The evolution of productivity is never-ending.
Author | Chenwen
Source | Insight New Research Society
At the recently concluded Shanghai World Artificial Intelligence Conference (WAIC), the “AI Productivity Wonder House” booth was crowded.
A corporate manager described the need in one sentence:“You act as an intelligent customer service assistant, responsible for answering customer questions based on documents,” and then uploaded a product document.
In less than a minute, an AI customer service capable of handling professional inquiries was set up, able to respond to customer questions instantly.This scene vividly demonstrates the revolutionary penetration of AI Agents in office scenarios—no longer just a conceptual demonstration, but a “digital employee” capable of meeting KPIs and integrating into core processes.
01
From “Showcasing Technology” to “Pragmatism”
In fact, the implementation of AI in the office sector has not been instantaneous, but has undergone a gradual development process.
Looking back two years, it coincides with the emergence of ChatGPT, as the industry continues to delve into the deep waters of large model technology while exploring the application of large models in various fields, represented by Microsoft Office Copilot and WPS AI 1.0, marking the official entry of AI into office scenarios.
At this time, AI in the office was just beginning, functioning as a feature plugin, providing text generation, format optimization, and basic data analysis assistance, characterized by “passive response”—the user initiates commands, and AI executes single actions, without forming a complete task loop. We can refer to this stage as the Copilot Assistance Stage.
With the deepening application of large model capabilities in platforms like DingTalk AI and BetterYeah, AI Agents gradually exhibit characteristics of task automation and initial autonomy. By mid-2024, AI office work will enter the Agent Task Stage.
At this stage, AI can understand context based on instructions and connect multiple steps to complete tasks. A typical example is the 3300 AI assistants created by rookie employees, capable of automatically handling 80% of HR inquiries; the “Travel Inquiry AI” can generate personalized reports in 3-5 seconds, saving thousands of person-days annually. Although AI begins to undertake standardized processes, it still heavily relies on manually defined rules.
Observing the recently concluded WAIC, we can see that AI Agents have undergone a new evolution in office scenarios, transforming into “digital employees” deeply embedded in business processes and taking on responsibilities.
EHGO’s LuminaSphere adopts an “Assistant/Bag” architecture, allowing for the deployment of dedicated AI assistants by department (Finance, HR, Legal) and setting role permissions, directly interfacing with DingTalk/WeChat to push results; the practical Agent at Hebei Telecom handles over 20 financial scenario operations, reducing single scenario processing time from 2 hours to 10 minutes; Yongsheng Property uses DingTalk AI to analyze morning meeting content from over a thousand national projects, reducing management personnel from 15 to 3.
From the above cases, it can be seen that AI Agents have already acquired domain knowledge, awareness of permissions, and execution feedback capabilities.
It is worth mentioning that under the comprehensive embedding of Agent capabilities in office platforms like DingTalk and WeChat, the platform ecosystem of AI Agents has already taken shape.
Taking the DingTalk ecosystem as an example, employees of Cainiao Group have created over 3300 AI assistants in DingTalk, with “Cai Xiaomi AI” solving 80% of HR inquiries with an accuracy rate of nearly 90%, reducing the need for 30% of knowledge base administrators; Belle Fashion’s “Bailian AI” is based on DingTalk, simulating scenarios to train sales assistants, with pilot brand sales in Tianjin increasing, and its multi-group linkage model improving replenishment efficiency, enabling over 8000 stores to achieve efficient collaboration through DingTalk.
02
Threefold Driving Force and Key to Breakthrough
The continuous evolution of AI office work and its explosion this year is primarily driven by threefold forces.
Firstly, on the demand side, the rising labor costs, coupled with the need to address the “three high” pain points of high-frequency operations, high error rates, and high repetition in specific tasks, are pushing AI Agents from the laboratory into the office.
Secondly, on the technology side, the integration of LLM + RPA + low-code has broken through the bottleneck of task closure, as seen in the ISSUT screen semantic analysis technology of practical Agents, achieving a tenfold leap in understanding.
Thirdly, on the ecosystem side: platforms like DingTalk/WeChat have become natural testing grounds, with low-threshold development tools allowing business personnel to independently build Agents.
In practical applications, how does AI Agent solve the actual problems of workers? Analyzing the aforementioned cases, we can see that the current implementation of AI office work has shifted from localized efficiency improvement to the reshaping of core business, with the key to breakthroughs being “precise targeting of pain points + deep integration of technology”.
When the practical Agent was implemented in over 20 financial scenarios at Hebei Telecom, it directly addressed the “three high” pain points of financial work (high-frequency operations, high error rates, high labor costs), with its core technology integrating generative AI and traditional RPA. The fully self-developed vertical process model TARS achieves intelligent understanding, combined with screen semantic analysis technology (ISSUT), covering automated scenarios that improve efficiency by tenfold, with scenarios like procurement data retrieval achieving “second-level responses” and a labor release rate of 90%.
Belle Group’s “multi-group linkage” model, relying on the BetterYeah platform to deploy over 800 business AI nodes, has bridged the data silos between various intelligent systems, reshaping business processes
The core idea of the “multi-group linkage” is low-code + seamless system integration, allowing business personnel to quickly create AI assistants and connect ERP, CRM, and other systems through MCP (tool protocol layer), achieving a closed loop from automatic inventory monitoring to automatic replenishment.
Similar to Belle Group’s “multi-group linkage” model, the privatized AI assistant “Zhao Xiaojun” of China Merchants Securities also achieves one-stop processing of multiple office scenarios through system integration.
The highest-level implementation paradigm of AI Agents is to delve into complex decision-making and human-machine co-creation.
For example, SenseTime’s office raccoon, based on the Riri Xin 6.5 large model, has broken through the technology of “interleaved thinking chains” and can handle complex multimodal inputs, conduct deep fusion analysis, and output results in multimodal forms. In practical office scenarios, SenseTime’s raccoon can analyze complex Excel spreadsheets, construct a global analysis through multimodal thinking chains, and ultimately generate structured reports.
Its technical foundation aligns early visual and linguistic representations, enhancing perception efficiency and the depth of modality fusion, allowing AI to evolve from “executor” to “analytical partner”. Belle’s “Bailian AI” trains sales assistants through simulated scenarios, significantly increasing sales for pilot brands in Tianjin, demonstrating AI’s decision-making empowerment in unstructured scenarios.
03
Defect Repair and Ecosystem Reconstruction
From the analysis above, it is clear that the successful implementation of AI Agents is significant; however, from the actual user experience and feedback, there are still unresolved defects in the implementation of AI Agents in the office.
Firstly, there is a contradiction between development efficiency and implementation depth, as many enterprises face the dilemma of “one week to produce a demo, but six months to use it poorly” when implementing AI Agents.
In the early stages of development, sorting out workflows and requirements takes a long time, and later, due to AI’s lack of business understanding, it requires manual feeding and training like “guiding an intern,” which adds a workload. Platforms like BetterYeah have lowered the threshold through “one-sentence generation of Agents,” but the customization of complex business flows still relies on professional development.
Secondly, there is a contradiction between data fusion and system isolation. We know that enterprise data is often scattered across ERP, CRM, IoT, and other silos, and LLM cannot call key information in real-time. The development cost of traditional interfaces is also relatively high, leading to a lack of contextual support for AI decision-making. Some vendors’ AI Agent products use privatized deployment to solve this problem, but the new issue is that this design’s workflow loses the potential for cloud collaboration.
Finally, there is a contradiction between task closure and execution breakpoints. Currently, most LLMs can only generate suggestions and cannot execute final operations such as approvals or dispatches. There was a case where an automotive company missed compliance checks due to AI, leading to a batch of products being reworked, which was later resolved by solidifying the “ISO standard verification” node to achieve closure.
Poor task decomposition and delayed execution feedback hinder AI from becoming a true “responsible entity” rather than just a “suggestor”.
Through the lens of WAIC, we also glimpse the evolution direction of AI office work.
In terms of technical architecture, the “golden triangle” of MCP + LLM + Agent is becoming the new standard.
MCP serves as a standardized connection tool for data, LLM is responsible for task planning, and Agent schedules execution and feedback status. The data flow module of Volcano Engine HiAgent 2.0 is designed for this, supporting full-process automation from cleaning to optimization.
Interactions cannot be singular; they must be multimodal. Visual, textual, voice, and video interactions should not only become mainstream but also seamlessly connect.
SenseTime’s Riri Xin 6.5 achieves interleaved reasoning through “visual encoder optimization + deep narrow backbone model,” allowing humanoid robots to smoothly explain PPTs and interact in real-time; the popularity of DingTalk’s flash notes in cross-industry meeting scenarios marks a breakthrough in office interactions beyond text reliance.
In terms of practical applications, AI must elevate from “tool” to “organizational member.” BetterYeah’s Nova Agent supports agents to collaborate like human teams; HiAgent 2.0’s “digital employee dispatch station” can customize, manage, and assess AI performance.
It is conceivable that in the future, enterprises may operate in a “human director + AI execution team” model, even giving rise to “one-person companies” where entrepreneurs rely on AI teams to support core operations.
04Conclusion
In the finance center of Hebei Telecom, digital employees on screens are automatically entering invoice information into systems—this work, which once took human employees hours, now only requires a click. This seemingly small scene is a microcosm of the dramatic change in office logic, as AI Agents transition from processing spreadsheets to taking on KPIs, from executing commands to proactive collaboration, permanently rewriting the “human-machine relationship” in the office.
Just as DingTalk AI takes root and grows in property management, retail, and education, or as practical Agents save every labor hour in telecommunications, the essence of the AI office revolution is not merely the accumulation of tools but the reconstruction of production relationships. The future competitiveness of enterprises will depend on their ability to integrate the “brain” of LLM, the “hands and feet” of Agents, and the “nerves” of MCP into an organic whole—where there is no boundary between humans and machines, only the symbiotic evolution of intelligent agents.
* Image source from the internet, please contact for removal if infringing
Previous Recommendations
Who Will Dominate the AI Gaming Era? Tencent, NetEase, and Mihayou’s Technical Dark War
Mobile Manufacturers Reignite Overseas Expansion, What’s Different This Time?