Text by She Zongming

Although the year is not yet halfway over, 2026 is destined to be a year of dramatic change for AI assistants.

If we view the “AI Red Packet Wars” at the beginning of the year as a buildup and the subsequent “lobster craze” as a breakthrough into the mainstream, then the successive moves by two iconic AI assistants are signs of a qualitative shift in the industry.

These two AI assistants are Doubao and Wukong. The former is China's leading AI-to-Consumer (AI2C) player, while the latter is a benchmark enterprise-grade AI-native work platform.

Last night, Alibaba released its latest financial report, revealing that its full-stack AI technology investment has officially moved beyond the initial cultivation stage and entered a cycle of positive, large-scale commercial returns. In the enterprise AI sector, the enterprise-grade Agent platform “Wukong” has gradually scaled up its deployment in the near term.

About 10 days ago, Doubao began testing a tiered paid subscription model, with premium features targeting complex tasks and productivity scenarios, including advanced needs such as PPT generation, data analysis, and film and television production.

One is a personal AI assistant raising its fees, while the other is an enterprise AI assistant undergoing large-scale validation. On the surface, they seem unrelated.

However, if we apply Newton's mindset—that “by considering simple things very complexly, new fields can be discovered; by viewing complex phenomena very simply, new laws can be found”—and look at many things and phenomena together, new signals can be discerned.

Doubao charging fees and Wukong scaling up convey at least two signals:

1. AI assistants are diverging, with distinct paths for personal efficiency and enterprise efficiency.

2. As high-productivity scenarios become the key focus for AI commercialization, the value of enterprise AI is being recognized.

01

What should AI competition ultimately focus on? From the internet lexicon, many point to “traffic first.” However, Jensen Huang has offered a new perspective: “Tokens reign supreme.”

In essence, Doubao charging fees and Wukong scaling up are both branches sprouting from the trunk of “Token economics.”

Under the diseconomies of scale, where “more users mean higher costs,” Doubao must keep track of its Token usage.

Doubao's response is to target high-value computing power consumption and charge personal users.

Without shifting from a strongly entertainment-focused ChatBot used for chatting, poetry writing, and answering questions to an efficient productivity tool, it cannot balance its return on investment.

Wukong also has to account for Tokens: In the grand strategy outlined by Alibaba in March this year—“create Tokens → distribute Tokens → apply Tokens”—Wukong serves as the primary consumption engine for the B2B side.

It was inevitable that Wukong would scale up, bringing its capabilities from demo showcases to real business battlefields. After all, it must support Alibaba's Token ecosystem by establishing a commercial foundation on the B2B side.

It should be noted that both Doubao and Wukong are heavy Token consumers and will inevitably compete for the “entry point to human-machine collaboration platforms for carbon-based employees,” but they are taking different approaches.

While Doubao charging fees undoubtedly accelerates its transformation from a mass-market chat assistant to a “universal productivity tool + lightweight enterprise AI service provider,” it remains, overall, a personal efficiency tool.

Wukong scaling up, on the other hand, means moving from a group of early-adopting enterprises to countless businesses across all industries. From its inception, Wukong was not designed for “chatting” but for “getting work done”—it has always been an enterprise efficiency tool.

Now, both Doubao and Wukong have taken significant steps, demonstrating the trend of divergence among AI assistants.

In the past, many believed that AI assistants were all competing in the same arena, differing only in intelligence levels.

But now, a dividing line has emerged, with AI assistants heading down two distinct paths: one flowing into the vast ocean of personal work and life scenarios, the other diving into the rock layers of enterprise workflows.

Seeing this, you might recall Silicon Valley's “archrivals”—ChatGPT and Claude. Indeed, a similar divergence is unfolding in Silicon Valley:

ChatGPT follows a “personal super entry point” route, enhancing search, integrating shopping, and optimizing multimodal chat experiences, all with a focus on the consumer side. Although it offers features like ChatGPT for Excel and meeting assistants, its overall emphasis is on improving personal efficiency.

Claude, meanwhile, delves deep into enterprise scenarios. Artifacts enable AI to directly generate interactive documents and code, Computer Use allows AI to operate computers for complex tasks, and the Projects feature provides collaboration spaces for enterprise teams—all aimed at boosting enterprise efficiency.

Chinese AI assistants have their own “enterprise context,” but a cross-comparison cannot help but evoke a sense of déjà vu: The AI development paths on both sides of the Pacific may not repeat, but they certainly rhyme.

02

Whether focusing on personal efficiency or enterprise efficiency, there is a need to build To Prosumer (professional producer) Agent platforms to enhance office productivity.

This raises a question: What kind of Agent platform best improves human-machine collaboration efficiency for carbon-based employees?

ChatGPT and Claude can serve as reference points here.

ChatGPT's strength lies in its general capabilities and ecological breadth. Its native Context density is weaker than Claude's, relying more on “external RAG + plugins” for supplementation.

Claude's core strength lies in its extreme polishing (extreme refinement) of native Context density: a million-scale native context window, low Prompt caching costs, Claude Code, Dreaming memory mechanisms, and more, which form its competitive moat.

The result is that ChatGPT, with its “strong model + multiple plugins + massive user base,” excels in general scenarios. Claude, with its “long window + memory governance + native scenario integration,” has deeper barriers in deep knowledge work and complex business process scenarios.

The key points here involve two keywords: model capability and native Context density. Model capability is straightforward; native Context density sounds abstract and is, in fact... not particularly easy to understand either.

Let me put it this way: Your organization's relationships, supply chain information, database records, historical work orders, product manuals, operational audit logs, real-time business processes, and other information combined make up the Context. Native Context density refers to the richness, accuracy, structure, and reusability of the Context that a platform can provide to an Agent without relying on “external plugins,” solely through native integration.

Although the actual effectiveness of an Agent = model capability × Context quality, considering that gaps in model capability can be bridged through technological iteration and API access, high-density Context—which enables Agents to stably understand business and accurately execute tasks—requires long-term accumulation. Thus, the core barrier for Agent platforms lies not in model capability but in native Context density.

Among domestic AI assistants, C-side-oriented products (such as Doubao and DeepSeek) share a common trait: Meeting professional needs relies on general model capabilities and a rich plugin ecosystem.

Take Doubao, for example. Although Doubao and Feishu are both under ByteDance, they are “decoupled” products. Theoretically, Doubao can integrate with Feishu, but this integration is external: Doubao serves as the front-end general Agent entry point, while the back-end supplies the Context.

If Doubao wants to access industry data from Feishu to generate an operational report, it must go through steps like API interfaces, permission reviews, and data format conversions, resulting in a lengthy process.

Wukong takes a different approach: It is built into DingTalk, combining the “Agent entry point” and “back-end context middleware” into one. This means Wukong does not need to “look outward.” By CLI-fying all of DingTalk's underlying capabilities, Wukong can natively call enterprise data such as attendance, approvals, meetings, documents, and projects.

This enables zero-friction access—users no longer need to export data or switch software, and data is spared from loss during transfers, dock (docking), and pushbacks. It also ensures high permission controllability, keeping data within enterprise boundaries, and enables executable closures, where AI can not only provide answers but also send notifications, modify data, run processes, and make deliveries.

To have AI help draft a project review report, with Doubao, you would first copy meeting minutes and project materials from Feishu, clean up redundant formats, paste them into Doubao, wait for generation after providing prompts, and then copy the output back into a document for formatting. With Wukong, you could simply say, “Find the follow-up records in the multi-dimensional table (multidimensional table) from last Wednesday's DingTalk meeting minutes, combine them with ×× project data, and help me draft a review report and send it directly to my DingTalk document.”

One helps you do things faster; the other helps you get things done directly. There is no superiority here, but rather differences in scenario adaptation (adaptation): Doubao's speed and lightness are better suited for individuals and small teams handling standardized, high-frequency tasks; Wukong's depth and closed-loop nature are better suited for addressing complex enterprise pain points.

03

To some extent, Doubao charging fees is also an “indirect proof” of the value of enterprise-grade AI—the signal behind the signal it releases is not that “AI is now pay-to-use,” but that AI assistants only hold core value when they can meet advanced productivity demands.

And enterprise scenarios are precisely the main battlefield where advanced productivity demands are most intense and concentrated.

Wukong might say in response: Now you're speaking my language—when it comes to enterprise scenarios, many AI assistants are novices who need external coaches and floaties before they dare to take the plunge. Wukong, however, is a natural swimmer, born for the water, and thus “feels at home.”

From this perspective, the reason behind Doubao charging fees is not unrelated to the reason behind Wukong scaling up; there is a potential connection between personal AI assistants moving toward commercialization and enterprise AI assistants entering large-scale validation.

It should be noted that, at present, the most valuable AI scenarios are not for “chatting” but for “getting work done”—this has gradually become an industry consensus.

On the same day Doubao tested paid subscriptions, both OpenAI and Anthropic announced the establishment of on-the-ground service companies, sparking a close-quarters battle for AI's “last mile” into enterprise scenarios. Earlier still, OpenAI—which has traditionally focused on the consumer side—released enterprise workflow agents, Workspace Agents.

It is foreseeable that the battle among AI assistants in the B2B sector has only just begun, with enterprise-grade Agents inevitably sparking a contest for entry points into human-machine collaboration platforms.

This battle will not be won by parameters, daily active users, or hype but by who can first be validated in real enterprise scenarios.

And when it comes to whether an enterprise-oriented Agent is effective, businesses themselves have the final say. Non-standardization and low fault tolerance are inherent characteristics of enterprise scenarios. Only if an Agent is sufficiently useful, practical, and easy to use will enterprises pay real money for Token consumption.

Usefulness, practicality, and ease of use are not abstract concepts but are reflected in many enterprise-specific details: If I were an employee or boss with a need to monitor industry competition, could AI automatically crawl competitor data to generate a market radar? If I needed user insights, could AI perform intelligent analysis of massive UGC comments to precisely dissect user pain points, itches, and needs? If I required daily operational reviews, could AI automatically generate daily operational analysis reports, providing recommendations to double down/cut losses and early warnings on operational risks?

These are not groundless scenarios. Youkela, an enterprise in Yiwu, Zhejiang, once had these exact needs: As a domestic hidden champion in starry sky lamp categories and an “integrated industrial and trade” enterprise with “front-end stores and back-end factories,” Youkela has fewer than 80 employees but covers R&D, manufacturing, and multi-platform e-commerce operations. With long supply chains, a shortage of manpower, and high pressure,

CEO Wei Jun's solution was to adopt new technologies. In 2017, he moved the company onto DingTalk; in 2025, he became one of the earliest deep users of DingTalk AI Tables; and in March of this year, he became one of the first “monkey raisers.”

Sure enough, Wukong delivered reliable solutions: For market monitoring, there was no need to rely on manual tracking—AI could do it. For product R&D, there was no need to rely on gut feelings—AI could provide evidence. For operational reviews, there was no need to check data platform by platform—AI could automatically aggregate data across the board. The corresponding Skills were “Online Competitor Radar,” “Product R&D Guide,” and “Store Inspection Daily Report.”

A more tangible change: The company has only one HR, who used to spend two full days calculating salaries. After setting up a Skill, the HR could “teach” Wukong clock-in records, attendance rules, and enterprise-specific scenarios, and it would automatically clean and convert the data, reducing salary calculation time to less than 10 minutes.

This is not an isolated case but a microcosm of Wukong's early completion of practical validation across multi-industry enterprise scenarios. Such validation is not a demo but is backed by real business, real efficiency gains, and real revenue increases as “proof.”

In enterprise scenarios, whether AI takes one extra step or one fewer step in extracting, summarizing, and analyzing data is a true test of Agent capability. As Wukong scales up, more real enterprise users will inevitably use, provide feedback on, and refine its AI services within real workflows. Moving forward, whether a virtuous cycle of “more enterprises using it → better usability → even more enterprises using it” can form remains to be seen.

Usefulness, practicality, and ease of use are what truly matter, and this applies to both Doubao and Wukong. In essence, Doubao and Wukong are two mirrors of AI assistants. Neither is superior nor inferior; what matters is solving real problems for individuals and enterprises and validating AI's value.

Against the backdrop of high-productivity scenarios becoming the focus of AI implementation, it is especially important to deepen the value of enterprise AI assistants and improve their closed-loop capabilities—on a small scale, this involves cash out (fulfilling) the value of enterprise AI; on a large scale, it concerns the development height of China's AI productivity.

Time may be silent, but results speak volumes.