字号·· | 护眼
经济学人

人工智能代理将给企业带来冲击AI agents will cause disruption for companies

点「原文对照」整页切到原文,或双击某段只看那段的原文。

(《经济学人》)(The Economist)
(帕尔萨·穆斯塔吉姆)(Parsa Mostaghim)

自人工智能热潮开始以来,科技界的知名人士预测,能够既聊天又行动的智能体——即代理机器人——很快将在经济中广泛应用。OpenAI首席执行官山姆·奥尔特曼表示,智能体可能在2025年“显著改变公司的产出”。英伟达当时的首席执行官黄仁勋预测,代理人工智能将在当年“起飞”。还有人称2026年将是“智能体之年”。然而,AI智能体大多仍局限于软件开发者使用,开发者往往倾向于快速采用新技术。这种情况将在2027年改变。

Since the start of the artificial-intelligence boom, tech-industry luminaries have predicted that agents—bots that can act as well as chat—would soon spread across the economy. Sam Altman, the boss of OpenAI, said agents may “materially change the output of companies” in 2025. Jensen Huang, then boss of Nvidia, predicted that agentic AI would “take off” that year. Others said 2026 would be the “year of the agent”. Yet AI agents mostly remained the preserve of software developers, who tend to adopt new technologies quickly. That will change in 2027.

根据AI模型市场OpenRouter的数据,触发代理活动(例如使用工具)的AI提示词占比,从2026年4月的2%上升至9月的24%。智能体的更广泛应用将冲击企业成本、工作岗位、安全等诸多方面。

According to OpenRouter, a marketplace for AI models, the share of AI prompts that trigger agentic activity, such as the use of tools, rose from 2% to 24% from April to September 2026. Wider use of agents will disrupt corporate costs, job roles, security and more.

先说成本。与聊天机器人的来回对话不同,要求智能体执行特定任务会以惊人的速度消耗代币——AI输出的计量和计费单位——尤其是当智能体生成更多智能体来协助它时。2026年,鼓励员工尽可能多地使用AI(所谓“代币最大化”)的公司面临巨额账单。例如,优步在四个月内就耗尽了2026年的全部代币预算。

Start with costs. Unlike a back-and-forth with a chatbot, asking an agent to perform a particular task can burn through tokens—the units in which AI outputs are measured and billed—astonishingly fast, particularly if an agent spawns more agents to help it. In 2026 firms that encouraged employees to use AI as much as possible (so-called “tokenmaxxing”) were hit with huge bills. Uber, for example, burned through its entire token budget for 2026 in four months.

这迫使公司更加关注成本管理、成本效益分析,并为任务选择合适的模型。追踪企业支出的公司Ramp的数据显示,用户对2026年上半年推出的强大新AI模型采用较慢,倾向于继续使用更便宜、不那么先进的模型,除非是处理最复杂的任务。咨询公司麦肯锡称,五分之一的公司正在限制AI使用,以控制运营成本。

This has forced companies to put more focus on cost management, cost-benefit analysis and using the right model for the job. Data from Ramp, a firm that tracks corporate spending, suggest users were slow to adopt powerful new AI models introduced in the first half of 2026, preferring to stick with cheaper, less sophisticated ones except for the most complex tasks. According to McKinsey, a consultancy, a fifth of firms are limiting AI use to keep a lid on operating costs.

企业也在更多地使用开放权重模型,这些模型能以专有模型成本的一小部分取得令人印象深刻的结果。电信巨头AT&T在其25%的AI使用中采用开放模型,并设定了长期目标,即达到70%至80%。但企业需要更好地评估代理的输出,以检查它们是否使用了最合适的模型。

Companies are also making greater use of open-weight models, which can deliver impressive results at a fraction of the cost of proprietary models. AT&T, a telecoms giant, uses open models for 25% of its AI usage, and has set a long-term target of 70-80%. But firms need to get better at evaluating agents’ output, to check that they are using the most suitable model.

管理结构和工作角色也将受到影响。代理可以从内部系统收集信息,监控员工的产出,这可能意味着需要的经理更少。据就业咨询公司Korn Ferry称,五分之二的员工表示,自2025年以来,他们的公司部分原因是因为AI而精简了管理层级。

Management structures and job roles will also be affected. Agents can gather information from internal systems to keep an eye on workers’ output, which may mean fewer managers are needed. According to Korn Ferry, an employment consultancy, two-fifths of employees say their firms have trimmed management layers since 2025, in part because of AI.

与此同时,各类员工将花更多时间指挥代理。这可能比管理人类更容易:代理不会为认可或晋升而争斗,不会浪费时间开会,也乐于让其他代理使用自己的工作成果。但代理也会模糊工作角色之间的界限,例如让程序员完成设计师的工作,反之亦然。这可能会在人类之间引发紧张。

Meanwhile, employees of all kinds will spend more time bossing agents around. This may be easier than managing humans: agents do not fight for recognition or promotion, do not waste time having meetings and are happy for other agents to use their work. But agents will also blur the lines between job roles, for example by allowing coders to do the work of designers, and vice versa. That may cause tension between humans.

随着代理被授予更多系统的访问权限,并被要求执行更复杂的任务,它们将需要监督,以确保不会做出意料之外的事情。这凸显了代理在2027年扰乱企业的另一种方式:带来新的安全威胁。2026年发生了多起事件,代理群失控并闯入其他系统。这些并非蓄意攻击,而是AI测试的意外后果。但随着网络犯罪分子获得更强大的模型,代理群发动蓄意攻击的浪潮似乎已经开始,并在未来几个月将愈演愈烈。

As agents are granted access to more systems and asked to perform more complex tasks, they will need supervision to ensure that they do not do unexpected things. That highlights another way agents will disrupt companies in 2027: by posing new security threats. There were several incidents in 2026 in which agent swarms went rogue and broke into other systems. These were not deliberate attacks, but unintended consequences of AI testing. But as cyber-criminals get their hands on more capable models, a wave of deliberate attacks by agent swarms appears to have begun, which will intensify in the coming months.

最后,公司将不得不决定是否以及如何让外部代理更容易获取信息,以便促进信息检索、商务等活动。一些零售商已禁止代理进入其网站;另一些则欢迎它们。组织机构已经在为“代理洪流”苦苦挣扎:一场由人工智能生成的求职申请、投诉和法律挑战构成的洪流。以各种方式,这些新的人工智能工具确实将成为变革的推动者。■

Finally, companies will have to decide whether and how to make information more readily available to external agents, to facilitate information retrieval, commerce and so forth. Some retailers have banned agents from their websites; others welcome them. Organisations are already struggling with “agentic flooding”: a deluge of AI-generated job applications, complaints and legal challenges. In all kinds of ways, these new AI tools really will be agents of change. ■

阅读媒体原文 ↗