Most companies still think the next layer of AI work is a tool.
It is not.
The tool problem was solved a while ago. The model can write. The agent framework can call APIs. The orchestrator can coordinate. The infrastructure is sufficient for the work most companies actually want done.
What is missing is something different. Something procurement cannot help you with. Something that does not show up in the Gartner quadrant.
What is missing is a model of the company.
The agent acts on a reconstruction
When I talk to teams trying to put agents inside their organization, the conversation almost always travels the same arc.
First, the team builds the agent. They focus on the model, the prompt, the toolset. They make impressive demos. Then they try to put the agent somewhere real.
That is where it fails.
Not because the agent cannot reason. Not because the model is too small. The agent fails because it is acting inside a company it cannot see.
It reads the wiki, but the wiki is wrong. It reads the CRM, but the CRM contradicts the wiki. It reads the Slack history, but the Slack history is performative. It reads the email, but the email omits the decisions that happened in person. It builds a theory of the company out of artifacts that were never meant to be a model of the company.
So it produces work that is plausible and wrong.
This is not a model problem. It is a representation problem.
Every agent builds a world model. The only question is whether they build one each, or whether they build one together.
The missing layer is not another tool
Most organizations have too many tools already. Their context is scattered across project management systems, docs, CRMs, drives, chats, dashboards, decks, spreadsheets, and the memories of the people who have been around long enough to know which of those things are lying.
The missing layer is not another place to click.
The missing layer is a representation.
A world model is a continuously updated, machine-readable map of how the organization actually works. Not how the wiki says it works. Not how the deck says it works. How it actually works, today, with the decisions and exceptions and carve-outs that make work move.
That definition is not mine; it is the working definition behind everything we do. And once you have it in front of you, the three things it has to hold come into view.
Processes. How work gets done, with exceptions and handoffs. Not the documented flow — the actual flow, including the carve-outs that nobody admits to in the kickoff meeting.
Roles. Who is responsible for what, including the accountability relationships that are invisible on the org chart. The person who has the title and the person who actually owns the outcome are not always the same person; the world model has to name both.
People. Who knows what. Who works with whom. Where tribal knowledge lives. The org chart shows the lines on the wall; the world model is the actual network of expertise that keeps the company alive.
Those three are the spine.
Legibility is the prerequisite
I named these surfaces in The Legible Organization — workflows, mandates, canon, permissions, policies, records. Plus org structure and people, the two layers most companies already have in some form.
The world model is what happens when you stop treating those layers as a wiki and start treating them as a queryable representation — machine-readable, versioned, kept honest by the work it sits under. Legibility is the prerequisite. The world model is what you build with it.
The minimum viable world model is one workflow
The temptation, at this point, is to map everything.
That is how organizations turn useful ideas into dead programs.
The world model does not have to start universal. It starts with one workflow.
Pick something real enough to matter and bounded enough to see. Renewal of mid-market accounts. The new-product launch process. The escalation flow when customer support has to involve legal. Something that already exists, that already costs you something, and that you suspect breaks more often than your team admits.
Then map that workflow. Name the trigger. Name the steps. Name the owner for each step. Name the artifacts the work depends on — the canon it has to align to, the policies that gate it, the records that explain why it is shaped the way it is. Name the carve-outs. Name who is allowed to invoke them. Name where the invocation gets recorded.
That single workflow, fully mapped, is the minimum viable world model. It is small. It is dense. It is enough to put one agent inside without that agent reconstructing the company badly.
Then you do another one. Then another.
A real world model accretes the same way a city does — block by block, around the people and processes that already function. Not master-planned. Not delivered as a single thing. Built.
Maintenance is the product
Here is the failure mode I see most often.
A team builds the first slice of a world model. It is good. It is useful. They put an agent inside it. The agent works. Six months later, the workflow has changed three times, the mandate has moved between two roles, the policy has gained two carve-outs, and the world model has not been updated.
The agent is now operating on a fiction.
This is the same failure mode that kills documentation. The difference is that the cost shows up faster, because the agent acts on the world model without asking around first.
A world model is not infrastructure. It is a discipline. A map ages out the moment maintenance stops.
The discipline is its own kind of work, and it is the kind of work most companies have never had a role for. The job description does not exist yet on most org charts. It is closer to a librarian than a project manager — someone whose job is to keep the company's representation of itself accurate, who has the authority to push back when the actual work diverges from the documented work, who can tell when a workflow has drifted past the point where the agent inside it is still safe.
Call it a librarian, an ontologist, a chief-of-staff-for-the-machines. The title isn't settled. The work is — and the companies that staff it early will be the ones whose agents stay accurate past month six.
The cost was already there
A new hire can survive a partial world model.
If something is missing, they ask. If something is wrong, they eventually find out from a coworker. If a workflow has drifted, they learn the new shape by trying the old one and getting corrected.
This is expensive — onboarding takes the time it does for exactly these reasons — but it is survivable.
An agent has none of those affordances. It cannot ask. It cannot watch. It cannot get corrected gracefully in the hallway. It acts on whatever model of the company it has, immediately and at scale, and the cost of being wrong shows up in a customer email, or a compliance question, or a number on a dashboard.
The agent does not absorb the company by osmosis. The agent absorbs the world model. If the world model is wrong, the agent acts wrong.
This is why the world model is not optional once agents are real. It was always being paid for — in onboarding time, in re-litigated decisions, in the fraction of every meeting spent figuring out who to ask. Agents do not invent the cost of illegibility. They surface it.
The hinge
If there's a hinge in this work, it's the moment a company stops treating its representation of itself as documentation and starts treating it as an instrument the company can act through.
Until then, every AI investment compounds locally and decays globally. The individual pilots work. The system-wide gains never arrive. The dashboards look fine and the substrate is rotting underneath.
After it, even a small surface — one workflow, one department — starts to compound. The pilot has somewhere to land. The agent has a place to live. The work, for the first time, sits on a representation that does not drift the moment you look away.
Closing
A world model is less infrastructure than instrument.
It is not a database. It is not a system of record. It is not another place to click boxes. It is the thing the company uses to see itself — to direct itself, to delegate to itself, to act through itself.
You cannot direct a company that cannot see itself.
You also cannot scale agents through a company that cannot see itself. They will see something — they have to — and what they see will be a reconstruction, built locally, in fragments, from artifacts that were not designed to be a model.
The next layer of work is not a tool.
The next layer of work is a model of the company.