Tezign's Fan Ling: Enterprise AI needs two systems—a subjective world model and an enterprise world model

QbitAI | WeChat Official Account QbitAI

A somewhat counterintuitive reality:

In the 2025 enterprise generative AI report released by MIT NANDA, there is a set of figures that are often looked at together:

Among the surveyed enterprises, 95% of enterprise generative AI pilots brought no measurable profit-and-loss return; at the same time, employees at more than 90% of enterprises are already using personal AI tools to get work done, yet only 40% of companies have actually purchased official enterprise subscriptions.

Behind these figures is actually the same question:

This morning, at the WTCC forum in Shanghai's West Bund, Fan Ling, founder of Tezign, gave the answer:

What enterprises ultimately need to fill is not a stronger model, but two world models.

The term "world model," put simply, is whether a system can answer: if I do this, what will happen next.

Applied to an enterprise, this question has at least two directions:

One outward, understanding customers. Why customers prefer a certain choice, what makes them change their minds, what kind of expression and products are more readily accepted. At this layer, Tezign's phrasing at the forum was "customer world model," which is the "subjective world model" that appears more often in later materials.

One inward, understanding the company. After a decision enters the enterprise's interior, what processes it goes through, which resources, rules, and collaborative relationships affect it, and whether it can ultimately be implemented. This part is closer to the "enterprise world model."

Both appear to be answering "how to make decisions," but they are essentially not the same thing.

The subjective world model that "understands customers" deals with subjective preferences, motivations, and choices, and relies on external user data;

The enterprise world model that "understands the company" deals with processes, assets, rules, and tacit knowledge, and relies on context accumulated over the long term inside the organization.

Regarding the subjective world model, QbitAI already did a detailed explanation some time ago, so it will not be expanded here.

And precisely because these two things differ in data sources, update cadence, and ownership, they are hard to crudely merge into one large model.

Merged together, they are often not stronger but duller on both sides; lacking either side, the problem immediately surfaces:

So this is not a matter of "one more model is safer," but that the enterprise agent architecture has inherently been missing two completely different things.

If the customer world model is closer to external perception, then the hardest part of the enterprise world model is actually the identification, organization, and utilization of internal knowledge.

Because much of what is most valuable in an organization is precisely not in documents.

At the forum, Tezign split enterprise knowledge into two categories:

In fact, quite a few enterprises are already trying to build their own context systems based on model capabilities.

But whether the context is built well or not, this layer may not yet have formed a unified criterion for judgment.

The context system an enterprise truly needs must at least satisfy several conditions:

Summed up in one sentence: what the agent calls is not data, but context.

If the enterprise world model solves internal context, then another change underway is that expert judgment is beginning to enter training data.

What is discussed here is not a third world model, nor is it equivalent to the customer world model.

It is more like a shared capability foundation beneath the two world models.

According to disclosures at the forum and related technical materials, Tezign has 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 only on general data but begin to include judgment paths in business contexts (Creative Trajectory). At the same time, some partner institutions are also advancing research in a similar direction, and the accumulation of related judgment paths has reached the order of ten thousand entries.

Why does this matter?

Because the most valuable knowledge in an enterprise, such as judgment criteria, trade-off logic, and risk sensitivity, is not on the internet, nor in general models.

It originally existed only among people, projects, and archives.

What is being done now is actually turning these judgment processes, which were originally hard to retain, into an asset that can be accumulated.

From "renting intelligence" to "sovereign intelligence."

Finally, there remains an unavoidable question: how to evaluate?

General benchmarks measure whether a model can answer the standard answer correctly.

But in enterprise scenarios, the more important question is often not "is it right," but: did it do it right in this company's way.

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 models.

So the real value of a business-vertical benchmark is not in giving a model a score, but in turning the organization's judgment criteria into a test set that can be continuously regressed, continuously compared, and continuously corrected.

Once this thing exists, what capability changes model upgrades, process adjustments, and team changes actually bring no longer need to be judged by "it feels better."

On this point, the Tezign team mentioned a very typical calibration method:

In the Atypica scenario, by having real people and AI personas complete the same economic behavior experiments, and then using the differences between the two to quantify the calibration of the simulation effect, thereby feeding back into model training and optimization.

When QbitAI reported on-site at WAIC in July this year, it mentioned a result: the subjective world model reached 85% accuracy in behavior simulation.

Of course, such simulation and benchmarks are essentially still calibration tools; they can make research and validation happen earlier, but they cannot replace real-human research itself.

What kind of mechanism does an enterprise need to build for agents to truly run?

Once the foundation is in place, how can the mechanism truly start turning?

The formulation given at the forum is that for enterprise agents to truly run, a complete closed loop must turn: Signal, Plan, Create, Evals.

Behind this actually corresponds to the difference between enterprise agents and the previous generation of Copilot.

Copilot is more like a tool invoked one time: you ask, it answers; you start a task, it gives you a result.

But the agent an enterprise truly needs is not like this. It is long-horizon, can run continuously, must enter workflows, and must retain goals, state, and feedback across time, rather than starting from zero every time.

In his speech, Fan Ling said:

The agent pushes the work to 80 points, and then people take it from 80 to 120.

The important part of this sentence is not the two numbers "80 points" and "120 points," but that it redefines the division of labor between people and agents.

The two are not a relationship of replacement, but a staggered division of labor: AI is responsible for pushing forward a large amount of work that can be tokenized, while the truly valuable part of experts remains in that final stretch of high-density judgment.

One set understands the outside, one set understands the inside, and in between runs a long-horizon agent capable of continuous advancement.

It is worth mentioning that at the WTCC forum, Tezign simultaneously released the Agentic Transformation co-building plan, bringing in universities, research institutions, enterprises, model and technology partners, and professional practitioners. The reason is also simple: this path has not yet been fully validated, many questions have no ready-made answers, and one can only grow while doing in open collaboration.

If an enterprise wants to turn its own judgment into an asset that can be accumulated, then what should the first step be?

*This article is published by QbitAI with authorization; the views belong solely to the original author.