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

Aaron Levie agrees that application companies are becoming the hottest part of the AI market. He says the gap between an LLM’s raw capability and an enterprise’s actual workflow is much larger than many researchers assume: agents still need access to other data systems, human-in-the-loop steps, delayed processes, legacy software, and business-process change management. That creates opportunities both for applied companies such as Box and for infrastructure providers serving them. Levie compares this moment with AWS, GCP, and Azure enabling companies such as Snowflake and Databricks, which cloud providers could not simply replace.

Levie says model labs face a strategic tension: moving up the stack brings them closer to customers, but leaving room for an application ecosystem builds trust and distribution. A model-agnostic company can optimize each task for cost and accuracy without favoring the model it sells, while model providers may subsidize inference only temporarily. He expects costs per token to fall as competitors such as Nvidia, China, and other well-capitalized actors pursue different objectives. He doubts one or two labs will capture 95% of AI’s value, partly because excessive concentration could invite antitrust intervention or nationalization.

Levie explains that Box began by securely storing and sharing enterprise data in the cloud, then expanded into managing unstructured files such as contracts, research documents, financial records, and marketing assets. Box experimented with image classification, OCR, and other AI applications around 2015, but those systems were expensive and required a separate model for each workflow. The ChatGPT launch changed the company’s view of what was possible, prompting a company-wide pivot and the development of Box’s AI stack and Box Agent. Because Box sits on hundreds of billions of files, agents can search, extract metadata, answer questions, and automate workflows around information companies rarely had the labor to read.

Levie identifies a straightforward “hero” use case: an enterprise can ask Box to inspect a million contracts or research documents, extract structured data, and make it searchable or usable in automated workflows. The more ambitious goal is a long-running agent that executes an entire process, such as onboarding a bank customer. Instead of a one- or two-week sequence of manual exchanges, the agent would automate every possible step and escalate only unusual cases for human review, potentially reducing the process to about an hour. Other examples include discovering research data and flagging security events faster.

Discussing “workslop,” Levie says code is treated largely as a utility: for most users, the goal is to automate something, so AI-generated code is preferable if it removes a bottleneck. Content is different because a presentation, memo, or strategy document is often used to judge whether someone understands a subject and can execute. Levie admits he now reads material twice—once for its substance and once to calculate whether the author wrote it. He compares the transition to using a calculator, but says the analogy is imperfect because AI may perform many tasks at once. He would not want to be a college professor deciding what students actually did.

Box built an agentic harness around its file system, permissions, search engine, and knowledge of how users select relevant documents. The harness can search repeatedly, rerank results, extract text or document chunks, generate embeddings, and decide when to read a full file. Levie says Box sees better accuracy and latency than simply sending the same task to Claude or an OpenAI API. Box evaluates hundreds of scenarios through its Complex Work Eval across areas such as life sciences, financial services, government, and technology, as well as a holdback evaluation based on how Box employees use their own data.

Levie says model performance on Box’s use cases generally tracks coding benchmarks, although Gemini is disproportionately strong in some knowledge-work and tool-use tasks. He describes the frontier race as close among several model families and says customers usually accept Box’s default model but sometimes evaluate multiple models for specific cost-and-accuracy requirements. Open-weight adoption is higher than many expect but lower than enterprises ultimately want; some of the current demand reflects the appeal of experimenting with new systems. Over time, mature workloads can be peeled off to cheaper open-weight models, while difficult tasks are escalated to expensive frontier models.

Levie is interested in systems that incorporate user or company context into model weights rather than retrieving everything through RAG, but he warns that enterprise access controls make this difficult. A researcher may have broad access, while a lawyer may be restricted to five projects and barred from seeing a competitor’s work. Users can also gain or lose access daily, changing the context a personalized model is allowed to use. He expects progress in continual learning and domain-specific models, especially in areas such as drug discovery, but says the right split between information stored in weights and information retrieved at runtime depends on how quickly the data changes and how much performance improves.

Levie says SaaS companies cannot choose between building an excellent agent for their own product and making their data accessible to outside agents: they need both. A specialized agent should be 10 or 20 percentage points better than a generic agent because it understands the product’s workflow and can uncover new use cases, such as an always-on governance agent warning an employee that a document violates retention or legal-hold policies. At the same time, systems of record must expose APIs or agents through tools such as MCP so Claude, ChatGPT, and other systems can use their data. He says users will pay for this value; he would pay much more for LinkedIn if he could query it through MCP.

Levie attributes coding’s rapid AI adoption to several unusual advantages: code is text whose output directly creates software, labs benchmark themselves on coding, engineers can diagnose technical failures without waiting for IT, and even modest productivity gains are valuable in a highly paid field. Sales, by contrast, depends on customers responding, scheduling meetings, and having budget, so AI cannot remove those external constraints. Other knowledge-work domains also have fragmented data, legacy systems, and complicated permissions—there is no universal “give us your GitHub” connection. Levie therefore expects enterprise diffusion to take longer and says the applied companies that can implement AI and reach customers will capture a large share of the opportunity.

Levie says Box was well positioned because it had spent 20 years organizing unstructured enterprise data, but the company still had to bet on AI and respond to customer demand. Box uses an AI center of excellence, internal AI-focused engineers, usage leaderboards, and demonstrations from employees who discover effective workflows; when one employee showed a team how he worked with AI, Box organized a training session within hours. Levie’s personal advice is to join Twitter and follow roughly 20 to 100 informed accounts, because a person’s feed can leave them a year ahead or behind. He says young founders have extraordinary leverage, but must combine fast building with enterprise distribution and trust.