> Site index: [llms.txt](https://ortelian.com/llms.txt) with all pages and descriptions. # How to build a world model for your company Source: https://ortelian.com/notes/how-to-build-a-world-model-for-your-company/ Markdown: https://ortelian.com/notes/how-to-build-a-world-model-for-your-company.md Author: Laurens Nys Published: 2026-10-14 Build a world model for your company: start from one piece of work, audit it, write the ontology, connect sourced data, then grow it in two-week cycles. You build a world model for your company by starting from one piece of work, not from your data. Audit that work on site, write down the things it depends on, connect the sources that describe them, keep a source on every fact, and put the first part to use within weeks. Then grow it in short cycles as more work moves onto it. This is the order we follow at Ortelian. If the term is new, start with what a [world model for knowledge work](https://ortelian.com/notes/what-is-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. This note is about how you get one. ## Why you start from the work, not the data The tempting first move is to connect every system and model everything. It fails for a simple reason: nobody can say what the model is for, so nobody can say when it is right. A company holds thousands of kinds of facts. A single piece of work, such as quoting a custom order or preparing a sales call, depends on a few dozen of them. Starting from the work gives you a test. If the agent doing that work gets the right customer, the current price list and the latest accepted delivery date, the model is good enough for that work. If it does not, you know exactly which part to fix. The model grows from there, one piece of work at a time, and every part of it earns its place. ## The steps ### 1. Pick one piece of work that matters Choose work that is frequent, costly when wrong and spread across several systems. Quoting, order intake, planning, account research and month end reconciliation are typical. Avoid work that lives in one tool; that tool probably already holds what the agent needs. ### 2. Audit it on site for two weeks Sit with the people who do the work. Write down every decision they make, what they look at to make it, where that information lives and where it is wrong or missing. Most of the knowledge that matters is not in any system. It is the rule a planner applies without thinking, or the spreadsheet the sales team keeps next to the CRM. We also [redesign the work](https://ortelian.com/notes/redesign-the-work-before-you-automate-it/) at this stage, because an agent on a broken process makes it faster, waste included. ### 3. Write the ontology for that work List the kinds of things the work depends on: customers, products, quotes, orders, machines, suppliers, people. For each, write what it means, how it relates to the others and what can be done with it. This is the legend of the map. The [ontology how-to](https://ortelian.com/notes/how-to-build-an-ontology-for-ai-agents/) goes through it step by step. ### 4. Connect the sources and keep the source on every fact Build data pipelines from the systems that describe those things: the ERP, the CRM, email, shared drives, public registries. Every fact keeps where it came from and when. That is what lets a person check an answer, and what lets the model prefer a registry address over one guessed in a note. The pipelines need to be inspectable, so that when a fact is wrong you can see which step produced it. ### 5. Give every thing one identity The same customer appears as a CRM account, an ERP debtor number and a name in an email thread. Until those resolve to one thing, the agent sees three customers. Identity resolution is unglamorous and it is where most of the early accuracy comes from. ### 6. Separate accepted facts from guesses Agents will write knowledge into the model: a summary of a call, an inferred relationship, a flag on a late supplier. Mark what has been accepted by a person or a trusted source, and what is still a proposal. An agent that treats a guess as a fact is how [AI agents give confident wrong answers](https://ortelian.com/notes/why-ai-agents-give-confident-wrong-answers/). ### 7. Put the first part to use Agree on real examples what a good result looks like for the work, then let an agent do it on the model, with people approving where it matters. At the same time, give your team access to the model from the AI tools they already use. We set that up through MCP inside the engagement, so people can ask the model questions on day one of use, not after a long rollout. ### 8. Grow it in two-week cycles Each cycle adds the next piece of work, the parts of the ontology it needs and the sources that fill them. Because every piece reuses what is already there, each one is cheaper than the last. The map gets denser where the company actually works. ## How long it takes With Ortelian, it starts with a two-week audit on site. Then we build in two-week cycles. By the end of the first two-week cycle, a first part of your world model is live, your people can use it right away, including from their own AI tools through MCP, and a first piece of work is automated. Every next cycle extends the model and automates more. The [implementation timeline](https://ortelian.com/how-long-ai-implementation-takes/) has the detail. ## An example: a manufacturer that quotes custom orders Take a typical SME manufacturer that makes products to order. A quote 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 to others. Today that picture is assembled by a person who knows which spreadsheet to open and which colleague to ask. The first cycle would model customers, products, quotes and the capacity plan, sourced from the ERP, the CRM and the planning sheet. An agent drafts quotes on that model and a person approves them. The next cycle adds production orders and suppliers, so the model can flag a quote that depends on a late delivery. The one after that adds invoices. Nothing in that sequence needs the whole company modelled up front. ## Build it yourself or with a partner You can build a world model in house. It means building the data pipelines, the graph, the agent harness, the evaluations and the review flows, and then running all of it. That is a product, and it needs a team that keeps building it. The [comparison with building in house](https://ortelian.com/ortelian-vs-building-in-house/) sets out what that takes. Ortelian is that product plus a forward deployed team that builds the model with you. It runs on our platform, hosted in the EU. Your company owns the world model, the IP built on it and the data in it, and nothing is shared with anyone else. ## Mistakes we see Modelling the whole company before any work runs on it. Treating it as a data warehouse project with an AI layer added later. Dropping the source of each fact to save time. Letting agents write into the model with no line between proposal and fact. Choosing a first piece of work nobody cares about, so nobody checks the results. ## Questions people ask ### What do we need before we start? One piece of work that matters, the people who do it and access to the systems it touches. You do not need a clean data warehouse or a finished data strategy first. The audit shows what is actually needed. ### Is a world model the same as a knowledge graph? No. The graph is one layer of it. The model adds the pipelines that keep it current, a source and a time on every fact, and the line between accepted facts and guesses. See [world model vs knowledge graph](https://ortelian.com/notes/is-a-world-model-for-knowledge-work-a-knowledge-graph/). ### How is it different from a context graph? The world model is the maintained map of the company. A context graph is the slice an agent draws from it for one task. See [context graph vs world model](https://ortelian.com/notes/context-graph-vs-world-model/). ### Who owns the world model? You do. The model, the IP built on it and the data stay yours, and nothing is shared with anyone else.