Box CEO Levie: AI Diffusion Slower Than Expected

Sequoia Capital's podcast "Training Data" host Sonya Huang recently recorded an episode with Box founder and CEO Aaron Levie. Huang noted that during the recording, current hot topics like Sequoia partner Doug Leone's painless root canal surgery were trending, rather than advancements in AI, marking a "simpler era". However, she believes Levie's advice on how businesses and individuals can reshape themselves for the AI era remains timeless.

Box, founded by Aaron Levie 20 years ago, currently holds hundreds of billions of enterprise files. He has bet the company's future on AI agents that can read and process each of these files. According to Sonya Huang, Levie is one of the most deeply involved figures in the AI industry, appearing on the cap tables of nearly all companies in the field. He claims that 95% of his AI knowledge comes from Twitter and is regarded as one of the few CEOs who can "cross internet circles" while gaining the trust of enterprise CIOs.

In the discussion, Levie made two core points: first, the gap between current model capabilities and actual enterprise workflow needs remains vast, and bridging this gap requires significant software engineering investment; second, the diffusion of AI outside the programming field will be much slower than the current general expectations in Silicon Valley, and this "slowness" is precisely the true source of value for applied layer companies.

The conversation also covered a range of specific topics: why applied layer companies are the hottest "new labs" today, and why the "LLM-wrapper" business model has finally started to work; the potential conflict of interest if model providers directly manage enterprise token routing, akin to a "fox-guarding-the-henhouse" scenario; why enterprises accept AI for coding but are resistant to AI-generated documents/PPTs (the so-called "work slop" phenomenon); and how Box is building its own agentic technology framework, claiming it outperforms direct calls to underlying model APIs in terms of accuracy and latency.

Additionally, they discussed the "open weights paradox"—the simultaneous exponential growth of closed-source model labs and open-source model ecosystems; the challenges that continual learning technology must address before landing in enterprise scenarios, such as meeting the specific needs of a lawyer handling five cases simultaneously, with "Chinese wall" requirements for information isolation; Levie predicts that in five years, over 90% of tokens consumed by enterprises will come from tasks not initiated by humans; his key judgment for current entrepreneurs is that those who can truly deliver AI capabilities to customers will win this competition.

From a market mechanism perspective, the funding and value distribution logic pointed out by this discussion is: if applied layer companies (like Box) truly master the distribution channels and scenario data (hundreds of billions of enterprise files) to end customers, their bargaining power will be on par with or even stronger than that of underlying model vendors—this is also why the "LLM-wrapper" business model, previously viewed skeptically, is now seen as viable; and the reason why "model vendors directly routing enterprise tokens" is likened to "fox-guarding-the-henhouse" is that once model vendors control both the underlying capabilities and the end distribution channels, applied layer companies will lose bargaining chips, and customer data and usage behavior may be exploited by upstream model vendors. Who benefits: established software companies with existing enterprise data and customer relationships; who is under pressure: lightweight application startups that rely solely on calling underlying model APIs and lack their own scenario data and distribution channels.

Source: Public Information