Kuaishou Breaks Down the Enterprise AI Productivity Practice Path: From R&D Efficiency Gains to the iWork Organization

From engineers to all employees, from individuals to the organization: Kuaishou breaks down the practical path of "enterprise AI productivity" Phoenix Tech Download the app | Exclusive first look October 22, 2026, 19:39:38 From Beijing

Phoenix Tech News, October 16 — At the QECon Global Software Quality & Efficiency Conference, Wang Xu, Vice President of Enterprise AI Productivity at Kuaishou, delivered a keynote speech titled "Kuaishou Enterprise AI Productivity: From R&D Efficiency to Organizational Leap," systematically reviewing Kuaishou's practical path over the past two years, from R&D efficiency improvement and AI for all employees to the enterprise-level AI productivity organization iWork.

In his speech, Wang Xu said that individual efficiency gains do not equal organizational efficiency gains. As large model applications move from personal tools to enterprise deployment, how to convert individual efficiency into organizational efficiency is becoming a key question for large enterprises' AI transformation. In September this year, Kuaishou established the "Enterprise AI Productivity" (iWork) organization, attempting to consolidate AI practices scattered across teams into company-level capabilities and to coordinate the unified development of general-purpose Agents, enterprise context systems, and enterprise information systems.

"I give you a Ferrari, but if you drive it on a road full of red lights, it's useless." Wang Xu used this metaphor to describe Kuaishou's conclusion: if collaboration chains such as R&D, design, and testing do not change in sync, efficiency gains in a single link cannot translate into overall delivery speed. In his view, R&D efficiency is not a single-point problem but an organizational collaboration problem.

Over the past two years, Kuaishou's AI productivity practice started first from R&D scenarios. In early 2025, Kuaishou brought together the CTOs of its various business lines to form the "CTO Efficiency" mechanism, discussing how AI can empower R&D once every two weeks, and distilled a "trinity" strategy: first, individuals continue to practice; then, the advanced capabilities distilled from practice are solidified; finally, the organization drives transformation from the top down.

From 2025 to the first quarter of 2026, Kuaishou as a whole introduced AI tools following the classic R&D chain of requirements, design, coding, testing, and operations and maintenance. Wang Xu introduced that Kuaishou observed that engineers' use of AI is gradually moving from code snippet assistance toward task-level development and full-chain development. Internally, this process is divided into three levels, L1, L2, and L3: L1 is snippet-level code modification, where engineers hand code snippets to AI for modification and then embed them back into the codebase; L2 is task-level delivery, for example, defining a database table, where after describing the requirement, AI generates the complete statement; L3 is handing the PRD or Spec directly to AI and using the SDD approach to guide it through autonomous full-chain development.

Based on the consideration that the foundation models and Harness capabilities at the time were not yet mature, Kuaishou did not directly pursue L3 full-chain automated development, but instead chose L2 coverage as a phased metric, with the goal of reaching 80% by the end of 2026. On the tool side, Kuaishou connects each link through products such as the Coding IDE, the requirements and CR management platform KATE, and the heavy-duty coding assistant Kwaipilot, embedding AI capabilities into existing processes.

In the second quarter of 2026, the craze for the lower-barrier general-purpose agent "Crayfish" showed Kuaishou a bigger opportunity for AI to move from R&D personnel to use by all employees. Kuaishou transformed and packaged its internal programming assistant CodeFlicker into the general-purpose agent MyFlicker, inserting a large amount of domain knowledge and Skills and opening up more scenario capabilities to all employees. HR can use the agent for recruitment and job-person matching, while the administrative team has integrated services such as meal ordering, shuttle bus queries, and haircut appointments into it, so that ordinary employees can also directly enjoy AI-enabled services.

During the rollout, Kuaishou also innovatively designed a Token points system, issuing points to employees, who can choose different models to consume on the general-purpose agent and arrange their usage themselves. Wang Xu spoke about the original intent behind the system's design, which had two main goals: first, to encourage employees to prioritize high-quality products; second, to avoid duplicate development, since it is clear at a glance whether an AI product is actually being used and how many Tokens it consumes. Through these efforts, MyFlicker's weekly active penetration rate within Kuaishou has reached about 70%.

In the process of promoting "AI for all," Kuaishou also held a series of activities such as the "AI Freeloader Assessment Officer," the "Company Anniversary AI Snake Game Competition," and the "PDE Assembly Call," in order to lower the barrier to AI use, help employees improve their AI awareness and practical skills, and verify the possibility of non-technical people building tools. Among the top three in the AI Snake Game Competition, the third place came from a non-technical team; the PDE Assembly Call ultimately received more than 4,000 submissions. Wang Xu believes that AI is redrawing the boundaries of capabilities within organizations, and tool production that was previously mainly done by engineers is opening up to more business personnel.

After the "AI for all" craze, Kuaishou did a review in H1. The good news was that AI usage was quickly ramped up, awareness initialization was basically completed, and a large number of Skills were accumulated internally. But as AI use spread rapidly, new problems also began to emerge.

Wang Xu and his team found that after "AI for all," three types of problems remained: Skill quality varied widely, with some becoming "Token black holes" that consumed thousands of points; individual efficiency gains were limited by existing workflows, with PRDs written in half a day but the downstream collaboration chain unchanged; and individual AI capabilities differed greatly, making it difficult for expert experience to be consolidated into team assets.

In response to the above problems, Wang Xu abstracted the issue into an "individual efficiency × organizational efficiency" coordinate chart during the mid-year review. The core judgment was that if the organizational level does not change, the ceiling for individual efficiency remains low; having only super-individuals without collaboration will create efficiency islands. The truly valuable path is to improve individual capabilities and organizational capabilities in tandem.

Kuaishou thus further clarified the evolution path of AI productivity: first complete the cold start of employee usage, then consolidate expert experience into a team baseline, subsequently transform team public knowledge, business assets, and collaboration rules into context that Agents can consume, and ultimately make Agents independent work units that enter collaboration chains of human-Agent and Agent-Agent.

In the third quarter of 2026, Kuaishou shifted the focus of AI efficiency from individuals to teams. On one hand, it pushed knowledge base construction, consolidating past projects, code, and core workflows onto a public infrastructure platform so that Agents can directly consume team context. On the other hand, it encapsulated public knowledge and Skills into Agents capable of working independently, and through a super Agent responsible for global coordination, registered sub-Agents such as product managers, designers, and full-stack developers into the pipeline, realizing a collaboration network of continuous 7×24 iteration. At the same time, safety gates for quality, permissions, and hallucinations were set up; when Agents encounter risks or uncertain issues, human intervention is triggered.

"Since April, the team has completed more than 2,000 tasks in total, with an AI first-pass task rate exceeding 60%, and 80% of tasks can be completed within two passes, that is, one AI pass plus one human calibration," Wang Xu revealed. Non-technical scenarios are also being advanced in parallel. The content operations team has integrated the relevant knowledge base and high-frequency workflow Skills behind Flicker for tasks such as hot-topic breakdown and content analysis, and is promoting its use across the content operations line. The value of AI is moving from "personal assistant" to "team-exclusive productivity tool."

From "Team" to "Organization": Moving Toward Enterprise AI Productivity (iWork)

As AI practices continued to emerge across teams, new challenges also arose: how can excellent experience become company standards? If every team builds its own AI applications based on its own business data and system exits, it may solve problems quickly in the short term, but in the long term it may lead to duplicate construction, fragmented enterprise context, and security risks.

Facing these problems, Wang Xu regards the establishment of iWork as the start of the fourth stage of Kuaishou's AI productivity path. He said Kuaishou has already completed three stages of this path and has just begun exploring the fourth stage. In September this year, Kuaishou established the "Enterprise AI Productivity" (iWork) organization, integrating the company's internal communication platform, online document collaboration, general Agent, and internal information system capabilities, and advancing AI practices scattered across teams into company-level capability building. Its construction approach includes two directions: first, build a stronger general Agent with stronger reasoning capabilities while being secure, controlled, and monitorable, continuously improving individual employee efficiency; second, establish a unified enterprise semantic framework to organize scattered information such as documents, data, code, process states, and meeting records into context that Agents can understand and invoke.

Wang Xu pointed out that in the past many Agents consumed local data within their respective systems, which easily formed "small chimneys." iWork aims to solve this by enabling Agents to see a more complete enterprise context. In the future, underlying information systems will gradually open callable interfaces, allowing AI to truly enter business processes. Kuaishou will also empower business through the FDE approach, focusing on exploring more scenarios such as content operations, content understanding, security, and livestreaming.

At the end of the speech, Wang Xu summarized and emphasized that R&D efficiency has never been an individual issue, but an organizational collaboration issue. The efficiency of a system is ultimately limited by its slowest module; only by connecting AI technical capabilities, internal business scenarios, and organizational collaboration mechanisms can AI truly enter business processes and become a force driving organizational efficiency upgrades.

It is worth mentioning that, to give the industry a more intuitive understanding of this complete path from frontline practice to organization-level leap, Kuaishou will hold a technical salon on October 31 themed "From R&D Efficiency to Enterprise AI Productivity," further disclosing the architecture breakdown and implementation cases of iWork.

"Special statement: The above content (including videos, pictures, and audio if any) is uploaded and published by users of Dafeng Hao, a self-media platform under Phoenix Net. This platform only provides information storage space services.

Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services."