在AI代理（agentic AI）部署方面，企业仍处于起步与试探阶段。KPMG《Global AI Pulse》显示，只有约9%的企业在组织内已有主动运行的代理，约20%正在探索，17%处于试点阶段；企业普遍属于“快追随者”而非“先行者”。以色列公司Upstream的市场Slack实例中，首个代理加入后出现了长达24小时无人发言的谨慎期，随后逐步被接受，这反映出有AI可执行（agency）后人们先警惕、再试用的行为规律。代理可规划并执行复杂工作流，对流程重构与员工协作方式带来比传统问答式LLM更明显的影响。
以规模化案例看，位于以色列的Upstream（约150人）将代理用于跨系统数据提取与例行任务，把以往耗时且易遗忘的问题快速串联定位，并在汽车安全场景中协助在数分钟内进行分诊、关联多工具信息与故障归因。Adecco Group（约35,000名员工）约一年前在英国开始试点，在一年内确认预筛选环节可由代理承担，并在全年推进至目标：到今年底其收入中50%由agentic AI驱动。实际效果显示，候选人中有50%与代理的对话发生在非工作时段（高峰在23:00至03:00），招聘流程节省时间约20%，公司因此把人工更多转向候选人辅导而非裁员。
成本与组织重构是第二个关键约束。代理常按“token”计费，类似Deloitte所谓“AI经济新货币”，按量合同下忘记关闭代理会导致账单飙升；Adecco采用Salesforce无限制访问协议以控费。Prosus（约23,000人，已建约50,000个代理）通过Toqan让非技术员工自建任务型代理，18个月内形成文化转型。虽然每日活跃约5,000个代理，但据称效率提升对应超出1000个全职岗位，且员工在法务与外卖业务中可更快生成数据分析与报告。与此同时，生产率放大也放大了校验负担：Euro Beinat举例说“我创造了20倍工作量，就需要20人来复核”，企业需同步重设计团队与流程；如KPMG所言，许多管理者仍像“不会骑自行车的人给孩子讲解”——缺少一线体验会拉开实践差距。
Adoption of agentic AI remains early: firms are moving from experimentation to operational use, with limited deployment. KPMG’s Global AI Pulse reports only about 9% of organizations have active agents, about one-fifth are exploring, and 17% are running pilots. Leaders often prefer to be fast followers rather than pioneers. Human resistance is common early, as seen at Upstream where a marketing Slack channel had 24 hours of silence after an agent was added, then gradually gained adoption. The key difference is that agentic AI can plan and execute workflows, affecting team operations more deeply than traditional command-style LLM use.
Case evidence links deployment to measurable gains. Upstream, a 150-person company, improved team efficiency by applying agents to routine, multi-source tasks, and in mobility cybersecurity cases they now triage incidents, correlate tools, and speed root-cause investigation to minutes. Adecco, with about 35,000 staff, piloted agents in UK recruitment and is now scaling toward a target of 50% of revenue powered by agentic AI by year end. It found 50% of candidate conversations with the agent occurred outside normal office hours, peaking 11 p.m.–3 a.m., indicating prior candidate coverage gaps. So far, it has saved about 20% of recruitment time and has shifted staff effort toward coaching and support rather than cutting headcount.
The dominant obstacle is cost control. Most platforms bill by tokens—called by Deloitte the AI economy’s new currency—so pay-as-you-go users can see high bills if autonomous agents run unattended; Adecco mitigates this with an unlimited Salesforce access contract. Prosus, with around 23,000 employees and about 50,000 agents built, has encouraged non-technical staff to build their own agents through Toqan for broad adoption. Daily active usage is about 5,000 agents, but leadership frames this as a cultural shift over 18 months, not merely a technical upgrade. Euro Beinat estimates productivity gains equivalent to more than 1,000 full-time employees, yet says magnified output also requires magnified review: creating 20 times more work requires 20 people to check it. Broad implementation therefore demands redesigned team structures, and Stephen Chase notes many executives still discuss AI without having personally operated it, like instructing a child to ride a bike without riding one.