The debate on the #companyworldmodel has exploded, and it has all the hallmarks of the debates I enjoy: a strong insight at its core, and a wrong conclusion built on top of it. The strong insight is this: companies produce enormous quantities of operational data, yet they continue to describe themselves through static documents, org charts that are two years out of date, and policies disconnected from how the work actually gets done. An enterprise world model that represents dependencies, decision flows, risks, responsibilities and systems is needed. It's needed badly.
The wrong conclusion is treating it as an operating system. Describing the organisation as code, compiling it, running it. Daniel Rothmann did this systematically in Company as Code: mapping the relationships between people, policies and systems the way you map dependencies in a codebase, with a "staging environment" for testing organisational changes before implementing them. The technical insight is elegant. The problem is the assumption holding it all up.
Why the temptation is understandable
Anyone coming from software recognises the pattern. If I can describe infrastructure as code (Infrastructure as Code), why not do the same with the organisation? If AI agents need context in order to operate, why not give it to them in a structured, queryable, versioned form? It's a line of reasoning that holds up. Up to a point.
The exact point where it stops holding up is this: an organisation is not an infrastructure. A server has no opinions about how to allocate resources. A router doesn't build trust with the router next to it. A pipeline doesn't take intelligent shortcuts because it has understood the objective better than the person who designed it.
Organisations work in a radically different way, and the value - the real kind, the kind that makes the difference between a company that adapts and one that stiffens up - doesn't come from the correct assignment of responsibilities. It comes from what happens *between* the nodes, not from the nodes. It comes from the informal relationships that short-circuit three levels of hierarchy. From the productive ambiguities that leave room for judgement. From the managed conflicts that lead to solutions no policy had anticipated. From the accumulated trust that allows a team to move fast under uncertainty. None of these things can be codified.
The problem with "Company as a Code"
Reducing the organisation to a grammar of tasks, roles, policies and declared relationships is a move made by engineers accustomed to deterministic systems, applied to complex adaptive systems. A useful model should represent fragility, hidden dependencies, the density of relationships, the quality of conversations, the capacity to learn. Not just who reports to whom, or which role owns which task. The risk is not theoretical. If the company world model becomes the primary means of *governing* (not of understanding, but of executing) it sterilises precisely what makes an organisation capable of adapting. You codify the current state, and then you defend the current state. The organisation becomes its own specification. There's a reason why the most rigid organisations I know are the ones that have invested the most in RACI matrices, documented procedures and precise org charts: not because documentation is the problem, but because they've used it as a substitute for judgement rather than as a support for it.
Where the world model does make sense (and it makes a great deal of sense)
That said, the company world model as a *map* is exactly the tool needed to govern collaboration between agents and people in the era we are entering. Rohit Krishnan has written the most lucid thing I've read on this: with thousands of agents making thousands of decisions a day, management cannot work the old way. The metaphor is moving from a first-person shooter (you control every single movement) to Starcraft: you move units towards objectives, supervising by exception. The world model is the battlefield that lets you do it. It doesn't execute the organisation. It *represents* it with enough fidelity to enable three things:
- Giving context to AI agents. Without an organisational map, agents operate on fragments: they optimise isolated pieces, ignore dependencies, generate local coherence and systemic incoherence. The world model is the shared nervous system.
- Running counterfactuals before acting. Changing the structure? Reallocating responsibilities? Simulate it first. The world model is a space for simulation, not for execution.
- Managing by exception rather than by control. I don't check that every agent does the right thing: I build the capacity to detect when something goes wrong relative to the expected baseline. Fowler and Böckeler call this a harness: not the baton of command, but the network of sensors and constraints that keeps the system on course.
These three uses are powerful. None of the three requires *executing* the organisation as if it were code.
The distinction that really matters
Traditional management has been trying to execute organisations for decades. Procedures, RACI, org charts, cascading OKRs. The result is well known: in 95% of cases companies are still organised as they were in the Middle Ages (vassals, vavasours, petty chiefs) - not because anyone is stupid, but because the pyramid structure was the right answer to the reasons why firms came into being in the first place: economies of scale, supply chain integration and, above all, the concentration of knowledge.
When those three reasons give way (and AI is causing all three to give way at once) the pyramid isn't merely inefficient. It is structurally misaligned. But the answer is not to replace the pyramid with another form of centralised execution, this time codified. The answer is to decentralise autonomy and mastery to whoever is in contact with the problem (the edge of the work, not the centre) and to centralise governance through AI, not through layers of management. Within this scheme, the world model does its job: it is the instrument of distributed governance, the map that orients agents and people, the simulation field in which management tests hypotheses before putting them into play.
It is not the operating system. It is the nervous system that transmits signals, integrates information, and leaves the organism the capacity to respond in ways that no specification has anticipated. Priel Korenfeld is right: no AI agent will ever understand the sarcasm in a meeting or the implicit references built up over years of working together. Not because AI is stupid, but because that knowledge - tacit, relational, contextual - cannot be codified. And if it cannot be codified, it cannot sit in the world model. Which tells us something important: not all organisational value is representable. And what isn't representable is often what matters most.
So yes: let's build enterprise world models. Let's make them queryable, up to date, useful to the agents and to the managers supervising them. Let's use them as a map for navigating complexity and as a simulation field for high-impact decisions.
But if someone proposes compiling your organisation and running it, ask them how many times they've worked in a real one.
