国家数据局的统计显示，中国日均Token调用量从2024年初的约1000亿，增长到今年7月的超过500万亿，两年增长超过5000倍。Token是AI处理信息的最小单位，每一次调用背后，都是企业真实付出的算力、电力与费用。
个人层面的繁荣不需要论证，企业层面的图景却始终模糊。数千倍增长，意味着不再是尝鲜式的试用，而是一笔规模可观、且还在持续膨胀的投入。但如何评估与衡量，还没有公认的答案。
10月29日，2026腾讯全球数字生态大会在深圳开幕。腾讯集团高级执行副总裁、云与智慧产业事业群CEO汤道生在主题演讲中给出的解释是，企业落地Agent最大的问题，不是模型不够强，而是断点太多——AI要做分析、写方案，接不上企业内部的系统拿数据，就无从了解业务、渠道和客户的实际情况。
腾讯集团副总裁、腾讯云CTO王慧星则更进一步指出，上下文连接只是起点，企业要真正发挥Agent的价值，还需要解决业务接入、安全治理、效果衡量和规模化落地等问题，这也是WorkBuddy正在做的事情。
作为国内用户规模最大的AI办公智能体的运营方，腾讯的落地答案可以从WorkBuddy身上窥见一斑。过去八个月，这款产品从个人助手走向企业系统，这次大会公布的新进展，几乎都在这个方向上。
门槛已经从模型移到上下文，甚至是组织管理层面
个人层面的提效已经被数据确认，易观的数据显示，2026年9月，它以8534万次PC端月访问量位居中国桌面端AI原生办公智能体平台第一。
但落到企业层面，这种变化才刚刚开始。对“跑通AI生产力”这件事，汤道生在29日下午的交流中给了一个可以检验的说法。他先反问，今天还能不能不用AI工作？“我觉得挺难回去的，所以它有没有跑通，第一是它所创造的价值，你愿不愿意放弃。如果你放弃不了，那等于它已经成为我们工作的一部分，生活的一部分。”
在他看来，如果一家企业没有抓住根本，没有调整组织、让员工有能力用好AI，“那可能你跟别人付同样的成本，但是你没拿到别人的return的时候，自然你的体感就没跑通”。
同样的Token、同样的工具，产出取决于企业愿不愿意把流程、权限和判断标准一并交出来。因此，WorkBuddy企业版上线的可观测中心，解决的是效果怎么衡量的问题，它把Token花在哪个部门、哪个场景、哪一步慢了或贵了列出来，企业关心的已经从能不能用，转到管不管得住、投得值不值。
组织这一侧究竟卡在哪几处，大会前后的交流中，腾讯三位业务负责人回答的，都是智能体进入企业之前要过的条件。王慧星说，模型和Harness在产品和技术上已经迈过那道坎，接下来的工作是准备上下文，让智能体拿到企业真实的工作场景和负载。他把平台能力开放排在四大能力升级的第一位，目的就是让Agent接进企业上下文。
他将个人与企业的差别打了个比方，个人提效像是给每个人配一个分身或助理，处理邮件、日程、待办这类点对点的事；回到企业，需要的就不只是人手一个助理，而是让程序员、法务、财务这些专业角色变成可调用的数字员工，它们要接触客户数据、进入业务系统，权限没管住、结果没校验，影响的是一整条业务线。
会上公布的规模数字是，WorkBuddy企业版已覆盖中国企业500强中的一半，产品上线至今累计完成任务请求超过90亿次。
要把企业的系统和数据真正接进来，先得划清权限的边界。腾讯云副总裁、安全负责人董志强表示，Agent的行为跑偏真实存在，与传统软件不同，模型的行为是一个开放命题；对策是把行为关进沙箱、全程留痕可审计，外部材料一律当作数据处理，调用内部工具走网关白名单。
汤道生的演讲给出的正是这条上下文主线的完整版本。会上发布的Buddy One是一套面向个人用户的跨应用协同体系，以WorkBuddy为工作台，打通腾讯文档、腾讯会议、ima、腾讯网盘等产品，用户不必在应用之间搬运文件、复制上下文，一次提问即可检索已授权的文档、会议与知识库内容。
官方对这个变化的概括是，协作围绕任务协同，而不是围绕文件流转。文件流转只要求格式互通，任务协同要求上下文互通。
上下文分布在几处不同的地方，文档、会议和知识库沉淀的是企业做过什么，日常沟通里说的是正在做什么，ERP、CRM、OA与财务系统存的则是这家企业怎么运转。前两类靠产品之间打通，最后一类只能靠系统开放接口，WorkBuddy的做法是让管理系统开放MCP接口，再用skills描述清楚数据结构和关系，Agent才能自己取数、拆解任务。
同步发布的腾讯网盘承担统一存储，被定位为Agent时代的记忆空间，企业版存储空间从每席5GB扩容至50GB。加上与企业微信的深度融合，员工在企微中的沟通、文档和会议记录，在授权范围内成为Agent理解任务的依据，企业不用换系统、搬数据、重做流程。
这条路能不能被更多企业接受，还取决于它锁不锁定。汤道生在交流中说，Skills本身是可以带走的，它只是文本和对数据接口的描述；作为产品提供者当然希望产品有黏性，但真正要把一个产品服务到大面积用户，开放生态是一种方式，能接入很多腾讯不具备的能力和数据。
王慧星补充，WorkBuddy在与混元深度共建的同时，接入KIMI等主流开源模型，也支持企业自部署模型。数据、上下文和记忆的沉淀权留在企业一侧，企业才愿意把真实的经营上下文交出来。
用得越深，沉淀在平台里的企业上下文越厚，智能体的交付也越贴近企业自己的标准；前提是企业愿意把更多日常工作交给它。
组织变革没有捷径
王慧星在演讲中把企业落地Agent要跨的门槛概括为四道关，跑得通、可托付、可衡量、好落地，接不进才谈得上管得住，管得住才敢算投入，算得清才愿意复制。前三项分别对应平台能力开放、企业级安全方案与可观测中心，最后一关是复制，一个团队验证过的智能体，能不能推到更多团队和场景。可观测中心上线后，苏宁易购用它调整提示词与模型配置，Token消耗降低了10%。
最难的是跑得通这一步，一家头部制造企业在售后、产品、市场提出13个场景，WorkBuddy通过MCP直连其经销商管理系统，每月自动初审2万单索赔发票。真正的难题在人这一侧，此前已有上千名员工自发使用个人版，管理团队在禁用与放任之间两难，企业版的统一管控上线第二天，2000多个测试账号被员工领完，企业内的AI使用者扩大了10倍以上。接进系统是技术问题，把上千个自发使用的人纳入统一管控是组织问题。
WorkBuddy的客户案例中有三个可对照的样本，分别来自日化、通信和乳业。
中顺洁柔于今年10月完成1200人的WorkBuddy强制考核培训，洁柔CDO杨森林举例了供应链，他透露，从销售订单、发货、选仓到生产排产的整条链路交给智能体求最优解，成本节降达到四位数级。业务人员自己动手之后，原本分散在外包团队和IT部门手里的规则，开始沉淀在企业内部。
联通华盛是中国联通的全资子公司，开票这个场景，过去31个省级公司的财务人员每月各花两天，核对打款、订单与开票三张表，现在把三张表和匹配规则交给智能体，一名财务人员每月用两个小时即可完成。副总经理陈丰伟介绍，按约100个内部案例总结，重复性工作的效率提升约为10倍。
汤道生在演讲中举的第三个例子是伊利，场景在营销与销售这类前端业务。伊利组合使用腾讯云ADP、WorkBuddy与企业微信，由WorkBuddy汇聚业务数据、生成经营策略，再下发到上千名导购，由ADP的导购智能体生成推荐与话术。按演讲中的数据，运营提效超过50%，A/B测试中销售订单提升26%。
三个样本所处行业不同，共同点在于都没有推倒重来，都是在既有系统和工作方式上逐步把Agent接进来。
组织这一侧的难度，汤道生举了一个程序员的例子，腾讯内部最先用起来的是研发团队，代码几倍几倍地提效之后，整合与质量保证的压力压到测试团队，藏在深处的bug和安全漏洞可能要到很后面才出现。
“组织必须要变化。每个人在做的事，也在变化。”他说，这个过程会筛出谁用得好、谁用得不好，企业再据此提出要求。至于怎么改，他没有给出通行的方案，只说不同岗位都会经历类似的过程，先是充分拥抱工具带来的提效，随后发现新问题，再由企业提出要求。
规模领先之后，生产力领先是真正拐点
从规模看，WorkBuddy在办公智能体赛道已经领先，易观6月发布的报告显示，其月访问量已超过第二、三名的总和。访问量反映的是采用广度，与产出之间没有直接对应。
规模领先不意味着竞争减弱，钉钉依托阿里，将AI定位为组织流程的入口；飞书依托字节，侧重知识协作；企业微信的优势在于与微信的连接。三家的切入点分别是流程、协作与连接，最终都要证明Agent能不能进入企业的核心流程。
面对横向竞争，刘毅的回答是，“我们没想过守这个字”。
在他看来，市场还在早期，九成以上的企业还没有大规模使用Agent，增量远大于存量，此刻的份额说明不了什么。腾讯也不缺底座，企业微信和腾讯云各握着数百万客户，WorkBuddy可以进入这些客户的场景，连接他们的上下文。
剩下的差别在执行，他比较的是服务、模型协同与驻场落地三件事。他也判断B端不会出现一家独大，最后是平台、产品与模型组合在一起，分别服务不同的客群。眼下的领先因此仅限于采用层面。
流程深处的产出如何计量，目前没有统一答案。汤道生观察，过去几年行业关注模型的参数与评测分数，今年这类讨论明显减少，衡量标准转向“用了AI在某个场景到底得到了什么”。他也表示，很难用一个数字衡量整个行业，不同场景之间的成本差异很大，Token消耗量能反映用户愿意为任务付出多少成本，但只是一个粗略的口径。
在企业层面，性价比表现为流程能否整体运转。这场竞争最终由什么决定，是模型、Harness还是生态，拐点依靠什么条件成立？王慧星给出的答案是“开放”，这是腾讯的基因，也是优势。
王慧星认为，企业不应该为了使用Agent而推倒重建现有系统，也不应该被绑定在单一工具或模型上。WorkBuddy向内融合企业微信及腾讯办公产品，向外通过MCP、连接器等方式接入第三方协同工具和企业自有业务系统，并为企业提供多样化的模型选择。让Agent进入真实的业务环境，调用完成任务所需的数据与工具，才能真正从演示走向生产。（本文首发于钛媒体APP）
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
According to statistics from the National Data Agency, the daily average of Token calls in China increased from about 100 billion at the beginning of 2024 to over 5 trillion in July this year, and more than 5,000 times in two years. Token is the smallest unit in AI to process information, and each call is followed by real energy, electricity and costs.
Prosperity at the individual level does not need to be argued, but the picture at the enterprise level remains blurred. Thousands of times the growth means that it is no longer a taste of fresh trials, but rather a sizeable and growing input. But there is no accepted answer on how to assess and measure.
On 29 October, the 2026Tencent Global Digital Ecology Congress opened in Shenzhen. In his keynote address, CEO of the Tencent Group, CEO of the Clouds and Wisdom Industry, CEO Totson, explained that the biggest problem with Agent's downside was not that the model was not strong enough, but that there were too many break points — AI was going to analyse, write a plan, not get data from the enterprise's internal system, so it was impossible to understand business, channels and customers.
The Vice-President of the Tencent Group, the Tencent Cloud CTO King comet, went further, saying that context connectivity is only the starting point, and that business needs to address business access, security governance, impact measurement and scale-down, which is also being done by WorldBuddy, if it is to really be playing Agent's value.
As an operator of the largest AI office in the country, AI agent, Tencent's landing answer can be seen from WorkBuddy. Over the past eight months, this product has moved from personal help to the enterprise system, and almost all the new developments announced at the Conference have been in that direction.
The threshold has moved from model to context, even at the organizational level.
The effectiveness of the individual level has been confirmed by the data, and easy-to-see data show that in September 2026, it occupied the first place on the Chinese desktop end-of-the-top AI primary office AI agent platform with 8.5340 million PC end-month visits.
But at the enterprise level, this change is only just beginning. On the matter of “run-through-AI productivity”, Thomason gave a testimonial in his 29th afternoon exchange. He asked if he'd work without AI today. “I think it's hard to go back, so it doesn't run, first of all, the value it creates, and you're willing to give up. If you can't give up, it's already part of our work, part of our life."
In his view, if an enterprise fails to capture the fundamentals, to adjust the organization and to equip its employees to use AI, “it may be that you pay the same costs as others, but naturally you lose your senses when you don't get someone else's return”.
The same Token, the same tool, depends on whether the enterprise is willing to hand over the process, permission and the judgement criteria. Thus, the Observable Centre on the WorldBuddy Enterprise website addresses the question of how to measure the effects, which lists the sectors, scenes, slow steps or costs that Token spends on, and which business concerns have shifted from being useful to being unaffordable and undervalued.
The three Tencent business managers responded to the question on which side of the organization had been marked and on which side the exchange before and after the Congress had taken place, all the conditions that AI agent had required before entering the enterprise. According to Wang comet, the model and Harness have already passed that road in terms of products and technology, and the next step is to prepare context for the real work scene and load of the enterprise. He ranked the opening of the platform at the top of the four capacity upgrades, with the aim of getting Agent into the Enterprise context.
He gave an example of the difference between an individual and an enterprise, where the personal effect was to provide each individual with a split or assistant to handle e-mails, calendars, to-dos, etc.; to return to the enterprise, what was needed was not just an assistant, but rather to turn the professional roles of programmer, law enforcement, finance into a digital employee who could be called, who had access to client data and the business system, and permission had no control over the results and had no validation of the whole business line.
The scale figures published at the meeting were that the WorldBuddy enterprise version already covered half of the top 500 Chinese enterprises, and that the product had been online for more than 9 billion cumulative completion requests to date.
To truly integrate enterprise systems and data, the border of permission has to be drawn first. Tencent Vice President and Head of Security, Dong Zhiqiang, stated that Agent's behaviour was real and that, unlike traditional software, the model's behaviour was an open proposition; the response was to lock the behaviour into sandboxes, leave a mark on the entire process, and external material was treated as data processing and to use internal tools to walk through the gateway list.
Thomason's speech gave a complete version of context's main line. At the publish meeting, Buddy One was a cross-application synergetic system for individual users, using the WorkBuddy as a workstation, to access products such as Tencent documents, Tencent meetings, ma, Tencent webpads, without the need for users to move files between applications, copy context, and search for authorized documents, meetings and knowledge base content once asked.
The official summary of this change is that collaboration revolves around task synergy rather than document flow. The flow of documents requires only the interoperability of formats, and the synergy of tasks requires context.
context is distributed in several places, documents, meetings and knowledge banks are the business doing, daily communications are about what is being done, and ERP, CRM, OA and the financial system are about how it works. While the first two categories depend on product-to-product connectivity, the last one depends only on an open system interface, WorkBuddy ' s approach is to make management systems open to MCP interfaces and to use skills to describe clearly the data structure and relationships so that Agent can take out and dismantle tasks on his own.
The Tencent grid, which is synchronized with publish, assumes a unified storage, is located in the Agent-era memory space, and the enterprise version storage space is expanded from 5 to 50 GB each. In addition to the deep integration with the enterprise WeChat, communication, documentation and minutes of meetings of staff in micro-enterprises are, within the mandate, the basis of Agent ' s understanding of the mission, without changing systems, moving data and re-engineering processes.
Whether or not this path is acceptable to more businesses depends on its being locked. In his exchange, Tong Doo-sheng said that Skills itself could be taken away; it was only a text and a description of the data interface; as a product provider, of course, wanted the product to be sticky, but to truly serve a large number of users, open ecology was a way of accessing many of the capabilities and data that Tencent did not possess.
Wang comet adds that, while working with hybrid depths, WorkBuddy has access to mainstream open-source models such as KIMI and supports business self-deployment models. Data, context and memory sedimentation rights remain on the side of the enterprise, which is willing to hand over the real operation of context.
The deeper it is used, the thicker the company context, which is deposited in the platform, the closer AI agent is delivered to its own standards; provided that the enterprise is willing to give it more day-to-day work.
There are no shortcuts to organizational change.
In his speech, Wang comet summarized the threshold for the landing of Agent as a four-gauge, one that can run, payable, measurable and well, one that can't get in, one can't get in, one can't get in, one can count on the input, and one can't copy it. The first three, respectively, are open to platform capabilities, enterprise-level security programmes and observable centres, and the last level is a copy, and a team certified AI agent can move to more teams and scenes. When the Observatory Centre went online, Suning purchased it to adjust prompt and model configurations and token consumption was reduced by 10 per cent.
The hardest part was to get through this, with 13 scenes coming from a head manufacturing business after sale, products and markets, and WorkBuddy, through MCP, linking its distributor management system, automatically examining 20,000 claims invoices per month. The real challenge lies on the side of the human being, with thousands of employees using their own personal versions, the management team having the dilemma between disablement and release, the second day of the unified control of the enterprise, more than 2,000 test accounts were taken by the employees and the number of AI users in the enterprise increased by more than 10 times. Access to the system is a technical issue, and the integration of thousands of spontaneous users into unified controls is an organizational issue.
In the case of WorkBuddy, there were three contrastable sample cases, from the solar, communications and dairy industries.
In October of this year, China completed the WorldBuddy mandatory examination training for 1,200 people. The Jade CDO Yang Forest provides an example of the supply chain. He revealed that the entire chain, from sales orders, deliveries, silos to production discharges, was handed over to AI agent for the best possible solution, at a cost of up to four digits. The rules, which had been dispersed between the outsourced team and the IT department, had begun to sink within the enterprise after the operational staff had done so themselves.
In the case of the Chinese-connected wholly-owned subsidiary, ticketing, the finance staff of the 31 provincial companies of the past two days each month to check three forms of payment, order and billing, now hand over three forms and matching rules to AI agent, which can be completed by a finance officer in two hours per month. The Deputy Managing Director, Chen Fung Wei, presented a summary of some 100 internal cases, in which the efficiency of repetitive work increased by about 10 times.
The third example given by Tong Doo-sheng in his presentation was Ely, where the scene was at the front end of marketing and sales. The Ely portfolio uses Tencent CloudADP, WorkBuddy and the Enterprise WeChat, which brings together business data, generate business strategies and distributes them down to thousands of guides, and the ADP ' s guide buys AI agent to generate recommendations and words. According to the presentation, the operational effectiveness exceeded 50 per cent and the sales order was raised by 26 per cent in the A/B test.
The three samples are different in their trades, and the common denominator is that they have not been pushed back, and that they have gradually brought Agent in in the systems and working methods in place.
The difficulty of this aspect of the organization is illustrated by the example of a programmer, who was first used by the R & D team within Tencent, and whose code is several times more effective, the pressure of integration and quality assurance is placed on the testing team, hidden in the bottom of the bug and the security gap may have to come to a very late end.
“Organization must change. What everyone is doing is changing." He said that the process would sift out who was good and who was not, and the business would then demand it. As for how to change, he had not given a universal formula, but simply said that different jobs would go through a similar process, first of all with the benefit of a full hug, then with new problems and then with the company.
After the scale, the productivity lead is the real point.
On a scale, WorkBuddy has taken the lead on the AI agent track, with an easy-to-see June report from publish showing a monthly visit exceeding the combined of two and three. The number of visits reflects the breadth of use and does not correspond directly to the output.
The size lead does not mean less competition, pinned to Tori, positioning AI as an entry point to the organizational process; flying to byte, focusing on knowledge collaboration; and the advantage of WeChat, an enterprise, is connected to WeChat. The entry points for the three were processes, collaboration and connectivity, which ultimately required proof of Agent ' s access to the core of the enterprise.
In the face of horizontal competition, Liu Yi replied, “We did not intend to keep the word”.
In his view, the market was still in its early stages, and the fact that over 90 per cent of businesses did not use Agent on a large scale, with a much larger increase than stock, did not mean much at this point in time. Tencent also lacks a bottom seat, and the businesses WeChat and Tencent cloud hold millions of customers, and WorkBuddy has access to their scenes and links them to context.
The remaining differences were being implemented, and he compared services, modeling and fielding. He also judged that there would not be a single one at the end of B, and that it would end with a platform, a product and a model, each serving a different audience. The current lead is therefore limited to the adoption level.
There is currently no uniform answer on how to measure the output at the bottom of the process. Yudosheng observed that in the past few years there has been a marked decrease in the discussion of the parameters and scores of the industry-focused model, and that the measure has shifted to “what exactly has AI gained in a given scenario”. He also indicated that it was difficult to measure the entire industry with a single number, that the costs varied widely between scenarios, and that Token consumption reflected how much the user was willing to pay for the mission, but only a rough calibre.
At the enterprise level, value for money is demonstrated by whether the process works as a whole. What are the conditions under which this competition is ultimately determined, whether it is a model, a Harness or an ecology? The answer given by Wang comet is "open," it's Tencent's genes and it's an advantage.
In the view of Wang comet, businesses should not push down existing systems to use Agent and should not be tied to a single tool or model. WorkBuddy provides internally integrated WeChat and Tencent office products, external access to third-party synergistic tools and enterprise-owned business systems through MCP, connectors, etc., and provides a diversified model selection for enterprises. In order to truly move from demonstration to production, Agent can enter the real business environment and access the data and tools needed to fulfil its mandate. (This article was first published on titanium medium APP)
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.