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

Summary

Aaron Levie says the enterprise AI market is already proving out the application-layer thesis: value comes from bridging model capability into real workflows, not from the model alone. He describes Box’s strategy as building a highly tuned, model-agnostic agent harness around its file system, search, permissions, and document-processing tools, with internal evals showing better accuracy and latency than raw API use in Box-specific workflows. Levie argues that enterprise adoption is constrained by access controls, legacy systems, human review loops, and workflow redesign, which makes diffusion slower than many in Silicon Valley expect. He sees a split future where frontier models handle orchestration and harder tasks while cheaper/open-weight models peel off mature workloads, and he predicts that within five years most enterprise token usage will come from background agents that users only review after the fact.

Notes

• Application-layer AI is working in enterprise because the value is the bridge from model capability to real workflow.
• Box is already shipping internal AI experiences, but the current emphasis is model-agnostic agent quality.
• Enterprise AI is a stack: applied companies, infrastructure providers, and multiple layers above the model.
• Real enterprise workflows require many non-model components beyond raw intelligence.
• Box’s core AI asset is its unstructured enterprise file corpus.
• Agents can answer questions over documents, extract metadata, structure data, and automate workflow steps.
• Box has been testing AI since the 2015-2018 AI winter, when models were too expensive and too specialized.
• Box uses OpenAI and Anthropic for advanced document creation and PowerPoint generation.
• A high-value enterprise use case is extracting structured data from contracts and research documents.
• Long-running agents could compress bank onboarding from weeks to about an hour.
• Box’s agentic harness is built on deep domain knowledge of how users search, select, and use files.
• The harness includes search, file-system access, text extraction, chunking, and on-the-fly embeddings.
• Box says this harness beats direct Claude or OpenAI API calls on accuracy and latency for Box workflows.
• Box runs hundreds of model tests, plus a document-centric complex-work eval and a holdback eval using employee behavior.
• Model quality in Box use cases mostly tracks coding-model performance, with Gemini standing out in some workflows.
• Customers typically use Box’s default model, but high token volume comes from workflow agents and extraction jobs.
• Open-weight adoption in enterprise is meaningful today, but far below where it will be in five years.
• Open weights can win on economics for mature workloads, but can still have odd failure modes.
• Mature use cases will likely peel off to open source or open weights once they stabilize.
• Enterprise AI will likely blend frontier orchestration with cheaper long-tail tasks, or reverse that pattern.
• Box sees permissions and access controls as a major enterprise AI constraint.
• The open question is what belongs in context versus what belongs in weights.
• Box Labs is Box’s applied AI team, focused on product-specific improvement rather than frontier research.
• Its mandate is to push agents another 10 points of accuracy on specific customer problems.
• For bank loan documents, 70% versus 97% accuracy is the difference between demo and automation.
• Enterprise agents must outperform off-the-shelf agents by 10 to 20 points inside the native product.
• Agents can also encode governance logic such as legal-hold and archive decisions.
• Systems of record need to go headless and expose APIs to Claude, ChatGPT, and MCP.
• Levie says he uses Salesforce more now because he can query it through Claude or ChatGPT via MCP.
• Enterprise UI will remain search-driven for ad hoc asks, but most work happens in background workflows.
• Levie predicts that within five years, 90% of enterprise tokens will be user-uninitiated.
• Coding is unusually automatable because code value maps closely to generated text.
• Coding agents are easier to deploy because GitHub provides a clean source of truth and developers can self-debug.
• Knowledge work lacks the GitHub-style onboarding path that made coding agents easy to start.
• Most non-Box enterprises still have data trapped in legacy file shares and on-prem systems.
• Box’s revenue run rate is about 1.3 billion, which limits how much of the market it can directly normalize.
• Enterprise diffusion will take longer than Silicon Valley expects.
• Box is using an internal AI center of excellence and specialist staff to target high-leverage workflows.
• Box also tracks token usage and uses those patterns to identify useful workflows.
• Levie is interested in Claude Tag as a Slack-native agent form factor.
• The moat in a faster-building AI era shifts toward distribution and customer workflow placement.