November 16, 2026

World Models for the Economic World

Companies record what happened, but rarely why they decided it or what would have happened otherwise. The models that learn from business data have arrived. The record they need to answer "what happens if" mostly does not exist yet.

World Models for the Economic World
by
Phil Bronner
•
Industry Analysis

Large language models gave companies a single model that can read and write across their documents, emails, and contracts. Until recently, nothing comparable existed for the other half of a company's data: the tables and records of its customers, orders, prices, and transactions. That data is kept to run the business, and the questions asked of it are about decisions: what to charge, what to stock, whom to hire. A model for it is judged by how well it predicts what will happen when the business acts.

This year, two of the world's largest technology companies made large bets to close that gap. NVIDIA acquired Kumo AI, Fortune reported on June 3.¹ SAP closed its acquisition of Prior Labs on July 17 and committed more than €1 billion over four years to its work.² Both companies build foundation models for structured data: large pretrained models designed to work across many companies' tables, rather than a model built from scratch for each question.

What none of them can buy is a company's record of its own decisions. Most companies do not keep one.

Our second post on world models ended on that gap. The economic world has the record problem: companies log what happened, but not why each decision was made or what would have happened if it had gone the other way. This post takes the economic world through the same test we used for robots, and looks at three companies we have backed that are working on it.

What the economic world has been doing instead

Business software is very good at telling a company what happened. What it has never done well is answer the question that matters most when a decision is on the table: what happens if we do this?

Today, that question is answered with a spreadsheet, a consultant, or a machine-learning model built specifically for that question. The company finds out whether it was right months later, when the decision can no longer be undone.

Three properties of the economic world make this hard.

The state is hidden. Just as a robot cannot see the weight of an object or what is behind it, a company cannot see its own state directly. Transactions and dashboards are observations. Real demand, customer intent, and the actual effect of last quarter's price change are hidden and have to be inferred.

Feedback arrives late. A robot learns the cup slipped when it hits the floor. A company learns months later that a decision was wrong, after the customers have already churned.

The decision is not in the record. The transaction is logged. The decision behind it, the conditions that prompted it, and the options a manager passed over usually are not. And because managers made each decision in response to what was happening at the time, the record rarely shows what would have happened if they had chosen differently. That is the record problem: the record shows what happened, but not why each decision was made or what would have happened otherwise.

One price change, played forward

Take a question every pricing team eventually asks: if we raise prices 5%, how many customers will we lose? Our first post set out a four-part test for a world model: does the model represent the state of a system, accept a specified action, predict the resulting state, and keep simulating from its own prediction? Let’s run the price change through it.

The state of the company is its customers, orders, prices, contracts, and support history. Some of that sits in the tables. Some of it, like how close a customer already is to leaving, does not.

The action is the change itself: prices up 5%.

The prediction is the next state: who renews, who downgrades, who leaves, and what that does to churn and margin.

Then the model keeps going. It takes its own prediction as the new state and plays the next quarter forward, and the one after that. A company can compare that future with the one where it holds price, before anything real is touched.

Which models a company can actually use

Rules people write. Many companies already model "what if." A financial model or a planning tool lets an analyst write the rules: if prices go up 5%, assume churn rises 2%. Tools like Palantir's let a company test a change on a copy of its data using rules it defines.³ These tools work, and people trust them because they can see the assumptions.

Foundation models for business data. This is what NVIDIA and SAP bought. They are pretrained on data from many companies, so a team can ask a new prediction question without building a new model. Some are starting to answer “what if”: Kumo compares how a customer is likely to behave with and without an offer. But Kumo’s own documentation says that needs a history of past offers and their outcomes. That is the record problem again.

The next step is a model that learns a company's data once and can then answer many questions from it, including what a specific action will do. That is the difference between a forecaster, which tells you what is likely to happen, and a world model, which tells you what is likely to happen if you act.

We see two kinds of economic world models, and we have backed companies building each.

Enterprise world models learn how one company behaves from its own records: its customers, orders, prices, contracts, and support history. They answer questions like "if we raise prices 5%, what happens to churn and margin?"

Wood Wide AI, an Ardent portfolio company founded by Pradeep Ravikumar and Varsha Raj, is building this kind of model.⁴ Pradeep was a professor in Carnegie Mellon's Machine Learning Department and co-editor-in-chief of the Journal of Machine Learning Research, and his research on reliable, interpretable, and causal machine learning is the foundation of Wood Wide's approach. The company describes itself as a world model for structured data. It learns a company's own tables, events and time series once, so the company can ask many questions of the same model, from predictions to explanations to what-if analysis, without starting a new modeling project for each one.

Ferris AI, another Ardent portfolio company, is starting with one industry.⁵ It calls its product a Delivery World Model: a living picture of every implementation project at professional services and delivery teams, assembled from meetings, documents, and email threads. Today, it flags risks before they become problems. The world model is the goal: showing how a change to a project will play out. What Ferris is building now is the record that goal depends on.

Market world models learn about the environment in which a company operates from public and third-party signals about economies, supply chains, and geopolitics. They answer questions such as "if a key supplier region is disrupted, what happens to prices and lead times?"

Sooth Labs, an Ardent portfolio company founded by Yaser Sheikh, Russ Salakhutdinov and Chuck Hoover, is working on this kind of model.⁶ Yaser was the founding director of Meta's Reality Lab in Pittsburgh, where he led work on codec avatars, photorealistic digital versions of people, and his lab at Carnegie Mellon's Robotics Institute created OpenPose, a widely used system for tracking human movement. Russ is a Carnegie Mellon computer science professor who studied under Geoffrey Hinton, led AI research at Apple, and later led research at Meta. Sooth describes itself as a world model for global foresight: it brings together economic, geopolitical, and operational signals to understand how risks move across markets and regions, so that a company can see how a disruption in one part of the world is likely to reach its own business.

Most real decisions need both: an understanding of how your own business responds, and of the world it responds to.

The record problem

The record problem runs deeper than missing logs. A model trained on a company's history learns from decisions that managers already made, and those decisions were not made at random. Managers chose them in response to what was happening at the time, such as strong demand or a competitor's move. Even when the data records those conditions, the model may never have seen the same decision made under different ones, so it cannot separate the effect of the decision from the effect of the conditions that prompted it.

Suppose a company has only ever raised prices when demand was already strong. A model trained on that history will see price increases followed by growth and conclude that raising prices drives growth. It has learned the pattern, not the cause. Then, ask it what happens if you raise prices during a slowdown, and it may give you a confident and wrong answer.

Medicine has the same problem. The Cell paper we drew our four-part test from finds that a model trained on health records can learn how care was delivered rather than how patients respond to treatment.⁷

Research suggests that size alone does not teach a model cause and effect. Models do better when they are designed and trained with causal structure in mind,⁸ for example, by learning from many simulated examples of cause and effect before they see a company's data.⁹ In our view, causation works alongside scale rather than replacing it, and the teams that understand both will have an advantage.

The other ingredient is data from real decisions. That can come from experiments a company runs, from decisions logged alongside what happened afterward, or from forecasts committed in advance and scored once the outcome is known.

Sooth works this way. Its forecasts are committed before outcomes are known and scored afterward, which builds a record of how well its model actually did.⁶

Ferris builds its record from inside the business. Because every project it supports ends with a known result, each one adds to what the model can learn. Data like this accumulates slowly, which is why a competitor cannot easily buy or copy it.

The test we use

When we meet founders in this space, we ask three questions:

  1. Does the model learn one lasting picture of the business, rather than starting over for every question?
  2. Can it take a specific action and predict its effect?
  3. Does it capture what happens after decisions are made, so its answers can be checked and improved over time?

The third question is the one most companies cannot answer yet. It is also the one that compounds once they can. Wood Wide, Sooth and Ferris are each early, and each is working toward this standard from a different starting point: a company's own data, the world outside it, and the work of delivering projects.

What the economic world teaches the others

  1. Keep the decision, not just the transaction. A world model learns from actions paired with consequences. Most companies log only the second half.
  2. Learn once, then ask. A model that starts over for every question is a forecaster. A world model holds one picture of the business and answers the next question from it.

A learned world model has to earn the trust that planning tools get from visible rules. Its advantage lies in businesses with too many interacting factors for anyone to write down the rules. If learned models cannot show that their answers beat the rules people write, they will become a feature inside existing planning tools rather than a new category. That is the question we will be watching most closely.

What comes next

The next post covers digital world models: the software that AI agents increasingly operate in, where some actions, such as a payment or a deployment, are hard to undo.

If you are building a model that can predict what will happen when a company or market acts, we would like to hear from you.

Disclosure: Ardent Venture Partners is an investor in Wood Wide AI, Sooth Labs and Ferris AI.

Phil Bronner is Managing Partner at Ardent Venture Partners. ardent.vc

Companies and researchers mentioned

Companies

  • Acquirers: NVIDIA, SAP
  • Acquired: Kumo AI, Prior Labs
  • Planning and rules: Palantir
  • Ardent portfolio: Ferris AI, Sooth Labs, Wood Wide AI

Researchers

  • Geoffrey Hinton
  • Pradeep Ravikumar, Varsha Raj (Wood Wide AI)
  • Yaser Sheikh, Russ Salakhutdinov, Chuck Hoover (Sooth Labs)
  • Marinka Zitnik and co-authors ("World models for biomedicine," Cell)
  • Noah Hollmann, Frank Hutter and co-authors (Do-PFN)
  • Rahul G. Krishnan and co-authors (CausalPFN)

Notes

  1. Fortune, "Exclusive: Nvidia snaps up Kumo AI in latest acquisition," June 3, 2026. Fortune reported that the financial terms could not be learned. fortune.com
  2. SAP News Center, "SAP Completes Acquisition of Prior Labs," July 17, 2026. SAP's figure is a commitment to invest more than €1 billion in Prior Labs over four years, not a purchase price. news.sap.com
  3. Palantir, "Ontology Scenario basics," Foundry API documentation. A scenario is a sandboxed copy of a company's data on which changes can be simulated without touching production. palantir.com
  4. Wood Wide AI, "The world model for structured data." woodwide.ai
  5. Ferris AI (tryferris.ai); Ferris AI, "Introducing Ferris' Delivery World Model" (worldmodel.tryferris.ai).
  6. Sooth Labs, "World Model for Global Foresight" (soothlabs.com); Ardent Venture Partners, "Why We Invested: Sooth Labs" (ardent.vc).
  7. Ayush Noori, Nic Fishman, Ada Fang, Lukas Fesser, and Marinka Zitnik, "World models for biomedicine," Cell, September 17, 2026. The paper notes that a model trained to reproduce health-record trajectories may learn the dynamics of healthcare delivery rather than how a patient would respond to an intervention. cell.com
  8. Jake Robertson, Arik Reuter, Siyuan Guo, Noah Hollmann, Frank Hutter, and Bernhard Schölkopf, "Do-PFN: In-Context Learning for Causal Effect Estimation," NeurIPS 2025. A tabular model pre-trained on millions of synthetic causal structures, including simulated interventions, estimated causal effects from observational data without being given the causal graph. proceedings.neurips.cc
  9. Vahid Balazadeh, Hamidreza Kamkari and co-authors, including Rahul G. Krishnan, "CausalPFN: Amortized Causal Effect Estimation via In-Context Learning," NeurIPS 2025. A model trained once on simulated causal data estimated treatment effects on new datasets without further training, and was also tested on real marketing experiments. arxiv.org/abs/2506.07918

‍