Company as Intelligence

The thesis

Jack Dorsey argues that a company can be rebuilt with AI as its structural core rather than its hierarchy, with every artifact composing a world model that anyone can query, and humans sitting at the edge to align that model toward customer outcomes. Companies are already intelligences in a functional sense, aggregating information and coordinating action toward goals, but they have been structured around the limits of human-scale information relay: a management hierarchy whose primary function was never power but bandwidth, solving communication across many people before information technology could. AI dissolves that problem. Every artifact a company generates, a chat message, an email, a pull request, code, a document, a recorded meeting, is already a structured data point about what the company is doing, building, failing, and learning, and an intelligence layer built on top of these artifacts creates a world model that anyone can query, from the chief executive to a customer support representative.¹

In this new structure, the AI world model sits at the center and humans sit at the edge, continuously aligning it toward customer outcomes; the chief executive's job shifts from managing the hierarchy to architecting and aligning the intelligence, and the board's job shifts from receiving relayed data to querying the model directly and spending meeting time on genuinely creative or existential decisions instead.

Why this is not just a productivity story

The common frame for AI inside a company, as a co-pilot that makes each individual more productive, misses the structural shift. An organization where every person is far more productive but still arranged in a five-layer hierarchy still has the same information-relay bottlenecks, the same politics, and the same distance between customer signal and product decision. The stronger claim is that the coordination cost that limits how large a firm can efficiently grow can be eliminated rather than merely reduced, by collapsing the layers between the intelligence layer and the people executing against it.¹

A company's roadmap is normally its own hardest constraint: a static bet on what customers want, built from inference such as customer research, interviews, support tickets, and social feedback. Conversational interfaces dissolve this. When customers interact with an AI interface rather than a fixed navigation structure, their queries and requests become direct, high-fidelity signals that the intelligence layer can read, and the roadmap becomes endogenous, defined by customers through interaction rather than fixed in advance.¹

Conditions for applicability

This model applies most cleanly to a business that generates a persistent, deepening signal about human behavior, something that understands human nature and gets a little deeper with every interaction. Financial transaction data is close to the purest version of that signal, since a transaction tends to tell the truth about a life or business in ways a stated preference cannot.¹ The argument extends to any business with genuinely rich behavioral signal, but for a business that does not generate this kind of signal, AI is more likely to remain an add-on than a structural core.

The implementation gradient

The practical advice for a hundred-person company runs in three steps: start querying, by putting the information the company already generates into an intelligence layer and beginning to have conversations with it, which alone can roughly double or triple understanding immediately because it stops relying on people to relay information;¹ question the hierarchy, since a company of that size is probably only two or three layers deep, and now is the moment to ask whether those layers are information-relay intermediaries the intelligence layer can replace; and build toward legibility, treating every tool, conversation, and output as a contribution to the world model.

A technical vocabulary for the world model

Fei-Fei Li's functional taxonomy of world models adds precision to this organizational metaphor. Her taxonomy identifies three kinds: a renderer, which outputs something for humans to see; a simulator, which outputs a state such as geometry, physics, or dynamics; and a planner, which outputs actions given an observation and a goal.² Mapped onto the company-as-intelligence idea, a company's intelligence layer functions as a simulator, building a state representation of organizational activity from every artifact it ingests, and its aspirational endpoint is a planner, a layer that takes actions based on its understanding of that state, such as proactively surfacing customer opportunities or flagging execution drift. Companies generally have plenty of data; what most have not built is the simulator layer that turns that data into an actual model.

Corroboration from Ramp

Ramp's 2026 fundraise framing corroborates that this thesis has entered mainstream fintech and venture discourse: business has historically been built on two pillars, capital and labor, and a third pillar, intelligence, is now the least governed cost and the single greatest opportunity.³ That observation names precisely the gap the company-as-intelligence thesis addresses: companies track and optimize capital and labor closely, but intelligence spend remains largely untracked and ungoverned.

A shipped implementation at Sierra

Sierra offers the clearest example of a second, independently arrived-at company building this thesis into a working system. Its co-founder describes running the company on an internal agent stack, and the build maps closely onto Dorsey's own implementation gradient. Starting to query becomes a single internal gateway aggregating every internal system, from chat to documents to operating reviews, permission-scoped so each employee's agent sees only what that person is allowed to see, letting anyone interrogate essentially the entirety of the company. The simulator plus a read interface becomes a purpose-built harness over that gateway, paired with a shared library of employee-built skills. The planner becomes a strategy agent grounded in a company document and every recent board letter and operating review, used to reason about what the company should be doing next, exactly the proactive, ask-what-to-do layer this thesis aims at rather than a layer that only answers what already is.⁴

Sierra's own account of the same build restates the hierarchy-collapse argument in operator language: role-specific agents, one per department, were tried first and failed, because the most important work happens across teams rather than within them, and companies are better understood as collections of jobs to be done, with departments existing only because one team cannot complete a whole job alone.⁵ Collapsing to a single shared agent is the tactical version of collapsing the relay hierarchy, where the unit of work becomes the cross-functional job rather than the organizational box. Two caveats accompany that account: business context, not raw intelligence, is the actual bottleneck, since frontier models are already capable enough for most business needs and the real constraint is making a company's own context queryable; and an outcomes-not-activity discipline is needed, since a legible, agent-run company can still generate an impressive-looking adoption chart without anything downstream actually improving, and measuring the outcomes that matter remains an open problem even for the company that built the system.⁵

Tensions

The signal prerequisite may be too high for most companies: rich transaction data is one thing, but email and chat logs may only produce a very detailed log that nobody meaningfully queries rather than a genuine world model. Legibility also creates political exposure, since an entirely legible company means every failure, drift, and internal conflict becomes visible to the intelligence layer and to anyone who queries it, which may chill candid internal communication. Regulatory complexity is a separate layer entirely from AI legibility and does not compress the same way. And world model quality compounds from data quality, so a company that feeds its existing dysfunction into the intelligence layer risks amplifying that dysfunction rather than resolving it.