May 29, 2026

The Takeaway: AI's breakthroughs are arriving so fast that they paradoxically *slow down* enterprise rollout — every new model obsoletes the architecture you just spent a year deploying.

Aaron Levie, CEO of Box, sells to the Procter & Gambles and Morgan Stanleys of the world and has talked to a couple hundred Fortune 500 CIOs just this year. His whole job is bridging frontier tech to the messy real world, and his read on agentic AI in the enterprise is refreshingly grounded. The mood, he says, is more optimistic than cynical — CIOs see their engineers shipping faster with Claude Code and Codex, and the business is now pulling them to bring those gains everywhere else.

The hottest button is token cost. The old "$20 per user per month" model breaks when a single coding task can burn $1,000 of compute. Worse, frontier token prices are *rising*, because a decade of compute scaling got compressed into eighteen months and the labs have pricing power. Levie's sharpest prediction: AI spend is escaping the IT budget (a fixed 3–7% of revenue) and moving into line-of-business OpEx — but marketing and sales teams have no FinOps discipline for compute. He floats "a $5 billion startup waiting to happen, just ERP for your AI compute."

Why does coding race ahead of the rest of knowledge work? Structural reasons: technical users who can fix a stuck agent, verifiable output, the codebase holds all the context, and clean access controls. Knowledge work has none of that — context lives across 20 systems and entitlements are a mess. That's why diffusion takes a decade and why the "internal FDE" becomes a durable new job, not a casualty. On jobs, he's a "complete Jevons paradox pill person": cheaper expert work lights up far more projects, so demand for people rises.

His most memorable line on why rollout drags: "the technology is getting so advanced that it makes obsolete the prior thing that you implemented, which actually means that the rollout takes longer, because there's no stable environment to roll things out in." His advice to anyone wanting to future-proof: spend $50–100 a month, actually use the tools, and treat an agent like an unlimited chief of staff — then notice all the places you'd want to *hire* a human to run with what it produces.

核心要点： AI 的突破来得太快，反而*拖慢*了企业落地——每出一个新模型，就让你刚花一年部署好的架构变得过时。

Box CEO Aaron Levie 的客户是 Procter & Gamble、Morgan Stanley 这种全球最大的企业，光今年他就和几百位 Fortune 500 的 CIO 聊过。他的本职就是把前沿技术嫁接到混乱的真实世界，所以他对企业级 agentic AI 的判断格外接地气。他说现在的情绪是乐观多过怀疑——CIO 看到自己的工程师用 Claude Code 和 Codex 出活更快，业务部门反过来催着把这种收益推广到其他岗位。

最敏感的话题是 token 成本。当一个 coding 任务就能烧掉 1000 美元算力时，过去那套"每用户每月 20 美元"的模式就崩了。更糟的是，frontier token 的价格还在*上涨*，因为本该十年完成的算力扩张被压缩进了十八个月，而 labs 手握定价权。Levie 最犀利的判断是：AI 开支正在逃离 IT 预算（固定在营收的 3–7%），转移到各业务线的 OpEx——可市场和销售团队根本没有管算力的 FinOps 能力。他顺手抛出一个点子："这里有一家估值 50 亿美元的创业公司在等着诞生，就做你 AI 算力的 ERP。"

为什么 coding 会跑在其他知识工作前面？是结构性原因：技术型用户能在 agent 卡住时自己修，产出可验证，codebase 本身就装着全部 context，访问控制也很干净。而知识工作完全没有这些——context 散落在 20 个系统里，权限一团乱。这就是为什么扩散要花十年，也是为什么"internal FDE"会成为一个长期存在的新岗位，而不是被淘汰的对象。谈到工作，他自称是"彻底吃了 Jevons 悖论这颗药丸的人"：专家级工作变便宜，会点亮多得多的项目，对人的需求反而上升。

他关于落地为何拖沓最点睛的一句话："技术变得太先进，先进到让你之前部署的东西过时，这反而意味着推广更慢，因为根本没有一个稳定的环境让你去铺开。"对想给职业生涯上保险的人，他的建议是：每月花 50 到 100 美元，真正去用这些工具，把 agent 当成一个无限量的 chief of staff——然后你会注意到，有太多地方你其实想*雇一个人*来接手它产出的东西。