中国互联网巨头全面转型：以AI工作台重塑业务架构

NextFin News — China’s large internet groups spent the first phase of the generative-AI boom adding chat boxes and side features to products they already owned. They are spending the next phase rearranging the companies behind those products. Separate teams that once chased overlapping agent ideas are being merged. Familiar apps—documents, meetings, knowledge bases—are being asked to behave as tools inside one AI-led workflow rather than as destinations that each need their own growth story.The pattern is easiest to see in software people use at work. On October 29, at Tencent’s global digital ecosystem event in Shenzhen, senior executive Tang Daosheng introduced Buddy One, a suite built around the WorkBuddy agent desk and linked to Tencent Docs, meeting products, and knowledge tools so an assistant can work on material the user has authorized. The individual product names remain. Their relationship changes: they become resources a task can call, not only apps a user must open in sequence.Organization moved before the branding. In March, Tencent Cloud created a sixth product division to build AI-native coding tools (CodeBuddy) and broader productivity agents (WorkBuddy), under cloud CTO Wang Huixing and reporting to Tang’s CSIG group. Coding assistance had already spilled into general office trials for thousands of employees. A separate line, QClaw—an agent aimed at controlling a PC through conversation, born in the same season as a wave of open-source computer-use tools—was later shifted toward the same division. By late September the company confirmed QClaw would shut in December, with users pointed to WorkBuddy. Two experiments, one entrance.Alibaba’s path ran through similar congestion. After a period in which business groups were pushed toward sharper P&L boundaries and thinner central platforms, AI pulled several efforts back together. In June, newly installed DingTalk chief Chen Yusen combined the Wukong and MuleRun teams. Those projects had already wandered through online coding, agent marketplaces, and “build your own agent” tools before settling on general agents that take a goal and execute. QoderWork, aimed at desktop agents, was showing demos that read spreadsheets and produced charts and decks—the same job category MuleRun and DingTalk’s Wukong were targeting from different angles. By August, Alibaba had put those capabilities into a public test under Tongyi office (Qianwen Office), arguing that parallel invention had run its course.ByteDance compressed product and go-to-market in the same spirit: Feishu product teams were grouped with Doubao product teams, while sales and service were parked elsewhere—another way of saying the chat-and-agent surface and the workplace suite should not maintain fully separate product identities.None of this is only a Chinese instinct. Google’s combination of Brain and DeepMind, and the later placement of Gemini app work closer to DeepMind, followed the same pressure: model research, consumer surfaces, and old application businesses cannot stay sealed off when the customer experiences one assistant.Reorganizations also expose a conflict of clocks. Product teams need active users now; foundation-model teams need months of pre-training and evaluation. Tencent’s path with Yuanbao illustrated the trade. The consumer assistant was moved under the cloud-and-industry group so it could be run as a product, not only as a showcase for in-house models. When DeepSeek’s R1 appeared, Yuanbao connected quickly. Tang’s public line was blunt: most users do not care who trained the weights. Tencent’s chief AI scientist, Yao Shunyu, later described sending post-training talent to improve DeepSeek’s behavior inside Yuanbao before Tencent’s own pre-training was ready—an ordering some researchers disliked, justified as protecting a live user base that future models would need.On the Alibaba side, public reporting of a tense all-hands after a senior Qwen researcher’s departure showed the same collision in harsher form: questions about team redesign by training stage and modality, resource gaps for coding environments and infrastructure talent, and a CEO promising tighter feedback loops while forming a small group to coordinate group-wide support for base models. Whether the issue is “use a rival model to win the app store” or “who gets scarce CPUs for the next run,” AI forces choices that cut across old research prestige and new commercial scorecards.The strategic bet is consistent even when logos differ. Agents that only chat are easy to demo and hard to monetize. Agents that edit the file, update the sheet, pull the meeting transcript, and respect enterprise permissions need the company’s existing productivity stack to open up—and need internal systems rewritten so machines can call them. DingTalk’s reorganization language went as far as creating an IT unit tasked with making internal systems “easy for agents to use.” Tencent’s suite rules point toward shared quotas and cross-product retrieval once accounts are linked. Microsoft’s own progression—from Copilot that answered questions to agents that alter Word, Excel, and PowerPoint content—sets an external benchmark for multi-step edits that preserve user intent.For years, these companies expanded by launching another App for another job. AI is pushing them to treat jobs as workflows that cross Apps. That is why horse-race teams are being retired, why office agents absorb PC agents and coding agents, and why model groups are told to care about multi-turn users who speak in fragments. The technical problem is alignment of tools and weights. The management problem is alignment of P&Ls. Both are now on the same org chart. 更多精彩内容，关注钛媒体微信号（ID：taimeiti），或者下载钛媒体App