Box's Aaron Levie: On Reinventing Yourself in the AI Age and Enterprise Diffusion

Levie sees a large enterprise AI opportunity in the application layer, where software connects capable models to company-specific workflows through data access, integration, and domain expertise. Because legacy systems, human review, and change management can block deployment, he argues that a substantial implementation and diffusion ecosystem will be needed unless AI labs build workforces of roughly 100,000 to handle transformation themselves.

Levie argues that product-tuned agents can outperform generic model access by combining company-specific knowledge with tools for retrieving and ranking information: Box’s agent harness searches accounts with tens of millions of files and, he says, improves accuracy and latency over simply connecting an external model to Box’s API. He says accuracy can determine whether document workflows are automatable—for example, moving from 70% to 97% accuracy on 100-page bank loan documents—and that enterprise agents must also respect permissions and access controls.

Levie expects enterprise AI to move beyond chat toward agents doing background and asynchronous work, and predicts that in five years 90% of enterprise tokens will come from tasks users did not initiate directly. He illustrates the potential with bank onboarding: agents could handle a complex process that currently takes one or two weeks in about an hour, escalating steps that require human review or verification.

Levie says AI could make valuable enterprise information more usable because labor costs have left many contracts and research documents unread at scale; AI could extract information without requiring people to review every document manually. He also sees legal work as a promising automation area because it involves computer-based tasks such as reviewing and writing documents and processing large volumes of information.