Box CEO Aaron Levie 谈 2026 年企业 AI 现状：Token 成本、Agent 部署与 Headless 趋势Box CEO Aaron Levie on 2026 Enterprise AI: Token Costs, Agent Deployment, and Headless Trends
核心要点：企业正处于 Agentic AI 采用的早期阶段，模型突破速度之快反而减缓了组织部署，因为新能力不断让先前实现变得过时。Box CEO Aaron Levie 凭借服务 Global 2000 企业的深刻洞察解释道，虽然编码 Agent 已实现逃逸速度，但将 Agentic 工作流扩展到一般知识工作面临数据架构、访问控制和实施复杂性等重大阻力。随着 Agent 执行更复杂、长时间运行的任务、更大上下文窗口和更强大模型，Token 成本大幅上升，将 AI 支出从 IT 预算转向业务部门 OpEx，并催生新的 FinOps 实践。Levie 指出了一个 bittersweet 的动态：“技术变得如此先进，以至于让先前实现的东西过时，这实际上意味着部署需要更长时间。”他强调了内部“FDE”角色的兴起——嵌入业务单元的技术专家，负责将 Agent 与正确的数据上下文和安全连接起来。Headless 软件将显著增长，用于 Agent 交互，但图形界面对于复杂的人类工作仍不可或缺。尽管面临挑战，Levie 对通过 Jevons 悖论创造就业保持乐观，因为生产力提升使各行业能够开展更雄心勃勃的项目。企业应关注数据准备、访问控制，并利用补贴期实验工具以建立熟练度。
The Takeaway: Enterprises are still in the very early stages of agentic AI adoption, where rapid model breakthroughs paradoxically slow down organizational rollout as new capabilities constantly obsolete prior implementations. Box CEO Aaron Levie, with deep insight from serving Global 2000 companies, explains that while coding agents have achieved escape velocity, extending agentic workflows to general knowledge work faces major friction from data architecture, access controls, and implementation complexity. Token costs have surged dramatically as agents perform more complex, longer-running tasks with larger context windows and more powerful models, shifting AI expenses from IT budgets into line-of-business OpEx and creating needs for new FinOps practices. Levie notes a bittersweet dynamic: "the technology is getting so advanced that it makes obsolete the prior thing that you implemented, which actually means that the rollout takes longer." He highlights the rise of internal "FDE" roles—technical specialists embedded in business units to wire up agents with proper data context and security. Headless software will grow significantly for agent interactions, but graphical interfaces remain essential for complex human work. Despite challenges, Levie remains optimistic on job creation through Jevons paradox, as productivity gains enable more ambitious projects across industries. Enterprises should focus on data readiness, access controls, and experimentation with tools while leveraging subsidized periods to build fluency.
查看原文 →