● AI ROI-First AX Strategy
Why Our Company Must Build a “World Model”: The Enterprise AX Strategy That Must Come Before AI Adoption
The core point of this piece is not simply “let’s adopt AI agents.”
It is about saying that if a company wants real results from AI, it must first design its own work world—that is, a world model.
No matter how intelligent AI models become, if the company’s rules, states, permissions, exceptions, and accountability structure are unclear, that AI may end up being nothing more than a machine that produces plausible-sounding words.
On the other hand, if a company’s world model is properly built, it can simulate countless failure and success cases that would be impossible in reality at almost no cost, and it can measure the ROI of AI investment far more precisely.
In particular, it is important that enterprise AX strategy is shifting toward 2027 from “how much AI is used” to “how much economic value AI creates and how productivity innovation is proven.”
Today’s core takeaway is the world model, the ESTC framework, neuro-symbolic AI, agentic workflows, and AI sovereignty for protecting enterprise competitiveness.
1. The AX paradigm is changing: ROI matters more than AI adoption rate
Enterprise digital transformation is now moving beyond simple automation.
In the past, just saying “our company uses generative AI,” “we added an AI chatbot,” or “we introduced an automation tool” could look like innovation.
But the market’s question has changed.
What matters now is not how much AI is used, but what value AI has actually created in revenue, cost reduction, customer experience, operational efficiency, and risk management.
In other words, the center of AX, or AI Transformation, is shifting from quantitative adoption to ROI-centered operation.
This change is also quite important from the perspective of global economic outlook.
Going forward, AI investment is likely to be evaluated not as a simple technology expense but as a core capital expenditure that determines corporate productivity innovation.
The problem is that simply introducing an AI model does not automatically increase productivity.
Rather, if AI is introduced before the work world is organized, confusion within the organization, unclear accountability, incorrect automation, and rising costs can all occur at once.
2. What is a world model: building “our company’s world” for AI to operate in
A world model is, simply put, a structure that explicitly designs the company’s work world so AI can understand and act within it.
Humans learn context implicitly while working in a company.
We know from experience whether a customer is an important customer, how far a refund is allowed to go, which cases require a manager’s approval, and what state truly counts as completed.
But AI does not inherently know this tacit knowledge.
LLMs may seem to understand context within sentences, but they do not accurately model the actual state and rules of a company.
So a world model is a blueprint that tells AI, “In our company, there are these entities, these states exist, only under these conditions can you move to the next step, and these constraints must never be violated.”
Only with this blueprint can AI agents move safely in a real business environment.
This is exactly the core point emphasized by Juhwan Lee, CEO of Sweet Technology.
No matter how smart an AI model is, if it does not define the world in which it will operate, it can become not a useful machine but a dangerous probabilistic machine from the company’s perspective.
3. The value of an agentic twin: even 15 million failures cost almost nothing
Once a world model is built, a company can create an agentic twin.
An agentic twin is a concept close to moving the real work world into a virtual world where AI can experiment.
The key value here is simulation.
Just as Doctor Strange in the movie simulated countless futures to find a path to victory, a company can also test countless decision paths inside a virtual work world.
If 15 million failures happen in reality, the cost, customer trust, employee fatigue, and legal risk would be unbearable.
But inside a world model, repeated experiments can be conducted at almost no cost.
For example, customer refund policies, marketing budget allocation, inventory operations, pricing strategies, and sales response scenarios can all be repeatedly tested virtually.
This makes a huge difference from the perspective of AI investment.
Before real execution, a company can identify failure patterns in advance and select the paths with the highest ROI.
Ultimately, a world model becomes both a risk management tool in the AI era and a strategic asset that strengthens enterprise competitiveness.
4. The minimum grammar of a world model: the ESTC framework
The core grammar of the world model presented by CEO Juhwan Lee is ESTC.
ESTC is an acronym for Entity, State, Transition, and Constraint.
These mean entity, state, transition, and constraint.
It may sound complex, but in practice it is the most basic framework for turning enterprise work into something AI can understand.
4-1. Entity: the object AI must act on or judge
An Entity is an object that exists in the business world.
Customers, orders, products, contracts, inquiries, refund requests, deliveries, campaigns, advertising budgets, employees, and approval documents can all be entities.
The important point is that, in the agent era, an entity is not just a simple data item.
It is both the object on which an AI agent performs actions and the subject that moves from one state to the next.
For example, an “order” entity can have states such as payment completed, preparing for shipment, shipping in progress, delivered, refund requested, and refund completed.
4-2. State: defining what state an entity is currently in
State is the current condition of an entity.
Just as water can be solid, liquid, or gas, business entities also have multiple states.
The problem is that within a company, the same state is often defined differently by department.
For example, let’s look at the state of “customer issue closed.”
The sales team may consider it closed once the final message has been sent to the customer.
The operations team may consider it closed only when the closing process has actually been completed in the system.
The finance team may consider it closed only when the cost or figures are reflected in accounting.
People can gloss over these differences, but once AI agents are involved, such differences can lead to incidents.
Therefore, state definitions must be explicitly modeled.
4-3. Transition: under what conditions can something move to the next state
Transition is the rule for moving between states.
It is the rule for whether an order can move from payment completed to preparing for shipment, whether it can move directly from shipping in progress to refund completed, or whether it must go from refund requested to pending approval.
It is easy for AI to hear a customer and reply, “We’ll process your refund.”
But it must also verify whether that order is actually in a refundable state, whether shipping has started, whether cancellation or refund is the correct action, and whether the amount exceeds a threshold.
Without these transition rules, AI may be polite linguistically but wrong in business terms.
4-4. Constraint: the company’s constitution that must never be broken
Constraint refers to global constraints.
These are not rules that apply only to a specific state or department, but principles that must be followed across the entire work world.
For example, refunds under 700,000 won may be processed automatically, but above that amount may require escalation to a human manager.
Constraints also include boundaries that AI must never cross, such as privacy, legal, compliance, and accounting standards.
Without these constraints, AI automation becomes fast but dangerous.
5. Why introducing only AI models fails: because there is no world for the company
The reason many companies fail in AI projects is not just poor model performance.
The real problem is that there is no world for the AI to stand on.
If you attach multiple AI agents in a worldless environment, what increases is not efficiency but the surface area where incidents can occur.
If each department uses different terminology, different state definitions, and different policy interpretations, the agents will also make different judgments.
CEO Juhwan Lee describes this as structural drift, semantic drift, and policy drift.
Structural drift is the problem of systems or data structures becoming misaligned.
Semantic drift is the problem of the same word being understood differently.
Policy drift is the problem of different departments applying different standards to the same situation.
In such a state, building multi-agent systems means the AI is not collaborating but acting in separate worlds.
Ultimately, enterprise digital transformation becomes more complex, and the ROI of automation declines.
6. Why LLMs are weak at failure: because they are optimized for the “happy path”
LLMs are fundamentally strong at success cases and positive responses.
On the internet and in public data, successful corporate cases are far more widely shared.
Failed internal operations, complex exception handling, and detailed rules for handling customer complaints are mostly not disclosed externally.
That is why LLMs tend to lean toward “good answers,” “friendly responses,” and “positive resolutions.”
For example, if a customer complains, the model may offer a full refund along with compensation.
That may look good to the customer, but from the company’s perspective it may be a loss and a policy violation.
The reason this happens is that LLMs do not accurately know our company’s profit structure, policies, cost structure, and accountability system.
In other words, LLMs are smart linguistically, but they cannot stably judge a company’s failure conditions and exception rules by themselves.
Therefore, the rules for handling failures should be handled not by the language model but by the world model and symbolic rules.
7. Agentic workflows are not SOPs
Many companies say they are building agentic workflows, but then they bring in existing work procedures.
But procedures, processes, and workflows are different.
SOPs are closer to standard operating procedures that humans should follow.
Processes organize work flow into major stages.
In contrast, a true workflow must include the conditions, verification, permissions, and events involved in moving from one state to another.
For example, to move from state S1 to S2, you must first verify that the current entity is in S1.
Then you must look at observable data signals and remove noise.
And then a guard must determine whether it is permissible to move to the next state.
The guard is a mechanism that checks policy conditions, system conditions, permission conditions, and accountability conditions.
If the guard passes, the entity moves to the next state and generates an event that it has moved.
That event can then activate another agent.
This is much closer to the structure of a true agentic workflow.
8. How the world model works in practice: the refund example
Let’s assume a customer says, “Please refund this item.”
The AI should not immediately reply, “The refund has been processed.”
First, it must verify whether the item the customer mentioned maps accurately to the actual order entity.
Then it must verify whether there was an actual purchase.
It must also check whether the product is non-refundable.
It must verify whether it is before shipment, in shipping, or delivered.
If it is before shipment, cancellation rather than refund may be the correct action.
If it is in shipping, the customer may need to be informed that a return label will be provided after delivery.
If a certain period has passed after delivery, refund eligibility may change.
If the amount exceeds a certain threshold, manager approval may be required rather than automatic processing.
None of these judgments should be made by LLM intuition.
The neural model can recommend the customer’s intent and the appropriate candidate next actions.
But the final judgment should be made by symbolic rules, that is, the company’s business logic.
9. Idempotency also matters: the same request should produce the same result even if repeated
One important concept in workflows is idempotency.
Simply put, it means that even if the same request comes in multiple times, the result should not be executed multiple times.
If a customer says “Please refund me” three times, the refund should not be processed three times.
Even if there is a timeout or network error in the middle, an action that has already been processed should not be executed again.
This is not something LLMs should decide.
The state machine, the workflow, and the symbolic rules should make that judgment.
If this basic structure is missing from enterprise automation, AI may speak quickly but the system will become unstable.
10. Why neuro-symbolic world models are needed
World models can broadly be divided into neural and symbolic approaches.
The neural approach is strong at understanding patterns through data learning and expanding flexibly.
However, it is difficult to explain why a certain judgment was made, and accountability is hard to trace.
By contrast, the symbolic approach can clearly express rules, states, conditions, and constraints.
It is predictable, traceable, and fits well with enterprise accountability structures.
In enterprise environments, explainable and controllable AI is more important than merely smart AI.
In areas where accountability matters—such as finance, manufacturing, commerce, healthcare, HR, legal, and customer service—neural models alone are not enough.
That is why neuro-symbolic AI is needed.
The structure is one in which the neural side handles flexible understanding and recommendation, while the symbolic side handles rule-based judgment and accountability tracing.
This combination is likely to become a core part of enterprise AI strategy.
11. Can our company build a world model too: start small and solid
Many companies worry, “How do we model our entire company?”
But there is no need to build the whole company as a world model from the beginning.
What matters is building a small and solid world first.
For example, you can start with media mix optimization in the marketing department.
Define entities and states such as ad channels, budgets, campaigns, customer inflow, conversions, and ROAS.
Then create rules for when an AI agent can adjust budgets and when a human must approve.
If this small world model works properly, it can be expanded to the commerce team, D2C operations, inventory management, and customer service.
A world model is a company asset, so it can be reused.
Once a world is properly built, it can be merged and expanded without conflict with other domains.
Using a game analogy, it takes time to build the initial map, but after that the world can be expanded like an expansion pack.
12. The sense of world models shown by games, Roblox, and Minecraft
A world model is not a completely unfamiliar concept.
If you look at games like Roblox or Minecraft, world models are already operating.
Inside the game there are entities such as avatars, items, shops, schools, regions, scores, and ranks.
There are transition rules for obtaining items, increasing scores, or changing ranks.
There are also constraints on what can and cannot be done.
Because these rules exist, countless users can have a consistent experience within the same world.
The same is true for companies.
Every domain—customer service, marketing, manufacturing, military, healthcare, commerce, HR, and more—has its own work world.
If that world is explicitly designed, AI agents can move through it stably.
13. The real core point that other YouTube channels or news outlets rarely explain well
The most important point is that we must abandon the illusion that “AI will understand our company for us.”
If a company does not build a world model, AI will internally create its own tacit world and reason within it.
That world may not accurately reflect our company’s interests, policies, accountability structure, or competitiveness.
What is even more dangerous is that the organization may gradually begin making documents in a way AI can understand, speaking in ways AI likes, and working in ways that make AI’s judgment easier.
At first, AI is introduced for automation, but at some point the structure can become one in which humans act according to AI’s tacit world.
This is what CEO Juhwan Lee means when he says that if humans do not design it, humans can become the designed objects inside a world created by AI.
From the enterprise perspective, this is not just a technical issue.
It is a matter of AI sovereignty, data sovereignty, and operational sovereignty.
If every company uses similar general-purpose models, a company’s unique differentiation can weaken.
Ultimately, corporate competitiveness is likely to depend not on the general-purpose AI model itself, but on how high a resolution our company’s domain is designed with.
14. A crucial distinction in AGI discussions: Generalization and Generality
A word that often appears in AGI discussions is General.
There is an important distinction here.
One is Generalization, meaning generalization.
The other is Generality, meaning broad applicability within a specific world.
Companies that build AI models naturally want to create generalized models they can sell to as many businesses as possible.
But the competitiveness of an individual company comes not from generalization but from uniqueness.
Our company’s customer understanding, operating style, policies, risk standards, brand tone, and decision-making structure are what create competitiveness.
Therefore, what a company should do is not chase the abstraction of a universal model, but raise the resolution of its own domain.
The work of raising this resolution is precisely the building of a world model.
15. The organizations of the future are likely to become “world-centered organizations”
In the past, people often said that those who use AI well would replace those who do not.
But now, most office workers have started using AI to some extent.
The difference in the future may not lie in simple usage ability, but in the ability to explicitly design a domain.
People who can explain their work in terms of state, rules, exceptions, constraints, and accountability structure will become important.
Working-level staff, team leaders, division heads, and operations managers can become the key designers of the AI era.
You do not need to know AI deeply.
You do not need to understand the structure of LLMs.
What matters is that you are the person who knows your domain best.
The moment you turn that domain into something agent-friendly, the person with hands-on expertise becomes a strategic player on par with AI experts.
16. A world model build roadmap companies should start right now
First, define the world mission and world scope.
This is the stage where you decide which work world to model first.
You do not need to choose the entire customer service domain; you can start with refund handling only, or not the entire marketing domain, but ad budget adjustments only.
Second, define the core entities.
Organize the objects AI needs to handle, such as customers, orders, products, refund requests, campaigns, and budgets.
Third, organize the states of each entity.
State definitions that departments understood differently must be surfaced and aligned one by one.
Fourth, design the state transition rules.
Make clear under what conditions something can move to the next state.
Fifth, define the constraints.
Make the rules for amount thresholds, approval authority, legal standards, privacy standards, and exception handling principles explicit.
Sixth, design the guard and event structure.
You need to connect AI recommendations, system judgments, and the next actions of the necessary agents.
Seventh, validate in a small area and then expand.
Rather than aiming for company-wide automation all at once, it is more realistic to create small successes and grow that world into a reusable asset.
17. The next stage after the world model: the execution world and the operational machine
A world model is the work of building the world that AI will understand.
The next stage is the execution world.
The execution world is a structure that allows multiple AI agents to collaborate on the same world, share states and rules, and carry out real work.
In academia, this is also explained as the concept of a Shared World.
And after that, the concept of an operational machine appears.
Rather than AI agents being the subject of every action, the structure can evolve into one where agents propose, the world judges, and the operational machine executes.
This trend signals that enterprise AI automation is evolving beyond simple chatbots into true operational systems.
Going forward, the AI industry trend is likely to move not only toward competition among smarter models, but also toward who can design internal enterprise execution structures more precisely.
< Summary >
Enterprise AX strategy is now shifting from AI adoption rate to ROI and productivity innovation.
The reason introducing only AI models is likely to fail is that our company’s work world has not been explicitly designed.
A world model is the work of defining the company’s world in which AI can act through Entity, State, Transition, and Constraint.
LLMs are strong on the happy path but weak at failures, exceptions, accountability, and permission judgments.
Therefore, a neuro-symbolic world model that combines the flexibility of neural models with the traceability of symbolic rules is needed.
Rather than aiming for company-wide automation from the start, companies should build and expand from a small and solid work world first.
In the future, corporate competitiveness is likely to depend less on the ability to use general-purpose AI models and more on how clearly a company’s domain is designed and how AI-friendly its operations are.
[Related Articles…]
AI World Models Are Reshaping Enterprise Automation
2027 AX Strategy and the Outlook for ROI-Centered Digital Transformation
*Source: [ 티타임즈TV ]
– 우리 회사 월드모델 만들어야 하는 이유 (이주환 스윗테크놀로지 대표)
