Yunzhong, reporting from Aofeisi
QbitAI | WeChat Official Account QbitAI
A somewhat counterintuitive reality is this:
AI is already being widely used in enterprises, yet the productivity of many companies has not risen noticeably as a result.
In the 2025 State of Generative AI in the Enterprise report published by MIT NANDA, there is a set of figures that are often looked at together:
Among the surveyed enterprises,95%of the generative AI pilots brought no measurable P&L returns; meanwhile, more than90%of employees in enterprises are already using personal AI tools to get work done, while only 40% of companies have actually purchased official enterprise subscriptions.
Behind this set of numbers lies the same problem:
AI is being used, but it stops at the personal level。
So the question enterprises face today is no longer "whether or not to adopt AI," but another, harder one:
After adoption, what exactly should the enterprise accumulate on its own so that AI truly becomes productivity?
At this morning's WTCC forum on Shanghai's West Bund, Tezign founder Fan Ling gave his answer:
What enterprises ultimately need to fill in is not a more powerful model, buttwo world models。
First understand the customer, then understand the company
Simply put, the term "world model" means whether a system can answer this question:if I do this, what will happen next。
Placed in an enterprise context, this question has at least two directions:
One is outward, understanding the customer.Why does the customer prefer a certain choice, what would make them change their mind, what kind of expression and products are more easily accepted. This layer is what Tezign called the "customer world model" at the forum — what the materials later more often refer to as the "subjective world model”。
One is inward, understanding the company.After a decision enters the enterprise's internal operations, what processes does it go through, what resources, rules and collaborative relationships affect it, and can it ultimately be truly implemented. This part is closer to the "enterprise world model”。
Both appear to be answering "how to make decisions," but in essence they are not the same thing.
The subjective world model of "understanding the customer" deals with subjective preferences, motivations and choices, relying on external user data;
the enterprise world model of "understanding the company" deals with processes, assets, rules and tacit knowledge, relying on context accumulated within the organization over the long term.
Regarding the subjective world model, QbitAI already gave a detailed explanation a while ago, so it will not be expanded upon here.
Precisely because these two systems differ in data sources, update cadence, and ownership of responsibility, they are hard to crudely merge into one big model.
When combined, the result is often not stronger but dulled on both sides, and if either one is missing, problems will surface immediately:
- With only an enterprise world model, the agent knows how to do the task but not whether it is worth doing, and in the end it may justefficiently accomplish the wrong goal。
- With only a customer world model, the agent can understand the需求 but cannot land it in real organizational delivery, and in the end it becomesunderstands the demand, but cannot deliver it。
So this is not a question of "one extra model for extra safety," but rather that the enterprise agent architecture is inherently missing two completely different pieces.
An enterprise's most valuable knowledge often hasn't been written down yet
If the customer world model is closer to external perception, thenthe hardest part of the enterprise world model lies in identifying, organizing, and utilizing internal knowledge。
Because much of what is most valuable in an organization lies precisely outside the documents.
At the forum, Tezign broke enterprise knowledge into two categories:
- One isexplicit knowledge, which can be written into manuals, processes, and systems, and pursues being higher, faster, and more accurate;
- the other istacit knowledge, such as taste, judgment, a sense of boundaries, experience, and relationships, which cannot be fully expressed and can only gradually grow out of concrete projects and collaboration.
This also explains why "feeding all the documents to AI" does not equal "AI understanding the enterprise."
In fact, quite a few enterprises are already trying to build their own context systems based on model capabilities.
But as for whether the context is built well, there may not yet be a unified standard of judgment at this layer.
The context system that enterprises truly need must satisfy at least several conditions:
- It does not replace existing systems, but connects existing data lakes, business systems, and asset systems;
- It is not knowledge for humans to read, but structured objects and relationships that can be invoked by agents;
- Nor is it a one-time delivery; rather, it can leave behind review basis after each judgment, so that similar problems need not start from scratch next time.
To summarize in one sentence:what agents invoke is not data, but context。
Expert judgment is becoming training data
If the enterprise world model solves internal context, then another change that is taking place is thatexpert judgment is beginning to enter training data。
What is discussed here is not a third world model, nor is it equivalent to the client world model.
It is more like a shared capability foundation beneath the two sets of world models.
According to disclosures at the forum and in related technical materials, Tezign has already brought in a large number of experts from different fields to participate in the collection and annotation of real judgment processes, so that training signals no longer rely solely on general data, but begin to include judgment paths within business contexts (Creative Trajectory). Meanwhile, some partner institutions are also advancing research along similar directions, and the accumulation of relevant judgment paths has reached the scale of10,000s of entries。
Why does this matter?
Because the most valuable knowledge in enterprises, such as judgment standards, trade-off logic, and risk sensitivity, exists neither on the internet nor in general-purpose models.
It originally existed only among people, projects, and archives.
What is being done now is, in fact, turning these judgment processes, which were originally hard to preserve, into an asset that can be accumulated.
Therefore, the path ultimately leads not only to "models understanding the business better," but to a clear direction of enterprise AI evolution:
from "rented intelligence"toward "sovereign intelligence"。
What's truly hard to evaluate is not whether the answer is correct, but whether it "resembles this company".
Finally, there remains one unavoidable question: how to evaluate?
Generic benchmarks measure whether a model can answer with the standard answer.
But in enterprise scenarios, the more important question is often not "whether it's right", but rather:whether it got it right in the way this company does things。
A company's brand boundaries, a team's product selection logic, a manager's tolerance range for risk—none of these are in public evaluation sets, nor in the default values of general-purpose models.
So where business-vertical benchmarks are truly valuable is not in giving a model a score, but in turning an organization's judgment standards into a test set that can be continuously regressed, continuously compared, and continuously corrected.
Once this thing exists, whatever capability changes model upgrades, process adjustments, or team changes actually bring about no longer has to be judged by "it feels better".
On this point, the Tezan team mentioned a very typical calibration method:
In Atypica's scenario, real people and AI personas are made to complete the same behavioral economics experiments, and the differences between the two are used to quantitatively calibrate the simulation effectiveness, which in turn feeds back into model training and optimization.
When reporting live at WAIC in July this year, QbitAI mentioned one result: the subjective world model achieved an accuracy rate of 85% in behavioral simulation.
Of course, these simulations and benchmarks are still fundamentally calibration tools; they allow research and validation to happen earlier, but they cannot replace real human research itself.
The Start of Agentic Transformation
For AI agents to truly run inside an enterprise, what kind of mechanisms need to be built?
Once the foundation is in place, how can these mechanisms be set truly in motion?
The answer given at the forum was that for enterprise agents to truly run, a complete closed loop has to turn:Signal, Plan, Create, Evals。
- Signal and Evals rely more on the customer world model; they answer: what do users care about right now, and will customers actually buy what gets made.
- Plan and Create rely more on the enterprise world model; they answer: whether this can actually be done within the organization, and how to advance it given the resources, processes, and constraints at hand.
Behind this lies the difference between enterprise agents and the previous generation of Copilots.。
A Copilot is more like a tool invoked one time: you ask, it answers; you start a task, it gives you a result.
But the agents that enterprises truly need are not like that. They are long-running, can operate continuously, must enter the workflow, and must retain goals, states, and feedback across time, rather than starting from scratch every time.
Fan Ling said in his speech:
The agent pushes the work to 80 points, and the human then takes it from 80 points to 120 points.
The importance of this statement lies not in the two numbers, "80 points" and "120 points", but in how it redefines the division of labor between humans and agents.
The two are not in a replacement relationship, but in a staggered division of labor: AI is responsible for pushing forward large amounts of tokenizable work, while where experts truly add value remains in that final stretch of high-density judgment.
Putting these threads together, enterprise-grade AI has entered a new stage, and what the "two world models" offer is precisely the expression of this set of capabilities within an enterprise architecture:
one to understand the external, one to understand the internal, and in between, a long-running agent that can continuously push forward.
It is worth mentioning that Tezign simultaneously launched theAgentic Transformation Co-Build Programat the WTCC forum, bringing universities, research institutions, enterprises, model and technology partners, and professional practitioners into it. The reason is simple: this path has not yet been fully validated, many questions have no ready-made answers, and growth can only happen through doing amid open collaboration.
Rather than "how to put AI to use", Tezign is trying to answer a much harder question through this co-building approach:
For an enterprise to turn its own judgment into an accumulable asset, what should the first step be?
Reference link: https://game.atypica.ai/
*This article is republished by QbitAI with authorization; the views are those of the original author.