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

The Rise of Applied AI: Bridging Models to Enterprise Workflows

Aaron Levie, co-founder and CEO of Box, offers profound insights into Box’s transformation into an AI-first company and the broader AI landscape for a technical audience. He argues that the most significant value in AI is accruing not just to foundational model providers, but to application companies capable of bridging raw model capabilities to complex, real-world enterprise workflows.

Application Companies as the New Neolabs

Levie champions the concept of “application companies as the hottest Neolabs.” He observes that the market is validating “LLM wrappers” or “model wrappers” because enterprises fundamentally require a sophisticated bridge between a model’s intelligence and their specific operational processes. This “bridge” represents a vast, underappreciated area of value.

Companies that build deep expertise in this application layer—understanding domain nuances and workflow intricacies—will eventually accumulate sufficient data and insights to potentially develop their own specialized models, creating a powerful flywheel effect. This dynamic fuels opportunities for both “applied companies” (Neolabs) and their underlying infrastructure providers.

Model providers face a strategic dilemma: either extend their offerings up the stack to compete with application companies, or foster a robust ecosystem by leaving the application layer open. Levie draws parallels to the cloud computing era, where infrastructure (e.g., AWS) enabled immense value creation by applications (e.g., Snowflake, Databricks). He asserts that enterprises in regulated sectors like banking, life sciences, healthcare, and government demand tailored solutions, a need that AI has not eliminated. This necessitates a “diffusion economy” requiring extensive implementation and change management support.

The Economics of Tokens and Open-Weight Models

Levie addresses the potential for model providers to expand into the application layer, countering with two primary arguments:

1. Cost Optimization and Model Agnosticism: Enterprises prioritize cost-optimized task execution without compromising accuracy. A model-agnostic company, unbiased towards any single model, is best positioned to route tokens efficiently. This mitigates the “fox guarding the hen house” scenario where the token seller dictates model choice.
2. Temporary Subsidization: Levie believes the subsidization of tokens by model labs is transient. Publicly traded AI companies will eventually confront gross margin pressures and the high costs of training runs. Non-economic actors (e.g., Meta, SpaceX, nation-states, GPU providers like Nvidia), who can tolerate lower inference margins due to alternative strategic or revenue objectives, will drive down token costs. This reduction in cost per token will disproportionately benefit the application layer.

Consequently, Levie foresees a dynamic, competitive environment rather than a monopoly by a few large labs, arguing that a healthy ecosystem benefits all players and deters antitrust scrutiny.

He discusses the adoption of open-weight models, characterizing it as “higher than people think, lower than what enterprises actually want, and much, much, much, much lower than what it’ll be in five years.” While cost is a key driver, challenges such as token inefficiency and reliability issues remain. Referencing the Jevons Paradox, Levie suggests that as use cases mature, they will migrate to open-weight models, especially if they offer cost efficiencies or performance gains through post-training. This creates a blended approach: closed models for orchestration or complex tasks, and open weights for long-tail, cost-sensitive operations, potentially leading to significantly higher token consumption by open-weight models.

Box’s AI Reinvention: Unleashing Content with Agents

Box, leveraging two decades of experience with unstructured enterprise data, found itself uniquely positioned for the AI wave. Following early, expensive AI attempts in 2015, the “ChatGPT moment” in 2022 served as a pivotal catalyst. Box responded by fully pivoting, dedicating teams to build an AI stack and the “Box Agent.”

Box’s core use case capitalizes on its role as a repository for hundreds of billions of files (contracts, research, marketing assets) whose contents are often unknown without manual review. The Box Agent can:

• Query this vast dataset.
• Extract metadata and transform unstructured data into structured formats.
• Automate complex workflows.
• Generate new content, including advanced documents and presentations, often by orchestrating calls to external frontier models like those from OpenAI or Anthropic.

Hero Use Cases:

• Massive Document Processing: Analyzing millions of contracts or research documents to extract critical information, thereby solving a long-standing, prohibitively expensive enterprise challenge.
• Long-Running Agents: Automating entire workflows, such as streamlining a multi-week bank onboarding process to an hour by using background agents for automated steps and human-in-the-loop for exceptions. This extends to accelerating data discovery and security alerts.

“Work Slop” and the Differential Diffusion of AI

Levie highlights a stark contrast in AI adoption: it’s “embraced” in coding but often elicits an “allergic reaction” in content generation, termed “work slop.”

• Coding: Code is primarily viewed as a utility for automating tasks. AI-generated code is widely accepted, even preferred, for its efficiency. Developers, a highly technical audience, are adept at triaging and resolving issues with AI-generated output. Code data (e.g., in GitHub) is inherently well-structured and accessible for training.
• Content: Human-generated content is often a proxy for trust, unique thought processes, and execution capability. AI-generated content can dilute this, creating a “messy period” where recipients question the author’s intellectual involvement. Despite personal use of AI for idea generation, there’s a collective distrust of AI-generated content from others. This psychological factor differentiates it from fields like financial modeling, where macro-generated results are accepted.

This distinction underscores fundamental differences in knowledge work. Coding’s text-based, utility nature, direct correlation to productivity gains, and technically savvy user base make it highly amenable to AI automation. Other forms of knowledge work, such as sales, are constrained by external human factors and demand more complex change management and data integration.

Box’s Agentic Harness and Evaluation Metrics

Box’s “agentic harness” is a proprietary application layer designed to optimize AI performance on its platform. It leverages Box’s deep understanding of its file system, permission structures, and internal search capabilities. This domain knowledge allows the harness to significantly outperform raw API calls to models like Anthropic or OpenAI in both accuracy and latency. Its toolkit includes document chunking, embeddings, and intelligent search functionalities.

Evaluation Process: Box employs extensive evaluation to ensure agent effectiveness:

• Complex Work Eval: This test suite evaluates models against domain-specific, document-centric problems in sectors such as life sciences and financial services.
• Hold Back Eval: Utilizes Box’s internal data and employee usage patterns for continuous evaluation.
• Model Agnosticism: Box tracks incremental improvements across models, deploying defaults based on cost and accuracy. Customers retain the option to select models from a “model garden.”

Levie observes that model performance generally correlates with coding benchmarks, with exceptions (e.g., Gemini sometimes exhibiting disproportionate strength in specific Box use cases, possibly due to superior tool use). While most customers rely on Box’s default models, high-stakes applications like workflow agents or data extraction demand granular evaluation, often involving human-in-the-loop processes to achieve target accuracy (e.g., 97-98%).

Memory, Continual Learning, and Systems of Record

Levie delves into the future of AI beyond RAG (Retrieval Augmented Generation), exploring continual learning where model weights adapt to user interaction and data. While acknowledging its potential, he highlights significant enterprise challenges:

• Permissions and Access Controls: Unlike researchers with broad data access, enterprise users (e.g., lawyers) operate within highly segmented access protocols driven by legal, compliance, and competitive considerations. Training models on individual, dynamically changing access profiles presents immense complexity.
• Context vs. Weights: A critical question revolves around where to draw the line between information stored in context (for RAG) and information baked into model weights (for continual learning). This decision hinges on the rate of data change versus the performance gains derived from weight-baking.

He aligns with Karpathy’s concept of models encapsulating “reasoning capabilities” (e.g., a company’s ethos) while actual content resides in lookup systems. However, much value often lies in deep, specific IP knowledge (e.g., legal case details). Continual learning may prove more viable at a firm level, particularly for domain-specific applications like drug discovery, where broad internal information access is desired.

Regarding the role of systems of record (like Box) in an agentic world, Levie emphasizes a dual mandate for SaaS platforms:

1. Build a Superior Agent for Your Product: Develop an agent that is “insanely great” at using your own product, outperforming general agents by 10-20 points due to profound domain understanding. This will uncover unique, “white space” use cases (e.g., an agent proactively monitoring governance policies).
2. Go Headless: Expose robust APIs to external AI platforms (e.g., Claude, ChatGPT) and ensure your agent is accessible. This enables external agents to leverage your system of record, creating new value (e.g., querying CRM data via Claude).

Levie believes these systems of record, with their invaluable data, will attract volume-oriented business models. He envisions users paying significantly more for tools that allow AI agents to interact directly and intelligently with their core business data.

The Future UI of Enterprise AI: Beyond Chatbots

Levie contends that while universal chat systems will remain a powerful tool for one-off tasks, they will not be the dominant UI for enterprise AI. Enterprise work often involves complex, asynchronous processes and workflows that occur “behind the scenes,” frequently initiated by background agents rather than direct human chat commands.

• Background Agents: These agents will continuously perform tasks such as reading contracts, monitoring logs, or triaging security incidents, presenting results in dashboards, workflows, queues, or task lists for human review.
• Vertical vs. Horizontal: Attempting to integrate every workflow experience into a horizontal product will render it “heavy” and diminish its simplicity. Vertical players (applied companies) are better suited to comprehend specific workflows and deliver tailored user interfaces.

He predicts that within five years, 90% of all enterprise tokens will be consumed by tasks not initiated by a human user, with human interaction primarily focused on reviewing the results of these automated processes. This underscores the vast opportunity for applied AI companies to deliver domain-specific, workflow-aware solutions.

Staying AI-First and Winning the Race

Levie attributes his “wired-in” status to Box’s historical focus on unstructured data and his personal fascination with AI. He advocates for Twitter (X) as the “global town square for AI,” recommending it as the primary source for staying current with emerging trends.

For company building, Box benefits from a disciplined data architecture, where nearly all internal documents are stored and managed within Box, providing a clean, accessible dataset for AI agents. Key practices include:

• Identifying high-leverage workflows.
• Establishing an internal Center of Excellence with dedicated AI FTEs.
• Utilizing “leaderboards” (lists of top AI users) to identify effective use cases and facilitate internal training.

Levie acknowledges the incredible leverage new founders possess today (two people accomplishing what 40 once did) but also notes the intense competitive landscape. In an AI-accelerated world where building is faster, winning shifts to who can get it to the customer effectively. He believes the “trillions to trillions” of value available at the applied layer will be captured by companies with the teams and expertise to “get to the enterprise.”

• Tags:
• Feature
• Enterprise AI
• AI Strategy
• Workflow Automation
• Agentic Workflows
• Open-Weight Models
• Founder Advice
• Continual Learning
• Systems of Record
• Data Architecture
• Applied AI
• Token Economics
• Aaron Levie
• Box
• LLM Wrappers
• Model-Agnostic
• Work Slop
• AI Diffusion
• Content Management
• Customer Adoption