Box's Levie says diffusion, not intelligence, gates enterprise AI

Box CEO Aaron Levie says the rate limiter for enterprise AI is not model capability but the diffusion of technology into regulated workflows. On Sequoia's Training Data podcast, hosted by Sonya Huang, Levie argues the value now sits in "the bridge from a model's raw capability to the actual workflow" inside sectors like banking and pharmaceuticals, not in the raw model itself.

Box is building an agent platform tuned to its own file system, permissions structure and search. The episode description says that stack outperforms direct API access to Claude or ChatGPT on both speed and accuracy for tasks over Box content, a claim sourced to Levie rather than to a published benchmark.

Levie's forecast: "within five years, 90% of enterprise tokens go to work no human user ever initiated." He frames long-running agents doing contract analysis and similar document work as the substrate for that shift, in the knowledge-work domains where enterprise adoption has trailed developer adoption.