Aaron Levie：应用层才是AI真正的战场Aaron Levie: Application Layer Is Where AI Value Accrues
The Takeaway：模型本身很强，但真正把智能接到企业真实工作流的那一层，才是价值最大的地方。
Box创始人兼CEO Aaron Levie认为，开放源模型和超级智能的叙事让人低估了“桥”的厚度。企业需要的不是通用聊天机器人，而是能处理权限、遗留系统、人机协作、变更管理的深度应用。历史已经证明：AWS/GCP创造了基础设施价值，但Snowflake和Databricks才真正把数据用起来。同样的逻辑会在智能时代重演。
他直言：“模型会极其有价值，但把模型带进银行、生命科学、医疗、政府这些真实工作流的应用，那将是大量软件。”Box自己坐在数千亿文件上，正用agent把非结构化内容变成可查询、可自动化的结构化数据。长期看，90%的企业token可能都不是用户主动触发的，而是后台agent默默跑完流程后只让人审核结果。
扩散速度会比硅谷预期慢得多。编码agent爆发是因为代码几乎全是文本、工程师能自己修bug、且高薪。销售、合规、法务则完全不同——外部约束多、数据散落在遗留系统、权限复杂。能耐心做domain expertise和落地的应用公司，才是下一波赢家。
Box创始人兼CEO Aaron Levie认为，开放源模型和超级智能的叙事让人低估了“桥”的厚度。企业需要的不是通用聊天机器人，而是能处理权限、遗留系统、人机协作、变更管理的深度应用。历史已经证明：AWS/GCP创造了基础设施价值，但Snowflake和Databricks才真正把数据用起来。同样的逻辑会在智能时代重演。
他直言：“模型会极其有价值，但把模型带进银行、生命科学、医疗、政府这些真实工作流的应用，那将是大量软件。”Box自己坐在数千亿文件上，正用agent把非结构化内容变成可查询、可自动化的结构化数据。长期看，90%的企业token可能都不是用户主动触发的，而是后台agent默默跑完流程后只让人审核结果。
扩散速度会比硅谷预期慢得多。编码agent爆发是因为代码几乎全是文本、工程师能自己修bug、且高薪。销售、合规、法务则完全不同——外部约束多、数据散落在遗留系统、权限复杂。能耐心做domain expertise和落地的应用公司，才是下一波赢家。
The Takeaway: Models are powerful, but the real value sits in the thick bridge that connects intelligence to actual enterprise workflows.
Box founder and CEO Aaron Levie argues that the “LLM wrapper” dismissals missed how vast the gap is between model capability and real-world processes. Enterprises need agents that handle permissions, legacy systems, human-in-the-loop delays, and change management—not just pure superintelligence. History rhymes: AWS and GCP created infrastructure trillions, yet Snowflake and Databricks still built massive businesses on top. The same pattern will hold for intelligence.
“The models will be insanely valuable, but the application of bringing those models into real workflows in banking and life sciences and health care and government, that’s just going to be a lot of software.” Box sits on hundreds of billions of files and is shipping agents that extract metadata, answer questions across unstructured data, and run long-running background processes. In five years Levie bets 90% of enterprise tokens will be kicked off by agents users never see, only the resulting tasks.
Diffusion outside coding will be slower. Coding agents exploded because code is text, engineers self-triage, and the payoff is high. Sales, legal, and compliance face external rate limits, scattered data, and access-control nightmares. The companies willing to do the unglamorous domain work and implementation will capture the applied-layer trillions.
查看原文 →
Box founder and CEO Aaron Levie argues that the “LLM wrapper” dismissals missed how vast the gap is between model capability and real-world processes. Enterprises need agents that handle permissions, legacy systems, human-in-the-loop delays, and change management—not just pure superintelligence. History rhymes: AWS and GCP created infrastructure trillions, yet Snowflake and Databricks still built massive businesses on top. The same pattern will hold for intelligence.
“The models will be insanely valuable, but the application of bringing those models into real workflows in banking and life sciences and health care and government, that’s just going to be a lot of software.” Box sits on hundreds of billions of files and is shipping agents that extract metadata, answer questions across unstructured data, and run long-running background processes. In five years Levie bets 90% of enterprise tokens will be kicked off by agents users never see, only the resulting tasks.
Diffusion outside coding will be slower. Coding agents exploded because code is text, engineers self-triage, and the payoff is high. Sales, legal, and compliance face external rate limits, scattered data, and access-control nightmares. The companies willing to do the unglamorous domain work and implementation will capture the applied-layer trillions.