WPS Lingxi, a completely different Agent sample.

At the WAIC 2025, which opened on July 26th, the exhibition area was crowded with visitors, creating a bustling atmosphere. In contrast, there were far fewer sensational stories and viral hits about AI products this year.

The "low - key" nature of these products is a microcosm of the current stage of large - language models. In the initial stage of simply piling on parameters, many eye - catching products soared in popularity but quickly faded away. At this year's WAIC, it was evident that "practicality" had replaced "sensationalism." Whether a product can effectively meet user needs has become the focus of AI manufacturers.

For example, at the booth of Step Star, one of the large - language model unicorns, in addition to the regular AI writing functions, it also showcased the "in - depth verification" function, which specifically addresses the issues of untraceable content and false information in AI - generated works. Mobvoi launched the AI hardware TicNote, which is only 3 millimeters thick, can be magnetically attached to a mobile phone, and comes with functions such as voice - to - text transcription and thought organization that are integrated with the phone software.

In this context, Kingsoft Office also released its new "down - to - earth" products, including the native Office intelligent agent WPS Lingxi and a series of related products, such as WPS AI PPT, WPS Personal Knowledge Base, and Lingxi Voice Assistant.

Nowadays, the term "Agent" is frequently mentioned, but Kingsoft Office has proposed a different approach from other Agent - based products on the market.

Tian Ran, the assistant president of Kingsoft Office, said that there are currently two situations for AI software. One is that when users interact with software, the AI capabilities are hidden behind the software and cannot be fully utilized. The other is that when users interact directly with AI, which then calls software, users often struggle to clearly articulate their needs.

"Therefore, Kingsoft Office's approach is to integrate and transform software and AI with each other. The next - generation AI software is Agentic Software," Tian Ran stated.

In essence, Agent is a capability rather than a final product form. It is a long - term process for users to change their behavior patterns to adapt to Agent. As software products that are closer to users, they should evolve themselves to help users adapt to Agent.

Shifting from "Function Acceleration" to "Process Reengineering"

Using large - language models to create content such as articles is one of the most common applications of AI. However, there is an underlying pain point in this scenario - whether it's text or images, what's the probability that you'll directly adopt the first result generated by a large - language model? Almost zero.

Therefore, in reality, for content generation, users need to have continuous conversations and make modifications, asking the AI to deliver a second, third, and subsequent versions... This basically mimics the scenario of a user communicating with a colleague to revise content: not only do the two need to chat in real - time, but the document also needs to be in front of both of them.

This "colleague - like collaboration" interaction logic is the breakthrough of WPS Lingxi. Different from the one - way "one - sentence generation" mode of traditional AI, WPS Lingxi enables interactive, modifiable, and multi - round conversations.

During the on - site demonstration, Lingxi was presented in a form of on - screen collaboration, with the Office suite on the left and the AI assistant on the right. When users call Lingxi within the WPS components and directly state their specific requirements, the AI can modify the document area on the left, show the modification traces, and keep the document format intact.

Moreover, when making modifications, Lingxi can intelligently determine the user's intention and the scope of modification, and precisely modify only the necessary parts, avoiding accidental operations and unnecessary full - text reformatting.

In addition, Kingsoft Office also launched WPS AI PPT, which also adopts the interactive, modifiable, and multi - round conversation creation mode. Users can modify the PPT outline while chatting with the AI, and every node and step is controllable.

A differentiating feature is that in Kingsoft Office's AI PPT, if users want to make detailed secondary modifications to the AI - generated PPT, the templates at the modified positions will automatically remain consistent, eliminating the need for manual "framing."

Lu Weijun, the product manager of WPS AI PPT, said that the PPT modules created by other AI PPT products on the market are freely stacked objects, while WPS Lingxi has a memory of the PPT's format and structure, so it can make modifications based on a stable structure.

The memory function is one of the most practical features of AI chatbots. It makes conversations more personalized and allows users to bypass the "prompt - word learning" threshold. If one sentence can't clarify the need, users can have more conversations, using natural language to gradually make the AI understand their requirements.

Looking at the industry as a whole, although domestic office Agents have been booming this year and products with single - point capabilities are competing in terms of technical indicators, when these indicators are soaring, the "real needs" of a large number of users remain unmet. For example, it may only take one second to generate a document or PPT, but it still takes days to make modifications.

This situation is because the intelligent transformation of most office software has not penetrated into the workflow, and most AI manufacturers have not identified the most time - consuming and labor - intensive parts for users.

WPS Lingxi's shift from "function acceleration" to "process reengineering" is a practical approach to solving these problems.

The "All - Rounder" of Office Intelligence?

It's not surprising that Kingsoft Office has developed such products in the highly competitive office software market.

As an office software company with a 37 - year history, Kingsoft Office has accumulated a large number of high - quality office scenarios and has control over three major software components. Its corporate DNA determines that it starts from the actual needs of users in specific scenarios, rather than simply engaging in a technological race.

In essence, Kingsoft Office "adapts technology to scenarios" rather than "adapting scenarios to technology."

At the current stage of transitioning from training to inference, it's a good opportunity for software companies familiar with scenarios to enter the market. Tian Ran pointed out that the so - called "vertical models will replace underlying models and experience rapid growth this year" is a false concept. More accurately, it's the vertical scenarios that are growing.

When it comes to application implementation, those with a deeper understanding of scenarios will grasp one of the keys to success.

Another key factor is engineering capabilities. Technically, after the development of underlying large - language models reaches a relatively stable state, if you want to further improve the product, it all comes down to engineering capabilities.

With a deep understanding of scenarios and strong engineering practice capabilities, Kingsoft Office has been able to develop products that others can't.

Taking AI PPT as an example, Tian Ran said that at the engineering level, it's much more difficult for AI to create a PPT than to create HTML and Markdown because there is insufficient PPT data in the materials used to train large - language models.

So, how can we enable AI to demonstrate its typesetting ability and creativity while also creating according to human requirements, that is, both unleashing and controlling AI? This is the main problem that the WPS team is working on.

"Both aspects require a large amount of engineering work. We need to develop a new set of APIs specifically for AI within the components, and when the AI calls these APIs, its knowledge should also be able to interact with these APIs," Tian Ran explained.

Teaching AI to communicate deeply with software is one of the key technological barriers in Kingsoft Office's Agentic Software strategy.

In terms of understanding scenarios, Tian Ran also put forward a view different from the market consensus. "Many people think that an Agent replaces humans and acts as a digital employee, but we believe that an Agent serves each employee. It doesn't directly handle requests but works with employees to address them."

Kingsoft Office's product launch this time is very practical, focusing on solving user problems rather than showing off technology. However, judging from the effective product results, it's actually another form of "showcasing skills" - the accumulation over the past three decades is truly reflected in the products, which may be difficult for many competitors to surpass.

Lingxi is not perfect. According to the WPS team, an AI product can never be "perfect." "An AI product is not something that is launched; it grows," Tian Ran said. The launch of an AI product is just the starting point, and subsequent iterations and growth are more important.

"In the future, we'll not only modify the outline while chatting but also modify templates, covers, and each page of the PPT while chatting," Lu Weijun said.

This is a long - term journey. WPS Lingxi provides an industry example at the starting point. Perhaps after this, the battle for office intelligence is just beginning.