What is a company world model?

A company world model is a living picture of how a company actually works, built from the people who do the work rather than from the org chart or the procedure manual. It holds three things: every process as it really runs, the problems people raise and how many raise them, and the people and handoffs everything depends on. It stays current as the work changes, and it is the thing an AI agent needs before it can act inside the company. The term borrows from AI research, where a world model is the internal representation a system uses to predict what will happen next, and applies it to the one environment most companies have never described: themselves.

Where the term comes from

In machine learning, a world model is a compressed internal representation of an environment that lets a system predict the consequences of its actions. Ha and Schmidhuber (2018) built agents that learned to play games inside their own model of the game, and opened their paper with the observation that the image of the world we carry in our heads is itself only a model, made of selected concepts and the relationships between them. In 2026 the term moved into products: World Labs (2026) publishes a taxonomy of world models and ships one for spatial intelligence, so that AI systems can generate, edit and simulate environments rather than just describe them.

The idea transfers cleanly. An AI agent dropped into a company is an agent dropped into an environment. Without a model of that environment it can read documents and answer questions, but it cannot predict what happens when it approves an invoice, reroutes a fault or drafts a quote, because nobody has told it how those things actually flow.

What a company world model holds

A useful company world model has three layers, and each has to come from the people doing the work rather than from the systems recording it.

Blueprints: every process as it really runs. Who does what, in what order, where it stalls, and what the official version leaves out. A blueprint records that large quotes wait for a director who has no slot for them, not that "quotes above the threshold are approved by a director". It carries a process map and the observations it was written from.

Problems: what hurts most, counted. Not a list of complaints but a ranked set: how many people raised each problem, from which functions, and what it costs in the company's own hours and rates. Counting is what turns anecdotes into a decision.

People: who everything depends on. The handoffs, the workarounds and the quiet experts nobody put on a chart. The person who is the only one who can run the month-end macro is in the world model. She is not in the ERP.

The knowledge for all three already exists. It is just held by individuals. Panopto's survey of 1,001 American employees found people spend about five hours a week waiting to reach the one colleague who has the knowledge they need, and around six hours a week duplicating work someone had already done (Panopto, 2019). A world model is that knowledge written down once, checked with the person who said it, and kept current.

How it differs from a process map, a wiki or a knowledge base

A process map is one blueprint, drawn by whoever was in the workshop, usually the manager. A wiki is a place documents go to die; nobody updates the page when the threshold changes. A knowledge base answers the questions someone thought to write down.

A world model differs in three ways. It comes from everyone, not a sample, so the contradictions between how teams describe the same step are found rather than assumed. It is quantified, so problems come with a count and a cost rather than a paragraph. And it is alive: as people keep talking about their work, the model updates, and the blueprint from March is not the blueprint in September.

A snapshot of a company is an ontology: what exists and what is attached to what. A diary is a world model: what usually happens, what stalls, what breaks when a key person is away. Most attempts at a company world model start from systems, because systems are easy to read; they get the recorded world and miss the part that never touches a system, which in a mid-sized company is most of the dysfunction. Ours starts with the people who do the work, adds systems second, and learns from outcomes third. It gets richer with every layer, and as it fills with motion it stops describing the company and starts predicting it.

How it gets built

The only source that has the real process is the people running it, and the only way to reach all of them is a private conversation with each one. That is what Transcript, our product, does: it talks with every person in the company in their own language, a few minutes a day, compares what was said with everyone else, asks again where accounts disagree, and lets each person approve what was learned from their words before it enters the model. Nobody, including managers and including us, can read a conversation. Everyone sees the world model. Privacy is not a courtesy here; it is the mechanism that makes people describe the workaround instead of reciting the manual.

Our view

We think the world model is the missing layer in most AI programmes. Fortune's account of MIT's 2025 study puts the share of enterprise generative AI pilots with rapid revenue impact at about five percent, and names the reason as a learning gap: generic tools do not learn from or adapt to the company's workflows (Fortune, 2025). Anthropic's guidance on building agents makes the same point from the engineering side: invest as much effort in the interface between the agent and its environment as you would in a human interface (Anthropic, 2024). In our experience the environment is the company, and describing it is most of the work. Build the world model first, then decide which vision is worth building on it.

Questions people ask next

Is a company world model the same as a digital twin?

No. A digital twin mirrors a physical asset or a system from sensor and transaction data. A company world model describes how people actually work, including the parts that never touch a system: the phone call, the spreadsheet, the person everyone asks.

How long does it take to build one?

Weeks, not months. With everyone talking a few minutes a day, a company of a few hundred people has a first world model within weeks, and it keeps improving from there.

Who can see it?

Everyone sees the world model: the blueprints, the counted problems, the dependencies. Nobody sees an individual conversation. In Transcript that is enforced in the database, not in a policy.

Does it replace process mining?

It complements it. Process mining measures what touched a system. The world model describes the whole route, including the steps that live in email and hallway conversations, and tells you where to point the miner.

What do AI agents do with it?

They work from it. An agent that knows the real fault route, the people on it and its history can run the route; an agent without that knowledge can only draft text about it.

References

1. Ha, D. and Schmidhuber, J., World Models (2018), https://worldmodels.github.io/
2. World Labs, blog: A Functional Taxonomy of World Models; Atlas: A World Model for Spatial Intelligence (2026), https://www.worldlabs.ai/blog
3. Panopto, How much time is lost to knowledge sharing inefficiencies at work? (2019), https://www.panopto.com/blog/how-much-time-is-lost-to-knowledge-sharing-inefficiencies/
4. Fortune, MIT report: 95% of generative AI pilots at companies are failing (2025), https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
5. Anthropic, Building effective agents (2024), https://www.anthropic.com/research/building-effective-agents