Artificial Intelligence
Why We Need a World Model for Work
The next leap in enterprise AI isn't a bigger model. It's one that understands consequences.
Julie Colwell
Principal Strategist
Workday
In 2022, researchers studying large language models noticed something alarming. The systems kept getting bigger, faster, and more fluent, but they kept making the same kind of mistake. They could write a flawless reorganization plan, but it would violate three company policies. They could summarize a contract, but they had no understanding of what the contract governed. They could predict the content, but they had no concept of the context.
Think about it like this.
It takes a teenager about 20 hours to learn to drive. Not because driving is simple, but because they already know what a car is, what other drivers are likely to do, what else to look for when a ball rolls into the street, and that a yellow light means something different than a red one. Twenty hours is enough only because a lifetime of context did the rest of the work. (And it still takes time and maturity for them to be good at it.)
Compare that to the self-driving car. Autonomous driving can follow a route flawlessly. But it struggles with a construction worker stepping into the road or a handwritten detour sign; it falls apart the moment the rules of the situation are not in any dataset. As Stanford researcher Fei-Fei Li has pointed out, autonomous driving started learning in 2005, and two decades later it's really good, but it took a long time.
Report
AI Agents Are Here—But Don’t Call Them Boss
When What Happens in the Digital World Stays There
Today's AI functions primarily in a digital world. But the systems and people it manages and impacts exist in the physical world. AI understands one, but not the other. That's why AI needs what we call a World Model for Work.
In business operations, generic AI functions might play out like this. Ask an AI agent to fix an overstretched support team and you have a plan in seconds: promote two people, backfill three roles, move a regional manager to balance the load. The logic holds. But on a closer look you might find things like this:
The promotion skips a level the company's job architecture does not allow
The three backfills together run past the headcount budget finance approved for the quarter
The manager being moved is the only person certified to approve something critical to the month-end close
None of that appears in the plan, because none of it was visible to the agent. It knew the words. It didn't know the world.
Without a world model for work, AI can be impressive, but you can't really trust it.
Without a world model for work, AI can be impressive, but you can't really trust it. And an AI you can't trust with consequential actions won't transform the world of work; it's just a better search bar.
The question for anyone deciding what to hand over to AI today is this: are you trusting it with words or are you trusting it with consequences?
The Making of a New Model - Five Challenges
Trusting AI with consequences requires a world model for work. This is an AI system that acts as a live, internal simulation of a specific company environment — its rules, structures, relationships, and constraints — that allows an agent to predict the consequences of actions before taking them, and to plan reliably across the dependent steps that company decisions require.
Five specific challenges stand between today's AI and a world model for work:
Cause and effect — predicting what an action will actually cause, not just predicting patterns
Robustness — holding up when conditions change, instead of breaking on small differences
Planning over time — reasoning reliably across dependent steps rather than one step at a time
The data gap — learning an environment that, unlike text, was never fully written down
Acting safely — being trustworthy enough to take consequential actions, not just describe them
Finding the solutions for these five challenges defines the frontier of AI development today.
Let's take a closer look at each one.
1. AI Cannot Reason About Cause and Effect
Large language models are prediction engines. They learn which words follow which other words, at a staggering scale, across enormous amounts of text. That produces something that looks a great deal like reasoning, but isn't.
Yann LeCun, Turing Award winner AI researcher, argues that this is the central limitation of modern AI. He and others now contend that the alternative to simply scaling language models is a different kind of system — one that learns about physical reality and can predict, plan, and reason about cause and effect, which current models cannot reliably do.
Stanford scientist Fei-Fei Li draws this distinction:
A language model learns the statistical structure of text
A world model learns the structure of space and time — how light falls on a surface, how an object responds to force, what happens when an action is taken in a particular environment
An intelligent system cannot only read a book. It has to read the room.
An intelligent system, she argues, cannot only read a book. It has to read the room.
2. AI Breaks When Conditions Change Even Slightly
Ask an AI system to handle a task it's seen many versions of, and it performs well. Change a small detail the training data didn't cover, and performance can fall apart, even though the task is basically the same.
That's because fluency and understanding aren't the same thing. AI models are very good at producing language that sounds right. They're weaker at grasping what that language refers to in the real world. The output looks polished because the model is fluent. It breaks easily because the model isn't grounded.
A child who watches a tower of blocks get knocked over can immediately guess what would happen to a sandcastle hit the same way, or a row of dominoes. That's not because the child has seen those exact scenarios before. It's because the child understands how physical objects behave in general, and can apply that understanding anywhere.
Today's AI doesn't generalize that way. It recognizes patterns it's already seen. A world model would do what the child does, understand the underlying dynamics well enough to handle a situation it's never encountered.
3. AI Cannot Plan and Reason Reliably Over Time
There's a difference between completing a task and executing a plan, even though both can look the same on the surface.
A task has a clear start and a clear finish, and nothing in between depends on anything else: summarize this document, draft this email, flag this anomaly. Today's AI agents are good at this. A plan is different. It's a sequence of tasks where each one changes the conditions for the ones that follow — promote this person, which changes the budget, which changes what the next hire can look like, which changes who's available to approve the next request. Get one step wrong, or fail to register what a step actually changed, and every step after it is built on a false premise.
That's what a world model is for. It lets an agent stay accurate, updating the picture of the situation as it moves through a plan — what's true now, what a given action would change, and what's actually true once that action has happened.
This isn't something more training fixes. Even the best reasoning models today still fall short of reliable planning in complex situations. The gap isn't a lack of knowledge. It's the absence of a model of the world that updates as the agent acts in it.
4. AI Has Far Less Data to Learn the Physical World
The leap in language models came from training on a staggering amount of text, essentially the written output of the entire internet. A company doesn't offer anything close to that. There's a vast gap between the data available to train a language model and the data available to teach an AI how a company actually operates.
This shows up in what's written down versus what's actually known. Think about what a quarter-end close looks like. The procedure says what gets reconciled and when. What it doesn't say:
Why intercompany transactions get reconciled before accruals, even though the procedure doesn't require that order
Which variances are normal for a given business unit, and which ones signal something is actually wrong
A world model for work would allow AI to handle a situation no one wrote a rule for.
None of that lives in a document. It lives in the judgment of the people who've closed the books before, built from watching what happens after each decision. An AI trained only on the written procedure can recite it. It has no way to know the judgment that makes the procedure actually work. A world model for work would allow AI to handle a situation no one wrote a rule for.
5. AI Cannot Be Trusted to Act Safely in the Real World
Every problem above compounds into this one. A system that can't reason about cause and effect, breaks when conditions shift, can't plan across steps, and lacks real grounding in its company's environment is not a system you can trust to take consequential action.
Action is what raises the stakes. You don't want AI that clicks buttons faster than a human if it doesn't understand what those clicks will cause. That's why world models are increasingly seen as the core requirement for any system meant to perceive, decide, and act — not just describe.
An agent that drafts a memo is producing words. An agent that approves a promotion, releases a payment, or moves a manager is taking an action with consequences. Trusting it to act requires the same thing a person in that role would need: an understanding of the budgets, policies, and controls that determine what the action actually sets in motion.
This is why some of the most prominent researchers in AI treat world models not as a technical curiosity but as a foregone conclusion. It's just a matter of time. LeCun has said publicly that within a few years, world models, not LLMs, will be the dominant architecture for AI.
What This Means for Companies Today
Enterprise software has spent decades getting good at capturing the world — transactions, headcount, approvals, policies. The five challenges above are really one challenge from different angles: an agent that understands cause and effect, holds up under changing conditions, plans across dependent steps, and learns from more than what's written down is an agent that can finally be trusted to act. A world model is what turns a system of record into an agent that can do all of that.
A world model for work is AI you can actually hand things to without constant fact-checking.
A platform sitting on decades of structured, governed organizational data has a significant head start because that data is what allows a world model for work to learn how a company actually behaves. A world model for work is AI you can actually hand things to without constant fact-checking — a plan you don't have to verify, a recommendation you don't have to check for compliance. The work doesn't just get done quicker. It gets done in a way you can trust without checking it yourself.
Report
AI Agents Are Here—But Don’t Call Them Boss
A gap between effort and impact is the real ROI problem. Discover the solution.