’s AI and an organization it can actually understand. Everything in our work is organized around them.

  • Representation
  • Synthesis
  • Drift
  • Execution

Positions

No single representation is sufficient.

Vector retrieval flattens structure. Knowledge graphs freeze it. Fine tuning bakes in a snapshot that is stale by the next reorganization. Each holds one projection of an organization and loses the rest. Our work indicates that organizational understanding requires multiple representations operating together, layers that each hold what the others structurally cannot, with a synthesis architecture between them. The details of that architecture are the core of our unpublished work.

Organizations are moving targets.

A company learns from every market event, every failed initiative, every hire. Its beliefs at month twelve are not its beliefs at month one, and most of that change never gets written down. Any system trained on a snapshot is modeling a company that no longer exists. We treat the organization as always changing: the Institutional Model is built to move with the company, not to describe it once.

Memory does not aggregate.

Giving an AI memory works for one person. It breaks for ten thousand. A company’s understanding is not the sum of its employees’ memories. It is what emerges when expertise at every level, from operator to boardroom, is reconciled into a shared direction. Storing everyone’s context gets you an archive. Synthesizing it gets you an institution. This distinction, storage versus synthesis, is where we believe the entire field is currently stuck.

What we are building

A machine readable organization that maintains context across levels, changes with the institution, and governs how that context reaches execution.

Company Team Personal
Perceive Think Reason Learn Execute
Synthesis substrate unpublished
Institutional learning flywheel

Every governed interaction updates the model available to the next decision.

BYOC Customer controlled data plane
Redacted architecture. The synthesis substrate remains unpublished.

Hindsight. Insight. Foresight.

Three ways an Institutional Model must reason across the organization’s history, present state, and possible direction.

Hindsight

Reconstruct why a decision was made by connecting its artifacts to the people, constraints, assumptions, and prior events that shaped it.

Insight

Synthesize distributed expertise and current operating context into a shared view of what the organization knows now.

Foresight

Test possible direction against institutional memory, current intent, and observed patterns. Predictive institutional reasoning remains an open research problem, not a claim of certainty.

Open problems

How do you evaluate whether a model understands an organization, when the organization itself cannot fully articulate what it knows? How much of strategic intent is recoverable from interaction data alone? At what point does an institutional model become predictive, able to propose direction, not just execute it? These are the questions our next year of work is organized around.

Research partnership

We partner with enterprises who want to be the environment this research is developed in, typically with the CDO, CAIO, or CTO. Research partnership means your organization becomes measurably more AI native while the field’s hardest open problem gets solved inside your walls.

Discuss a research partnership