What is a world model for knowledge work?
A world model for knowledge work is a sourced, current map of a company that AI agents query instead of guessing. What it holds, its layers and examples.
A world model for knowledge work is a maintained, sourced map of a company and the world it operates in, which AI agents query instead of rebuilding that picture from tools on every run.
It holds the companies, people, places, regulations and events that shape the work, how they relate, what is true about each of them now and where each fact came from. It has three layers: the data that keeps it current, a graph of the things and their relationships, and the knowledge attached to those things.
Robots and video models learn a different kind of world model
In AI research, a world model usually means something else: a model that learns how a physical environment behaves so it can predict what happens next. A robot uses one to anticipate where an object will fall. A video model uses one to generate the next frames of a scene. Those models learn physics from pixels and sensor data, and what they know sits in their weights.
A world model for knowledge work holds a business instead of a physical scene: which hospitals are in the territory, which integrator serves each one, who changed role last week. It is not learned into weights. It is built from the company’s own records and from outside sources, and every fact in it can be read, checked and corrected. What the two share is the idea underneath: an agent acts better from a picture of its world than from whatever input arrived last. That is why we use the name, and why we add “for knowledge work”.
Some vendors now use “enterprise world model” or “world model for business” for related ideas. We say “for knowledge work” because the model is about the work, and the work depends on the world outside the company as much as on the records inside it.
It has three layers: the data, the graph and the knowledge
The data is everything that comes in from the company’s tools and from outside: the CRM, email, Slack, call recordings, public registries. Every fact derived from it keeps its source, so an address from a government registry and an address guessed in a rep’s note are not treated as the same fact. A fact that can change also records when it was true. When a person changes role, the new role supersedes the old one and the old one stays as history. A CRM field only shows the last value someone typed, which is how agents end up acting on stale facts, as Why coding agents work and company agents don’t shows.
The graph is the ontology filled in: the particular hospitals, integrators and people, each one a single identity whether it shows up as a CRM record, a registry entry or a name in an email thread. Without that, the same company in three systems is three companies, and the agent treats them that way. Things are joined by relationships that carry meaning and direction, such as “works at”, “installs at” or “regulated by”, which is what lets a signal travel from a regulation to a facility to the company that runs it to the open deal and its owner.
The agent’s job
Prepare Thursday’s call with St. Anna Hospital
Knowledge
- Supplied by Medcore, so the tender runs through them
- Tender likely in MarchAdded by this job, not yet reviewed
Graph
- St. Anna Hospital, hospital
- one hospital, from 3 records in 3 systems
- supplied by Medcore
- employs J. Peeters
- open deal Imaging renewal
Data · its sources
- CRM · St Anna Hosp.
- Email · procurement@st-anna.be
- Public registry · Sint-Annaziekenhuis vzw
The knowledge is what the company understands about those things, attached to the thing it describes rather than to the conversation it came from. A note about a regulation sits on the regulation, not on the call where somebody mentioned it. Agents write most of it and people review it, so an agent’s conclusion stays separate from accepted fact until a person or an agreed rule promotes it.
What is in a world model for knowledge work
A world model holds seven kinds of content. Most companies already have pieces of each, spread across tools and in people's heads. The world model puts them in one place, joined.
| Content | What it covers | Example |
|---|---|---|
| Things | The customers, products, sites, suppliers, people, contracts and regulations the work depends on | One customer, whether it appears in the CRM, the ERP or an email thread |
| Relationships | How those things connect | Which supplier delivers which material to which production line |
| Current state | What is true about each thing now | The open orders of a customer, the current price list, a person's current role |
| History | How things changed and when | Delivery dates that were promised and then moved |
| Sources | Where each fact came from and when it was captured | A registry entry, an ERP record, a reviewed call note |
| Knowledge and rules | What the company understands and how it decides | A planner's rule for which jobs go on which machine |
| Allowed actions | What agents and people may do, and who approves | An agent may draft a quote; a person approves any discount |
The first three make it a map. History and sources make it trustworthy. Knowledge, rules and allowed actions are what let agents act on it rather than only read it.
How it differs from a knowledge graph, an ontology and a context graph
These four terms get used as if they were interchangeable. They are layers of the same stack.
| Term | What it holds | The question it answers |
|---|---|---|
| Ontology | The kinds of things in a business, what each means, how they may relate and what can be done with them | What can exist here, and what can we do with it? |
| Knowledge graph | The particular companies, people, places and events, connected according to the ontology | What exists, and how is it connected? |
| World model for knowledge work | The graph, plus the data that keeps it current, the knowledge attached to each thing, time and a source on every fact | What is true about our business and its world right now, and how do we know? |
| Context graph | The slice one task needs: the things involved, their current state and the decisions already made about them | What does this piece of work need, and what was decided before? |
A knowledge graph is roughly the middle layer of a world model. The ontology is its legend. A context graph is what an agent draws from it for one task, plus the decisions made along the way. A world model is also not a wiki, a data warehouse or a vector index over documents. Each of those holds material about the company. None of them holds the things themselves, with one identity, a current state and a source.
An example from sales: a territory selling through integrators
At a deep-tech company that sells into hospitals through integrators, whether a hospital was worth a call depended on which integrator already supplied it. The CRM had no field for that, and the partner list was a spreadsheet of about 2,700 companies. We built the territory from government sources, hospital lists and integrator registries, normalised it and linked it to the CRM records they already had. The spreadsheet became 64 ranked accounts. 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 same model now feeds a meeting-prep agent that pulls previous calls, emails, deal history and similar companies before each external meeting, and a list of net-new accounts that rebuilds itself every morning.
An example from manufacturing: orders, quoting and production
Take a typical SME manufacturer that makes products to order. Quoting depends on the customer and their terms, the product configuration, current material prices, machine capacity in the weeks ahead and the delivery dates already promised. Production depends on the same orders, the suppliers behind each material and the rules planners apply to sequence jobs. In most companies that picture lives in an ERP, a CRM, a planning spreadsheet and a few experienced people.
In a world model, the customer, the quote, the order, the production job and the supplier delivery are connected things with a current state and a source. An agent drafting a quote can see that the earliest realistic date depends on a material that is already late, because the supplier delivery and the capacity plan are in the same map. A person approves the quote. When the order is placed, the same model shows planning and production what was promised and why. This is an illustration of the structure, not a client case.
Why it matters for AI agents
An agent is only as right as the picture it acts on. A world model changes what it can do in three stages. First correctness: the right company, the right person, the current role, the latest accepted fact. Then sight: connected facts show what no single source contains, such as a regulation change that touches several open deals. Then initiative: a standing goal, like that daily list, turns sight into work.
A smarter language model does not remove the need. It still does not know which integrators serve which hospitals in your territory or that the deal stages changed last week. That knowledge is not in its weights, and something has to hold it. The language model brings the intelligence. The world model brings the company.
Without one, agents rebuild the picture on every run and fill the gaps with what is typical, which is why AI agents give confident wrong answers.
How to build a world model for knowledge work
You build it around one piece of work, then grow it as more work moves onto it. These are the steps, in order.
| Step | What it involves |
|---|---|
| 1. Pick one piece of work | A job with a clear owner and a standard you can check, such as preparing a quote or a planning round. |
| 2. Write the ontology for that work | Name the kinds of thing the work touches, how they relate and what agents may do with each. |
| 3. Connect the sources | Agree access to the systems the work depends on, such as the CRM, the ERP, email and outside registries. |
| 4. Resolve identities | Make every source point at one record per company, person, product and policy. |
| 5. Put time and a source on every fact | Record when each fact became true, what it replaced and where it came from. |
| 6. Attach knowledge to things | Keep what the company understands on the thing it describes, and keep agent findings apart until a person or an agreed rule accepts them. |
| 7. Name an owner and grow it | An internal owner decides what comes next. Each new piece of work adds to the same model. |
With Ortelian this starts with a two week audit on site, then two week build cycles. The world model, the IP built on it and the data stay with the client. The full walkthrough is in how to build a world model for your company.
Going further
Building one starts from a single piece of work, not from your data: how to build a world model for your company goes through the steps. The phrase enterprise world model is also used for simulations and synthetic agent environments; what is an enterprise world model sets the three meanings side by side. For the neighbouring terms, see context graph vs world model and world model vs knowledge graph.
Questions people ask
Is a world model for knowledge work the same as a knowledge graph?
Partly. A knowledge graph is its structure. A world model adds the data that keeps the graph current, time and a source on every fact, and a line between accepted facts and guesses. The longer answer is in Is a world model for knowledge work a knowledge graph?
Is it the same as a world model in AI research?
No. Research world models learn how a physical environment behaves from pixels and sensor data, and hold what they learn in their weights. A world model for knowledge work holds a business, is built from records and outside sources, and every fact in it can be read, checked and corrected.
Does it replace the CRM?
No. The CRM, email and Slack keep doing their jobs and feed the model. The model holds what none of them hold alone: one identity per thing across systems, the relationships between things and the history of how they changed.
How does a company get one?
It is built around the work you want agents to do, starting with the ontology for that work and the sources that fill it, then grown as more work moves onto it. At Ortelian we build it during a deployment, and it runs on our platform, hosted in the EU, as part of the engagement.
Is a world model for business the same thing?
Usually close. Some vendors say enterprise world model or world model for business for a maintained map of the company; others mean a simulation that predicts outcomes, or a synthetic environment for training agents. We compare the three in what is an enterprise world model.
Why do AI agents need a world model?
Because the model they run on does not know your business. Without a maintained map, an agent rebuilds the picture from tools on every run and fills the gaps with what is typical. With one, it starts from the right customer, the current state and the source of each fact.
How long does it take to build one?
With Ortelian, a two-week audit on site, then two-week build cycles. By the end of the first cycle, a first part of the world model is live, your people can use it from their own AI tools through MCP, and a first piece of work is automated. Each cycle after that extends it.
Who owns it?
The company does. The world model, the IP built on it and the data stay with the client, and nothing is shared with anyone else. At Ortelian it runs on our platform, hosted in the EU.
