Box's Aaron Levie: On Reinventing Yourself in the AI Age and Enterprise Diffusion
In a Nutshell
Aaron Levie argues that the trillion-dollar opportunity in AI lies at the application layer—building agents that connect frontier models to real enterprise workflows—rather than in model development itself. Box's success stems from deploying domain-specific agents against hundreds of billions of unstructured enterprise files to automate contract processing, data extraction, and governance workflows that previously required expensive manual labor. The biggest winners will be companies that master enterprise diffusion: getting agents into complex legacy systems with proper permissions, change management, and measurable accuracy gains, as coding agents proved easier to deploy than knowledge work AI due to simpler integration and technical user bases.
These notes were generated by AI and may contain inaccuracies.
Aaron Levie emphasizes the importance of following specific Twitter accounts and joining the platform to stay connected with current developments. He notes that being wired in through your feed can put you a year ahead or behind in your career. He observes that many 20-year-olds still rely on articles being emailed to them rather than actively staying connected to information flows.
Levie agrees that application companies represent the hottest segment in the current market, noting this trend has become clearer over the past two years. He attributes this largely to open source developments, explaining that LLM wrappers or model wrappers are succeeding because enterprises need bridges between model capabilities and actual workflows.
He describes how a trillion dollars has been bet on whether this bridge between models and workflows will be limited or vast. The bet centers on whether organizations should focus solely on the model itself and super intelligence, or on the application tier that connects intelligence to real workflows.
Levie explains that there's a significant gap between model capabilities and actual enterprise workflows. Bridging this gap requires connecting to other data systems, incorporating human-in-the-loop interactions, managing workflow delays, handling change management of business processes, and dealing with legacy systems.
He notes that research organizations typically don't want to attack these operational challenges, creating opportunity for application-focused companies. He draws a parallel to infrastructure versus application layers, citing how AWS, GCP, and Azure created trillions in infrastructure value while Snowflake and Databricks emerged as successful application-layer companies.
Levie identifies a key strategic tension for model providers over the next two to five years: how much they need to move up the stack to compete at the application layer versus leaving that space open for the application ecosystem. This represents a tension between getting closer to customers versus maintaining an ecosystem that trusts them.
He observes that VCs face similar existential decisions about whether to invest in model companies or application-layer opportunities.
Levie argues that AI hasn't meaningfully changed the fundamental need for technology that bridges core capabilities to enterprise workflows and business processes. When visiting law firms, pharma companies, or banks, these organizations need solutions that understand their specific workflows and can access core enterprise data.
He contrasts chatbots, which can be universal and horizontal, with agentic workflows that require domain expertise, timing within processes, and access to specific data that chatbots cannot natively reach.
Levie predicts that unless model labs scale to 100,000 employees, the diffusion economy will be massive. Every company from 50-person firms to multi-hundred-thousand person organizations will need teams to help with transformation and change management.
He emphasizes that domain expertise and understanding of specific business processes will be critical, creating significant value for companies that can provide this expertise.
Levie identifies the conflict of interest where model sellers also meter and gate which tokens work best for each use case. For agentic systems, customers want tasks to be cost-optimized with accuracy held constant, which requires a company without preference among different models.
Levie argues that model company subsidization of tokens is temporary because public companies will face standard capitalist pressures. He notes that subsidization works up to certain spend thresholds, and companies may have exceeded these when reaching tens of billions in capital requirements.
He points out that model companies with high gross margins on inference can afford to price APIs higher than first-party products, creating an equilibrium they must strike between API revenue and applied product adoption.
Levie identifies four cohorts that don't need to make money on inference at the same margin structure as Anthropic or OpenAI: Meta, SpaceX, China, and Nvidia. These actors might accept 10% margins on inference because they want to pay for massive compute clusters, which would drive down cost per token and accrue more value to the application layer.
Levie predicts value will exist throughout the stack rather than one or two labs capturing 95% of value creation. He suggests the biggest labs would prefer this outcome to avoid antitrust issues, as being the sole source of intelligence could lead to nationalization.
Box started as a secure cloud storage and sharing solution, scaling quickly by pivoting to enterprise needs. The company helps organizations move from on-premises systems to the cloud for storing, sharing, collaborating, and managing unstructured data including corporate documents, financial documents, marketing assets, and research materials.
Box has been working on AI products since 2015, initially exploring early models for image classification and OCR. However, these were expensive and required hyper-trained models for each use case, leading them to shelve the effort.
After GPT developments, they conducted hackathons exploring applications like type-ahead in note-taking products and text detection for security use cases. The ChatGPT moment prompted a full company pivot with dedicated teams and daily updates.
Box sits on hundreds of billions of files containing critical enterprise information including contracts, research files, marketing assets, and loan documents. The core problem is that organizations rarely know what's inside these files without manual review.
Box's platform deploys agents against unstructured data to answer questions, pre-process documents, extract metadata, convert it to structured data, and automate workflow steps.
Box enables agents to generate content within their collaborative products and has partnered with OpenAI and Anthropic for advanced document and PowerPoint creation. They use frontier technology from these providers through agents that interact with their systems.
The primary enterprise use case involves processing millions of contracts or research documents to extract structured data, populate databases, and enable querying, analysis, and workflow automation. This addresses long-standing problems where human labor was too expensive to apply at scale.
More advanced applications involve long-running agents executing entire workflows, such as bank client onboarding where automation handles every possible step and only escalates to humans for review or verification, reducing processes from weeks to hours.
Levie distinguishes between code and content domains regarding AI acceptance. Code is viewed as a utility designed to accomplish tasks and automate processes, making AI generation preferable when it removes blocking obstacles.
In content domains, presentations carry symbolic weight about whether the presenter can be trusted to execute, deliver results, or understand topics. This creates resistance to AI-generated content that differs from the code domain's utility-focused perspective.
There's uncertainty about how much of thought processes come from humans versus AI, raising questions about whether it matters. This creates a collective issue where people use AI for their own brainstorms and decisions, but when receiving work from others, there's hesitation about trust. This is a societal challenge that will need to work through over the next three to five years.
The financial model analogy illustrates this shift. People don't question whether financial models were generated by macros rather than personally computed, yet the same acceptance doesn't currently exist for strategy decks. The current evolution involves determining what a person's role is and what content serves as a proxy for - whether it's meant to represent how much someone knows or what they can execute on.
The Stan Ducken Miller Wall Street Journal piece served as a counterexample to typical AI writing reactions. Unlike content that triggers an allergic reaction when clearly AI-generated, this piece didn't produce that response. The question remains whether trust comes from knowing the author (Stan) versus the writing quality itself.
Reading AI-generated articles now requires double the mental processing: first for substance, second to calculate whether the person wrote it or if it's essentially reading a Claude prompt. This affects judgment of both the person and the post.
College professors face particular challenges since it's increasingly difficult to determine what work students actually did themselves.
While the calculator analogy seems closest, it breaks down because AI handles at least 10 tasks simultaneously rather than finite calculations. The discussion transitions to harnesses and Box's approach.
Box built an agentic harness based on deep understanding of their file system, permission structures, and search engine. They offer labs access to this data to improve external agents. The harness incorporates knowledge about how people search Box, decide which files to select from search results, and their internal calculus for determining relevant documents.
The harness has access to search systems and file systems with mechanisms for pulling document text, extracting chunks, and generating embeddings on the fly. It functions as a harness for asking questions of large datasets - with tens of millions of files accumulated over 20 years in individual accounts.
The agent performs multiple searches, reranks results, quickly pulls relevant information, and in some cases reads full documents. Box sees meaningfully better accuracy and latency results compared to giving Claude or OpenAI their API directly, due to tuning for specific workflows.
Box maintains hundreds of different tests across every model with two main evaluation frameworks: the Complex Work Eval covering domain-specific work in life sciences, financial services, public sector, and tech as document-centric evaluations; and a hold-back eval using Box employees' actual data usage patterns.
Model performance generally correlates with coding benchmarks, with Gemini as an exception showing disproportionate strength in some use cases, possibly due to better tool use. Fable 5.1 was identified as state-of-the-art among tested models. Model selection involves rolling out defaults based on cost and accuracy thresholds while allowing customers to choose any model from their model garden.
By volume, customers use Box's default models due to ease and performance tuning. Token volume tends to flow through workflow agents or data extraction rather than simple search interactions. For data extraction use cases, customers often conduct their own evaluations to ensure specific cost profiles achieve 98% accuracy targets, with FTE support for model testing.
Open weight model adoption is higher than commonly perceived but lower than enterprises actually want, and significantly lower than projected levels in five years. Cost drives approximately 30% of decisions, with the remainder attributed to the "sexiness" of trying models like GLM. Fortune 500 CIOs experiment with open source even when established models would suffice at similar cost profiles, driven by hedging strategies.
Long-term, open weights will gain economic advantage as workloads can be peeled off based purely on inference costs. Current challenges include token inefficiency and occasional issues like randomly speaking Chinese mid-chain, which creates problems for banking applications.
The Jesse at Decagon post describes a paradox where closed models continue exponential improvement while mature use cases can be transitioned to open source. Once use cases achieve stability, they make economic sense to shift to open weights models if either literally cheaper or if post-training delivers meaningful performance gains.
This creates a duality where the pie grows fast enough that both approaches win: using Fable or GPT-5.6 for orchestration while farming longtail tasks to cheaper models, or using cheaper orchestration agents that occasionally escalate difficult tasks to heavier models.
Current dominant architecture remains RAG-based systems with context lookup rather than fundamental weight changes. Continual learning - where model weights adapt as they get to know users - represents the direction labs are exploring. Engram works with customers to bake context into weights themselves.
Enterprise environments present unique challenges due to complex permission structures and access controls. Lawyers working on competitive projects require hard barriers preventing document overlap, creating scenarios where daily additions or removals from access contexts would require constant model retraining.
The decision point between what goes into context versus weights requires careful consideration of massive performance gains from baking information into weights. A rubric might evaluate when data change rates aren't extreme but weight incorporation dramatically improves accuracy.
Karpathi's concept of removing memorized information from models while retaining reasoning capabilities, with actual content handled through lookup systems, represents an ideal architecture if achievable.
Enterprise value may lie more in depth of knowledge about specific cases and application to other projects rather than distinct execution styles. Domain-specific approaches make sense for certain verticals - pharmaceutical companies like Eli Lilly want models specifically for drug discovery workflows without the church and state problems present in other industries.
Box Labs functions as their applied AI team, focusing on taking agents and achieving another 10 points of accuracy improvement on given problems. Research includes building problem set maps with given data to best execute tasks, and developing auto research at the agent and harness level for hill climbing on accuracy when answering questions from client data.
For banking use cases with 100-page loan documents, improving from 70% to 97% accuracy represents a critical difference in automation feasibility, requiring substantial system work to achieve these improvements.
Every software company currently pushes their own agents, but customers want systems of record that work well with their chosen agents rather than additional agent offerings. For SaaS platforms with customer data and workflows, two requirements emerge: building agents that are provably 10-20 points better than off-the-shelf agents at using the specific system, and identifying use cases that haven't been productized because of direct customer interaction.
One example involves agents working in the background to determine when documents meet governance policies, identifying items needing archiving or legal holds. Enterprise heads of compliance and governance could theoretically oversee entire organizations simultaneously through such agent capabilities.
Agents sitting next to employees can automatically identify when actions would break governance policy, rather than requiring users to predict or understand policy constraints. An ongoing agent that automatically flags governance violations enables earlier and better identification than off-platform solutions, allowing companies with agents within their products to identify issues sooner and better than the rest of the market.
Companies must expose APIs to cloud platforms, ChatGPT, and other systems to enable agent integration. This requires either direct access to deterministic APIs that agents can call via MCP or making agents headless and exposed in external systems. The commercial model must make sense for systems of record, ensuring value flows appropriately to the platform.
The misconception that systems of record would lose value underestimated the new use cases and white space opportunities created by AI. Aaron Levie uses Salesforce an order of magnitude more today by connecting via Claude or ChatGPT to query CRM data.
Systems of record have depth of purpose in organizing workflows, managing data, securing data, and providing guardrails. The business model needs to be volume-oriented, but solving real customer problems naturally generates revenue. Levie mentioned he would pay 10x more for LinkedIn if he could MCP into it to always understand when a CIO is doing something relevant for outreach.
Systems with valuable data will find customers rewarding that value creation when companies deliver good results.
Universal chat systems for one-off agent queries will remain a mainstay across horizontal and vertical products, similar to how every product has a search box. However, enterprise work consists of processes and workflows happening behind the scenes with computers, other people, or tasks that should happen but couldn't be afforded.
This creates demand for agents working asynchronously in the background to review contracts, examine logs, and triage security incidents. Rather than constant chatting, users want dashboards, workflows, queues, and task lists to catch up on agent work.
Horizontal products risk becoming too heavy when trying to manifest every workflow component, while vertical players understand specific processes and can manifest appropriate buttons, tabs, and terminology. As agents do more background and async work, this favors applied companies that understand domain-specific workflows.
Harvey and other legal AI tools demonstrate this pattern, with similar emergence in security and long-running coding agents like Cognition and Factory. Levie predicts that in five years, 90% of enterprise tokens will be from work users never initiated, with results appearing as tasks requiring review.
Coding has unique properties enabling rapid diffusion: utility almost entirely represented by generated text, models hyper-trained on code, AI labs using coding as competitive benchmarks with daily evals, and the most technical audience that can fix bugs and triage problems themselves.
Coding represents a high-paying vertical where even 10-20% productivity gains justify investment. The continuum ranges from roles rate-limited by external factors like sales (customer response, meeting availability, budget) to coding where someone could sit at a computer typing text all day.
Knowledge work lacks the simple GitHub integration that coding agents have. Most enterprise data exists in on-premises systems, legacy file shares, and infrastructure that doesn't interface well with agents. Box's $1.3 billion revenue run rate means many organizations still use legacy systems.
Access controls present additional complexity - in coding, practitioners have access to most relevant materials, but knowledge work requires navigating permissions like asking Sally to open file shares or projects.
Silicon Valley must prepare for diffusion taking longer than anticipated, but this represents opportunity for the applied layer with patience, domain expertise, and work ethic. The trillion-dollar applied layer AI opportunity involves getting technology to lawyers, sales reps, life sciences researchers, and customer support teams.
Two factors enabled Levie's position in AI: 20 years working with unstructured data made the benefits of agents immediately apparent, and the product shape already aligned with agent use cases. Box customers requesting ability to read every document and answer questions created positive feedback loops.
Levie stays current through podcasts including his own and Doresh's, plus Twitter as the global town square for AI. He recommends following specific accounts and joining Twitter as the number one advice for staying current on AI developments.
Levie uses five different AI systems daily for 20-30 questions on research, talent acquisition, competitor analysis, market expansion, and related tasks. Personal assistant products show limits with browser use, suggesting the internet needs CLI interfaces for products.
Box benefits from data architecture tuned for unstructured data - meeting notes, project plans, roadmaps, presentations, and financial documents - making everything queryable. Data hygiene across Salesforce and core systems enables easier agent deployment.
Best practices include identifying highest-leverage impact workflows, establishing AI centers of excellence, hiring internal AI FTEs, and tracking token usage to identify both wasteful patterns and successful approaches for internal training. Mick was highlighted as an example user whose AI workflow was demonstrated to his team.
Box ships 2-3 times more customer-facing products in some stack areas compared to pre-AI periods. The meta constantly changes from IDE plugins to Cursor rollouts to Slack bot mentions for code work. Cloud tags represent an exciting form factor, though team constructs and data requirements must be addressed.
Young founders can start companies with AI-first design principles and achieve in two-person teams what previously required 40-person projects. Established companies like Box face constraints around code review and security practices that customers require for data security and compliance, creating productivity discounts.
The history of business during this AI transformation period will be written about as extraordinary, with the art of business requiring new frameworks beyond traditional approaches.
In today's environment, every great idea instantly attracts five competitors, creating a fundamentally different landscape than existed previously. Earlier, companies enjoyed a couple of years to focus on grinding out product development and experience improvements without facing constant pressure from new entrants. During that earlier period, there wasn't the same phenomenon of seeing Sequoia or Benchmark funding competing companies every three days. While the speaker acknowledges they had their own version of competitive pressure at the time, in retrospect it was not worthy of the level of concern it generated.
The consensus view is that in a world where AI builds things so much faster, the advantage shifts to whoever can actually get the product to the customer. Founders who understand this mandate recognize that success is fundamentally an enterprise diffusion play - requiring companies to get their solutions into enterprise environments. Those who mistake this mandate will lose, as there are quite literally trillions of dollars up for grabs at the applied layer. The companies that know how to build the teams and get to the enterprise will be the ones that win.
"It's game over. Sorry. Like there is quite literally trillion to trillions up for grab at the applied layer and the companies that that know how to build the teams and and get to the enterprise will be the ones that win. It's just like obviously guaranteed."
Keep Sequoia Capital in your library
Save the videos and channels worth coming back to, and find them again in one place.
