Why AI agents need a world model of your company

AI agents stall inside companies because they are deployed into an environment nobody has described. The model can read, reason and call tools, but it does not know that large quotes wait for a director who is rarely in the office, that the month-end pack depends on one person's macro, or that the fault route on line 2 skips the fitter. A company world model, a living description of how the work actually runs, built from the people who do it, is what turns a capable model into a useful agent. Build it first, and most of the reasons agent projects fail disappear before the first agent is written.

The failure is not the model

Fortune's report on MIT's 2025 study of enterprise generative AI found that about five percent of pilots achieved rapid revenue impact while the rest stalled with little measurable effect on the P&L, and named the cause as a learning gap rather than model quality: generic tools stall in enterprise use because they do not learn from or adapt to the company's workflows (Fortune, 2025). Stack Overflow's 2025 survey of developers, the people closest to these systems, found 66 percent naming AI solutions that are almost right, but not quite, as their biggest frustration (Stack Overflow, 2025). Almost right is exactly what you get from an agent that knows the general shape of accounts payable and nothing about yours.

PwC's 2026 predictions put the split plainly: technology delivers about 20 percent of an AI initiative's value, and the other 80 percent comes from redesigning the work so that agents can handle the routine cases (PwC, 2026). You cannot redesign work nobody has described.

What an agent actually needs to know

Anthropic's guidance for teams building agents is to invest as much effort in the agent-computer interface as we have always invested in human interfaces, and to build the right system for the need rather than the most sophisticated one (Anthropic, 2024). Inside a company, the interface is the company. An agent that is going to run a route needs four things about it.

The route as it really runs. Not the procedure, the practice. Who gets the first call, what is logged and what is not, where it waits.

The exceptions and who owns them. Agents handle the routine case. The value of the design is in what counts as an exception and which person it goes to, with a time window.

The people and the history. Who has done this before, what was decided last time, which repair matched this fault in August.

What the company itself says matters. Which problems people raise most, and what each one costs in the company's own hours and rates, so the agent is built where it pays rather than where it was easy.

Those four things are a company world model. They do not live in any system. They live with the people doing the work, which is why they are missing from every agent built on documents and databases alone.

Why the world model has to come from everyone

Three people will describe the same process three ways, and the differences are not noise. They are where the handoffs break, where the rework sits and where the money is. A world model built from a manager's account, or from the SOP folder, averages those differences away and hands the agent the tidy version that does not exist. A world model built from everyone keeps the contradictions, asks again, and records the route that actually runs.

That is why we built Transcript to talk with every person in the company, privately, in their own language, and to turn what they say into one world model: blueprints, counted problems, and the people everything depends on. No manager can read a conversation. Everyone sees the model. Every agent the company deploys works from it.

Our view

In our experience the order matters more than the tooling. Hear everyone, build the world model, let the company choose the visions worth building, then build the agents on the model. Teams that start with the agent spend their first months discovering the company by trial and error, one almost-right answer at a time, and the pilot dies of fatigue before it reaches the P&L. Teams that start with the world model know on day one which route to automate, who owns the exceptions, and what the saving is worth in the CFO's own figures.

Questions people ask next

Can we not just give the agent our documentation?

You can, and it will learn the official process. The cost sits in the unofficial one: the bypass for urgent invoices, the call that skips the fitter. Documentation rarely contains either.

Does the world model have to be complete before the first agent?

No. It has to be complete for the route the agent will run, and it should be counted well enough to know that route is worth it. The rest keeps building while the first agent runs.

How does the world model stay current once agents are running?

People keep talking about their work, and agents write back what they did. The blueprint updates; the agent reads the current version. A model that is not maintained decays into documentation.

Who owns the world model?

The company. It is built from its people and it belongs to them, together with every agent built on it.

Where do most companies start?

Where the world model says the problem is most widely felt and most expensive: often finance and accounting, procurement, customer support or operations, but the model decides, not the vendor.

References

1. 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/
2. Stack Overflow, 2025 Developer Survey, AI section (2025), https://survey.stackoverflow.co/2025/ai
3. PwC, 2026 AI Business Predictions (2026), https://www.pwc.com/us/en/tech-effect/ai-analytics/ai-predictions.html
4. Anthropic, Building effective agents (2024), https://www.anthropic.com/research/building-effective-agents