Box CEO Aaron Levie：应用层才是AI企业落地的关键Box CEO Aaron Levie: The Applied Layer Is the Real Gap for Enterprise AI
The Takeaway：模型再强，也填不满企业工作流与真实数据之间的鸿沟，应用层公司将捕获巨大价值。
Box创始人兼CEO Aaron Levie指出，两年前“应用公司是最热门的neo labs”还显得奇怪，但今天市场已经证明：开源模型让“模型包装器”真正可用，因为企业需要的是从模型能力到具体工作流的桥梁。这座桥梁极其宽广——连接遗留系统、人工介入节点、变更管理、权限与合规——而不是单纯的超级智能。
Levie强调：“模型可能是世界上最聪明的超级智能，但工作流仍然需要你去对接其他数据系统，需要人机协作时刻，需要处理业务延迟，需要改造遗留流程。”历史证明基础设施（AWS/GCP）创造了万亿价值，但Snowflake、Databricks这类应用层产品同样创造了万亿。AI不会改变这个规律。聊天机器人可以通用，但真正的agentic工作流需要领域专长、数据接入与组织变革，这正是应用层的机会。即使模型持续跃升，这个层反而更重要，因为更强能力会解锁更复杂任务，放大执行差距。
他还谈到“狐狸看守鸡窝”问题：企业希望agent在精度恒定下做成本最优，因此天然偏好不绑定单一模型的应用层。补贴不会永久存在，公开上市后资本规律会回归。开放权重模型长期会切走成熟工作负载，但闭源与开源会形成编排+长尾任务的混合生态。最终，企业里90%的token可能来自后台自动运行的agent，用户只看到结果与待办。
Box创始人兼CEO Aaron Levie指出，两年前“应用公司是最热门的neo labs”还显得奇怪，但今天市场已经证明：开源模型让“模型包装器”真正可用，因为企业需要的是从模型能力到具体工作流的桥梁。这座桥梁极其宽广——连接遗留系统、人工介入节点、变更管理、权限与合规——而不是单纯的超级智能。
Levie强调：“模型可能是世界上最聪明的超级智能，但工作流仍然需要你去对接其他数据系统，需要人机协作时刻，需要处理业务延迟，需要改造遗留流程。”历史证明基础设施（AWS/GCP）创造了万亿价值，但Snowflake、Databricks这类应用层产品同样创造了万亿。AI不会改变这个规律。聊天机器人可以通用，但真正的agentic工作流需要领域专长、数据接入与组织变革，这正是应用层的机会。即使模型持续跃升，这个层反而更重要，因为更强能力会解锁更复杂任务，放大执行差距。
他还谈到“狐狸看守鸡窝”问题：企业希望agent在精度恒定下做成本最优，因此天然偏好不绑定单一模型的应用层。补贴不会永久存在，公开上市后资本规律会回归。开放权重模型长期会切走成熟工作负载，但闭源与开源会形成编排+长尾任务的混合生态。最终，企业里90%的token可能来自后台自动运行的agent，用户只看到结果与待办。
The Takeaway: Even the strongest models leave a massive gap between intelligence and enterprise workflows—this is the applied AI layer’s trillion-dollar opportunity.
Box founder and CEO Aaron Levie argues that two years ago calling application companies the hottest neo-labs would have sounded odd. Today the market has spoken: open-source models made “model wrappers” viable because enterprises need a bridge from model capability to actual workflows. That bridge is vast—connecting legacy systems, human-in-the-loop moments, change management, permissions and compliance—not pure superintelligence.
Levie puts it bluntly: “The model could be the most intelligent superintelligence in the world, but that workflow still requires you to connect up to other data systems, still requires these moments where there’s a human in a loop, there’s delays… there’s change management of the actual business process.” History shows infrastructure (AWS/GCP) created trillions, yet Snowflake and Databricks also created trillions. AI does not rewrite this pattern. Chatbots can be universal, but real agentic workflows demand domain expertise, data access and organizational transformation—precisely the applied layer’s role. As models improve, this layer becomes more, not less, important because greater capability unlocks harder tasks and amplifies execution gaps.
He also flags the “fox guarding the henhouse” dynamic: enterprises want agents that cost-optimize while holding accuracy constant, so they prefer application layers not tied to one model provider. Subsidies will not last forever; public markets enforce normal margins. Open-weight models will eventually peel off matured workloads, yet closed and open models will coexist in hybrid orchestration + long-tail setups. In five years, Levie bets 90% of enterprise tokens will come from background agents users never explicitly launch—they simply see results and review queues.
查看原文 →查看原文 →
Box founder and CEO Aaron Levie argues that two years ago calling application companies the hottest neo-labs would have sounded odd. Today the market has spoken: open-source models made “model wrappers” viable because enterprises need a bridge from model capability to actual workflows. That bridge is vast—connecting legacy systems, human-in-the-loop moments, change management, permissions and compliance—not pure superintelligence.
Levie puts it bluntly: “The model could be the most intelligent superintelligence in the world, but that workflow still requires you to connect up to other data systems, still requires these moments where there’s a human in a loop, there’s delays… there’s change management of the actual business process.” History shows infrastructure (AWS/GCP) created trillions, yet Snowflake and Databricks also created trillions. AI does not rewrite this pattern. Chatbots can be universal, but real agentic workflows demand domain expertise, data access and organizational transformation—precisely the applied layer’s role. As models improve, this layer becomes more, not less, important because greater capability unlocks harder tasks and amplifies execution gaps.
He also flags the “fox guarding the henhouse” dynamic: enterprises want agents that cost-optimize while holding accuracy constant, so they prefer application layers not tied to one model provider. Subsidies will not last forever; public markets enforce normal margins. Open-weight models will eventually peel off matured workloads, yet closed and open models will coexist in hybrid orchestration + long-tail setups. In five years, Levie bets 90% of enterprise tokens will come from background agents users never explicitly launch—they simply see results and review queues.