What is a world model for a company?

A company runs on facts about things. Which hospitals are in the territory, who runs procurement at each one, which integrator installed the last system, what was promised on Tuesday’s call. Those facts are spread across a CRM, an inbox, a Slack channel and the heads of a few people who have been there long enough to hold the connections. That worked while people did the work. It stops working when an agent does it, because an agent arrives with no memory of the company and a context window that empties when the task ends. A world model is the fix: a live, linked model of what exists inside the business and in the world around it, the companies, people, places, regulations and events and how they relate, with provenance on every fact, that agents work from instead of working from raw text.

People tend to hear that and map it onto something they already have: a wiki, a data warehouse, a vector index over the documents or the language model’s own memory. It is none of those. Each of them holds material. None of them holds a world.

The model covers the company and the world it operates in

Inside the company are the accounts, the contacts, the open deals, the calls and the rules the company runs by. Around it are the companies it could sell to, the people who move between them, the facilities they operate, the regulations that apply to them and the events that change any of it. The model holds both, and the links between them, because most of the value is in the links.

Ortelius did not draw new maps. He took the maps that existed, redrew them to one standard, put them in one book and credited every source. A world model does that for a company: the CRM, the public registries, the email, the calls, redrawn to one standard with sources named, so that the things which were always connected become visible.

At a deep-tech company that sells into hospitals through integrators, the customer’s world was hospitals, the integrators that install in them and the government sources that list both. Their CRM held one slice of that. The territory we built came from government sources, hospital lists and integrator registries, normalised and linked to the CRM records they already had.

Five properties separate a world model from a database

The first is identity. Every object is one thing, whether it shows up as a CRM record, a registry entry or a name in an email thread. In a Belgian outreach build we started from about 4,437 companies in a public CO2 register and found that matching on company name was unreliable. The domain became the primary key. Without a stable identity, the same company in three systems is three companies, and the agent treats them that way.

The second is current state that supersedes rather than overwrites. A fact that can change records when it was true and when it was recorded. When a person changes role, the new role supersedes the old one and the old one stays as history. Nothing is overwritten in silence. At a planning-software company, a new head of sales changed the deal stages without telling the CFO. The CRM just showed the new stages. An agent working from that CRM that week would have been working from a fiction.

The third is provenance. Every promoted fact carries where it came from, and the evidence stays beneath the state derived from it. A company’s address from a government registry and the same address from a rep’s guess in a note are not the same fact. The model knows which is which, so the agent can rank them and a person can check.

The fourth is relationships that carry meaning. A join table says two ids co-occur. A relationship says “installs at” or “regulated by” or “works at”, with a direction. That is what lets a signal travel: an OSHA violation attaches to a facility, the facility belongs to a company, the company has an open opportunity and the opportunity has an owner. The schema that decides which concepts exist and which relationships are allowed is the ontology. The populated graph holds the particular hospitals and people.

The fifth is that knowledge attaches to the thing it describes, not to the conversation it came from. A note about a regulation attaches to the regulation, not to the call where somebody mentioned it. This is the opposite of a wiki, where the page is the unit and the thing it is about is a word in the title. It also means an agent’s conclusion does not become an accepted fact just because an agent wrote it. Interpretation stays separate from accepted truth until a person or an agreed rule promotes it.

It changes what an agent can do in three stages

The first stage is correctness: the agent acts on the right company, the right person, the current role, the latest accepted fact. At the deep-tech company, the partner list was a dirty spreadsheet of about 2,700 companies. Once the territory was modelled from government sources and integrator registries and linked to the CRM, that list collapsed to 64 ranked accounts, 48 of them with contacts already pulled. The language model did not get smarter between those two numbers. It went from working on 2,700 strings to working on 64 things that exist.

The second stage is sight: connected facts show what no single source contains. A changed regulation connects to the affected accounts, the open opportunities, the earlier calls and the people who must act. The same company now has a meeting-prep agent that runs at 9am in a shared Slack channel. For each external meeting that day it pulls previous calls, emails, deal history and similar companies from the CRM and the graph. Nobody told the rep about it. He found it on his own and now checks it before every external meeting.

The third stage is initiative: a standing goal turns sight into work. A top-50 list refreshes daily at 7am so the rep has it by 9am. It only shows net-new companies, a company can reappear after 30 to 60 days, and as accounts move stage new ones fill in. Each run leaves a trace on the graph, so the next run starts where the last one stopped. Correctness is proven. Sight is the larger bet, and initiative depends on it.

A smarter model does not remove the need for one

The obvious objection is that the next generation of language models will make this unnecessary. It will not. A much smarter model still does not know which integrators serve which hospitals in your territory, what was promised on Tuesday’s call or that the deal stages changed last week. That knowledge is not in its weights. It is in your world, and something has to hold it. A better model improves the product. The durable asset is the maintained world around it: identity, current state, history, relationships, provenance, permissions and outcomes.

You can rent the model, but not the world it has to work in.