Tencent's WorkBuddy hit 20.97 million monthly PC visits in China in June 2026, taking the top spot in the AI office race and exceeding the combined traffic of second and third place. But that number isn't the point—what truly deserves our attention is that the underlying power ledger has changed: from 2026 onward, the metric that ranks AI companies shifts from "how many people clicked" to "how many Tokens were consumed."
What this is
A Token is the smallest billable unit of a large model, priced by word count. Letting AI run a complete employee workflow (such as auto-organizing invoices) burns Tokens at tens of times the rate of a human writing a weekly report.
SiliconFlow CEO Yang Pan, in a 100-minute conversation, threw out this judgment: "Starting in 2025, we should stop developing software for humans." His reasoning: future Agents (AI that autonomously completes multi-step tasks) will call external APIs far more often than humans click screens. In other words, the primary "user" of enterprise IT is no longer the employee, but the AI doppelganger behind the employee.
Once we see this main line, both Tencent's WorkBuddy winning and ByteDance folding the Feishu team into Doubao make sense—this ToB (business-to-business sales) battle is competing for "the default entry point when AI works for you."
Industry view
The supportive side sees this as a redistribution of business models. When AI can generate features on demand, money flows from middle-layer SaaS (cloud software subscription services) to model makers—"whoever generates features that meet the need gets paid."
But cautious voices exist. First, Token consumption doesn't equal revenue—many Agent projects burn Tokens happily during PoC (proof-of-concept, i.e., small-scale feasibility testing), only to discover high error rates and steep maintenance costs in production; Andrew Ng has previously warned that 90% of Agent projects get stuck at the deployment stage. Second, open-source models (led by Zhipu AI's GLM 4.7) are crossing the "baseline inflection point"—when "a college student who can actually do work"-tier models cost as little as cabbage, Token scarcity gets diluted fast. Third, enterprise Agent deployment is the dirty work of mapping business processes and on-site debugging; big players may not want to build their own delivery teams, and the win may fall to the service provider ecosystem.
Impact on regular people
For enterprise IT: budgets will gradually shift from "paying SaaS annual fees" to "paying per-call model fees"—procurement and finance have to learn a new language.
For individual careers: core competitiveness shifts from "knowing how to use a certain software" to "whether you can decompose repetitive tasks for AI to run." The efficiency gap between people who can orchestrate Agents and those who only chat with ChatGPT will be far wider than in the tool-party era.
For consumer markets: ZARA, H&M, and Google are all scrambling around "virtual try-on" for body data—whichever runs this data flywheel successfully locks out smaller brands from entry.