Over the past five years in the TO C battlefield, model capabilities have largely determined product positioning, with R&D investment and card quantity being decisive factors. However, in the AI TO B arena, models and products are merely entry tickets; scenarios, ecosystems, and the overall commercial system truly determine success.

This is an all-out war.

Author|Pi Ye, Wu Kunyan

Produced by|Industry Insight

'(They) are now titled Enterprise Efficiency Consultants and can sell Doubao Enterprise Edition,' an OPC told Industry Insight. Recently, he received a marketing push from Feishu Assistant, showing a package that includes Doubao Enterprise Edition and Feishu Pilot Edition, integrating Feishu Suite and Doubao Enterprise Edition's AI assistant functions.

During the conversation, the salesperson emphasized the functions and implementation cases of Doubao Enterprise Edition to the OPC, including Seedream image generation, Seedance video generation, and Agent operations for browser content capture and dashboard generation. Doubao Enterprise Edition, positioned as a productivity tool, is now officially stepping into the spotlight, leveraging ByteDance's entire TO B reach.

This is just one aspect of the recent office AI entry battle.

On July 30th, ByteDance initiated organizational adjustments, with Feishu, Doubao, and Volcano Engine undergoing varying degrees of restructuring and integration. The Feishu product team was merged into Doubao, while the market and sales teams were integrated into Volcano Engine. A week before this structural adjustment, Alibaba also fully integrated three Agent products—QoderWork, Wukong, and MuleRun—launching a new product called 'Qianwen Office.'

Taking the timeline further back, according to LatePost, Tencent's 'New Crown Prince' WorkBuddy has an average daily active user base in the millions. Within Tencent, Ma Huateng attends all meetings related to WorkBuddy.

AI office has become the second strategic consensus among major internet companies in the AI era, besides foundational models.

The underlying logic of this entry battle is not difficult to understand. From a market demand perspective, whether it's the AI productivity needs of C-end users or the definite willingness of enterprise customers to pay for AI TO B solutions, existing entry points like DingTalk, Feishu, and WeCom struggle to fully meet these demands. Whoever can first occupy user mindshare with a new entry point will be able to capture clear AI demands at the front end, achieve the flywheel effect of 'product-model,' and convert Token into commercial value.

In other words, this is a must-fight battle that balances AGI and AI business models. So, what will determine the outcome of this battle? Or, what do WorkBuddy, Qianwen Office, and Doubao Enterprise Edition need to do right and do fully to become the ultimate winners?

Over the past five years in the TO C battlefield, model capabilities have largely determined product positioning, with R&D investment and card quantity being decisive factors. However, in the AI TO B arena (Note: C-end productivity users are considered small B users, collectively referred to as TO B), models and products are merely entry tickets; scenarios, ecosystems, and the overall commercial system truly determine success.

This is an all-out war.

1. What do companies really want behind these new entry points?

First, let's answer the first question: What are companies essentially vying for behind these entry points?

Objectively speaking, in the few years since the AI large model wave erupted, such uniformly coordinated organizational moves are rare, involving two core products from major players: DingTalk and Feishu. Both have adopted nearly identical models, integrating products into AI brand products and formally transferring market and sales to the cloud business unit.

Behind these consistent adjustments lie two key indicators: intelligence and commerce (Token economy).

First, let's look at intelligence. At this year's Tencent AI Industry Application Conference, Yao Shunyu, who made his first offline appearance and truly propelled Tencent Hunyuan to the forefront of the domestic market, shared his understanding of model evolution—an important mechanism in current model evolution is the Co-design of 'product-model.'

In layman's terms, the closer the model's evolution direction and evaluation criteria align with the real user needs at the front-end product level, the more valuable the model's evolution becomes. Therefore, during the retraining of the Hunyuan large model under his leadership, the model R&D team even prioritized collaborating with front-end product teams like Yuanbao to address user needs.

The same logic applies to the B-end. Whether it's DingTalk or Feishu, both have, over the years, constructed a sufficiently unique domain—scenarios close to or even within enterprises. On these platforms, millions of real TO B demands pour into (flood in) daily, processed by DingTalk/Feishu to output the desired results for users.

After the merger, these will quietly transform into existing scenarios for Qianwen Office and Doubao Enterprise Edition. Based on these already operational AI TO B closed loops, on the one hand, they can optimize their intelligent experience at the product level, such as by understanding more real contexts to build more professional AI assistants, skills invocation pathways, MCP ecosystems, etc.

On the other hand, Co-design will undergo valuable updates. Compared to the diversity and aimlessness of C-end demands, B-end scenario demands are often of higher quality and come with built-in standards and evaluation systems. This high-quality 'demand-generation' link will further radiate from WorkBuddy, Qianwen Office, and Doubao Enterprise Edition to model layers like Hunyuan, Qianwen, and Doubao Seed, achieving high-quality evolution of model-side intelligence and accelerating AGI.

Next, commerce. If intelligence corresponds to ByteDance, Alibaba, and Tencent's pursuit of AGI at the model and product levels, then commerce can be seen as the most direct purpose of these organizational adjustments, which is the 'Token economy' often discussed in the industry.

Regarding the Token economy, many people mistakenly believe it's about selling Tokens themselves, assuming that the more cards (or Proxy's card quantity , proxy card quantity) one has, the higher the profit. This is not entirely accurate or comprehensive; it's an explanation for Token agents, i.e., OpenRouter's valuation.

For major internet companies, the true understanding of the Token economy is that, beyond direct sales and model coding plans (currently with low profit margins), their core model is to process underlying Tokens at different levels of the application layer, such as through consumption in proprietary products and ecosystems. By converting Tokens into usable AI products and forms, they can achieve product-based pricing based on Token costs on the one hand and ensure more Token consumption occurs within their business boundaries on the other.

This explains the scene at the beginning of the article. Behind the market and sales adjustments of DingTalk and Feishu lies a core objective of major companies: to implement the 'Tokens strategy' proclaimed at the strategic level, consolidate all TO B touchpoints, i.e., 'cloud + collaboration platforms,' under one roof, and truly shoulder a common indicator: selling Tokens (as an add-on) based on new AI products.

Some industry insiders told Industry Insight that the first step in the organizational adjustments of DingTalk and Feishu is to adjust the sales system. Previously, everyone could only sell products from their own business lines, but now 'everything can be sold,' with sales KPIs and product pricing gradually shifting towards Token consumption. 'This adjustment is positive for the sales team as a whole, meaning they can sell all products now,' they told us.

The benefits for enterprises' AI business systems are also inevitable. Taking Feishu as an example, its revenue is expected to exceed 3 billion yuan in 2025; media reports claim that since 2026, over 90% of Feishu's new customers have simultaneously purchased Feishu AI products; similar data also apply to Wukong. Chen Yusen stated that Wukong will continue to operate for some time because 'a large number of enterprises are still using DingTalk Wukong's AI capabilities.'

It's foreseeable that when enterprises with clear AI demands on Feishu and DingTalk are placed within ByteDance and Alibaba's entire AI systems, their average customer price and Token consumption are bound to skyrocket exponentially.

This is currently one of the few directions in the AI range that can see a positive commercial closed loop. Everyone must give it their all.

2. How to fight and win this battle?

Now that the objectives are clear, how to deploy forces? Or, how can one take the lead in this war?

Over the past few decades in software development, whether it's WPS, DingTalk, Feishu, or a host of SaaS companies, there has always been a lack of a clear metric in the productivity track: all products are based on subscription fees because products are standardized, and their effectiveness depends on individuals and customized services.

However, under the Token economy model, a new product standard has emerged: effectiveness. In other words, whoever can truly meet user needs and achieve user requirements will see higher usage frequency and depth of AI products, leading to increased Token consumption.

So, how to achieve 'effectiveness'?

First, a factor that ByteDance, Alibaba, and Tencent cannot avoid is model capabilities. This element, which has been regarded as the core strategy by various companies over the past five years, remains a crucial factor determining 'effectiveness.' Corresponding product capabilities include sufficiently long and understandable contexts, rapid response capabilities (reasoning), etc. For example, WorkBuddy received good industry evaluations for its 'fast response speed' during its breakthrough, which depends partly on the card quantity used and partly on the model's reasoning capabilities.

Given the current strategic positioning and model capabilities of AI office products, it's challenging for companies to differentiate significantly at this level.

However, models are not the sole determinant of success, or rather, the most intuitive manifestation of models is in general-purpose tasks. When it comes to tasks in real industrial scenarios, what truly determines effectiveness is the 'system.'

Specifically, the 'system' includes both internal and external systems.

The internal system corresponds to the accumulation of contexts for different industrial scenarios embedded at the product's core level. This accumulation is reflected in the execution of various industrial tasks through main dialog box interactions, such as marketing copy generation, customer management data, retail data analysis, etc. It is also reflected in the expert and coding capabilities brought to the forefront alongside the main page. The former consists of partially encapsulated scenarios based on the accumulation of major companies, directly reusable by enterprises and users in corresponding tracks; the latter involves coding's understanding and development of different scenario tasks.

The external system corresponds to the ecosystem. Although the current AI office entry battle is still focused on the first stage, ecosystem construction will become a crucial factor determining the value of AI office entries in the long run. Major companies need to design corresponding Agent ecosystems that meet the demands of more professional scenarios by absorbing vertical Agent enterprises in various deep industrial scenarios, such as energy, industry, and retail, thereby more perfectly satisfying the 'effectiveness' of TO B demands.

In fact, regarding ecosystem construction, the actions of various companies are also evident. For example, DingTalk and Feishu were incorporated into the Qianwen and Doubao systems partly because, besides scenarios, they already possess a sufficiently comprehensive software ecosystem. These software ecosystems are now quietly evolving into AI ecosystems alongside the evolution of various software vendors, becoming extended touchpoints for Qianwen and Doubao. Additionally, WorkBuddy is also absorbing Agent partners from various directions through 'suites' to jointly grow the Token consumption pie.

''Effectiveness' is where visible gaps exist, but that's only part of the picture. Beyond these, crucial aspects like permissions and security in TO B must be standard features in AI office products; otherwise, you won't even get a foot in the door with many enterprises, let alone achieve 'effectiveness,'' a head of a leading domestic SaaS company told Industry Insight.

He believes that integrating Feishu and DingTalk into Qianwen and Doubao is also intended to provide TO B customers with a better trust foundation and accelerate the cold start speed of Qianwen Office and Doubao Enterprise Edition.

3. The AI productivity era is fully underway

In fact, from the perspective of AI supply and demand models, the launch of WorkBuddy, Qianwen Office, and Doubao Enterprise Edition, while satisfying major companies' pursuit of their own AI intelligence and current commercial value, also signals a new rallying cry: the AI productivity era is fully upon us.

'Major companies will have a window of two to three years to develop mature applications,' Wang Xun, founder of AI marketing company Xinyao Zhichuang, earlier predicted to Industry Insight.

Recently, with the consecutive enhancements of Qianwen Office, WorkBuddy, and Doubao Enterprise Edition, general-purpose Agents have extended their capabilities to content, marketing, and browser operations. When asked if his prediction had changed, Wang Xun replied, 'Actually, not much has changed.'

This certainty does not stem from ignoring technological progress. He acknowledges that specific AI marketing products like digital humans, text generation, and image generation are likely to face impacts from the improved model capabilities and general-purpose Agents of major companies. He even predicts that the application layer will eventually be consolidated by a few products.

However, from the beginning, he did not position his company at the application layer. He believes that software is only a small part of the overall delivery. What enterprises truly purchase is not digital humans, copy generation, or video tools but a comprehensive solution for selecting channels, producing content, acquiring customers, and closing deals.

Therefore, Wang Xun's 'two to three-year window' does not simply refer to how long it will take major companies to develop similar functionalities. More accurately, it's a productization window Reserved for vertical service providers (left for vertical service providers), during which they can organize industry experience into workflows and solidify human services into repeatable delivery capabilities before general-purpose capabilities continue to expand.

In other words, as model and tool invocation capabilities continuously improve (continue to improve), simple functional encapsulations will be continuously learned by models and expressed on AI products. At this level, coding will no longer be just a vertical capability for developers but will become the underlying executor of general-purpose Agents, gradually sinking from high-value front-end scenarios into the capability foundation of AI office products.

For example, QoderWork, which was integrated into Qianwen Office as mentioned earlier, provides a coding execution foundation. This involves identifying the skills and experts truly needed by users, using this feedback to improve scheduling effectiveness, and forming a basic product flywheel.

However, models or coding cannot execute everything.

A SaaS industry insider told us, 'The data models of the previous software era mainly served forms, reports, and fixed processes. Agents, on the other hand, require the business meaning behind the data.' In other words, although data is the nourishment for AI Agent evolution, the premise is that AI can read and understand the data. Just because enterprises have accumulated data does not mean they have accumulated contexts that AI can directly use.

This is precisely the direction that most domestic SaaS companies, or rather Agent-native companies, are currently aiming for—transforming years of know-how into contextual environments for corresponding scenarios, helping companies complete the last mile of AI implementation in the form of FDE.

For example, Wang Xun's team now provides accompanying services for large clients at a monthly fee of around 100,000 yuan: In the first month, their team manually helps clients run through processes and achieve specific growth results; in the second month, they encapsulate the verified methods into standardized workflows; and in the third month, they deliver them for use by the client's own team.

This logic of 'outcome-based delivery' also places new demands on front-end traffic entry points. While traditional ISVs place greater emphasis on the distribution and integration efficiency brought by platforms, Agent companies in the new era will weigh more factors, such as whether the traffic provided by the platform is sufficient to compensate for client relationships, task data, and even to enjoy entirely new pricing power.

For WorkBuddy, Qianwen Office, and Doubao Enterprise Edition, they must not only excel in the Co-design of their own 'product-model' but also in setting up a Co-design system based on the ecosystem, creating new traffic aggregation, Agent distribution mechanisms, and a new Token-based pricing system.

Only in this way can they truly navigate the path from AI productivity entry points to AI productivity platforms, moving from their own AI business closed loops to AI business closed loops across the entire industry.

The era of AI productivity, or more precisely, the era of productivity centered around Tokens, has taken another step forward.