By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Sequoia CapitalBox's Aaron Levie: On Reinventing Yourself in the AI Age and Enterprise Diffusion
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
CHAPTERS
- 0:00 – 1:55
Staying "wired in": why your information feed shapes your career
Aaron Levie argues that being current in AI and tech is largely a function of what you follow and where you spend attention. He describes his own habit of playing with new tools and staying continuously plugged into what’s happening online.
- •Aaron Levie says people can be “a year ahead or a year behind” based on their feed
- •He frames being “wired in” as a practical career advantage, especially for students and early-career folks
- •He contrasts actively tracking information with passively “seeing articles”
- •He emphasizes experimentation and regularly trying new products
- 1:55 – 6:22
Why "application companies are the hottest neolabs" and the real enterprise bridge
Sonya Huang asks whether application companies are emerging as the most exciting new AI labs. Aaron agrees and explains that enterprises need a deep bridge between model capability and real workflows, which creates large opportunity in the application layer and its supporting infrastructure.
- •Aaron Levie says the “LM wrapper” idea is working because enterprises need workflow bridges
- •He describes a major market bet: the bridge is either limited or vast—and he believes it’s vast
- •He argues bridging the gap can later enable building models through data and domain flywheels
- •He sees multiple startup layers: applied companies and infrastructure providers
- 6:22 – 12:34
Will model labs move up the stack? Ecosystems, trust, and cost optimization
They explore whether model providers will compete directly with application-layer companies. Aaron highlights the strategic tension: being closer to customers versus preserving an ecosystem customers trust, plus the importance of cost-optimized model choice across tasks.
- •Aaron Levie predicts enterprises will still need workflow- and domain-specific integration
- •He describes chatbot universality versus agentic workflows needing data access and setup
- •He argues customers want task routing that’s cost-optimized without vendor preference
- •He questions how long token subsidization can last, framing it as a gross-margin issue
- •He suggests multiple actors with different incentives can drive inference costs down
- 12:34 – 16:17
Box’s evolution and the company-wide bet on AI agents
Sonya transitions to Box’s history and how the company is reinventing with AI. Aaron recounts Box’s origins in secure cloud storage and enterprise collaboration, prior experiments with earlier AI waves, and the rapid pivot after the ChatGPT moment toward an AI stack and Box Agent.
- •Aaron Levie describes Box’s initial mission: secure cloud storage, sharing, and collaboration for enterprises
- •He recalls earlier AI efforts like image classification and OCR being expensive and use-case-specific
- •He describes renewed momentum with GPT-era models and the interactive chat form factor
- •He says Box went “all in,” built an AI stack, and focused on deploying agents over enterprise content
- •He emphasizes the scale of unstructured content and the need to understand what’s inside files
- 16:17 – 16:54
Content creation and partner models: documents and PowerPoint workflows
Sonya asks whether Box uses agents to create new content. Aaron says yes, both inside collaborative documents and via integrations with frontier model providers to generate advanced documents like PowerPoints.
- •Aaron Levie says agents can generate content within Box’s collaborative products
- •He describes work with OpenAI and Anthropic for advanced document and PowerPoint creation
- •He frames frontier model providers as likely to stay ahead on certain creation capabilities
- •He emphasizes agent-driven interaction with external systems to produce higher-quality outputs
- 16:54 – 18:42
Hero use cases: a million contracts, extracted structure, and long-running workflows
Aaron outlines the strongest enterprise wins today and what he expects to matter over the next few years. He highlights large-scale document understanding (contracts and research) and the rise of long-running agents that execute workflows end-to-end with human review at key checkpoints.
- •Aaron Levie cites scanning “a million contracts” to discover what’s inside as a standout use case
- •He describes extracting metadata and converting unstructured documents into structured databases
- •He anticipates long-running agents executing workflows with humans in the loop for verification
- •He gives an example vision: bank onboarding compressed from weeks to roughly an hour
- •He frames the goal as faster discovery, alerts, and process automation across enterprise workflows
- 18:42 – 24:08
Work slop: why AI code feels acceptable but AI-written content triggers distrust
They discuss the cultural pushback to AI-generated content compared with the embrace of AI-generated code. Aaron argues code is often treated as utility output, while documents and decks act as proxies for trust, competence, and ownership of thinking—creating a messy transition period.
- •Aaron Levie says code is largely viewed as a utility, making AI assistance feel preferable
- •He contrasts this with decks and writing that signal trust and understanding of the author
- •He describes the psychological “double work” of reading for substance and for AI-authorship detection
- •He predicts society may need years to adjust norms around AI-generated professional content
- •He notes analogies like calculators partly fit, but break down as systems perform many tasks at once
- 24:08 – 26:15
Box’s agentic harness: tool access, search, reranking, and why tuned workflows win
Sonya asks how the Box Agent works. Aaron describes an “agentic harness” that leverages Box’s deep understanding of permissions, search behavior, and document retrieval, plus toolkits for extracting text, chunking, embeddings, and multi-step querying to improve accuracy and latency.
- •Aaron Levie says Box knows user heuristics for search and document relevance inside its system
- •He describes a harness with tools for search, file access, text extraction, chunking, and embeddings
- •He explains the agent does multiple searches, reranks results, and sometimes reads full documents
- •He claims this tuned harness can outperform simply giving an external model Box’s API
- •He frames it as a specialized system for asking questions across very large file collections
- 26:15 – 27:23
Evals that matter: Complex Work Eval, holdbacks, and the model garden
Aaron explains how Box evaluates model performance and selects defaults. He describes domain-specific document-centric tests, internal holdback evaluation based on employee usage, tracking small changes in performance, and allowing customers to choose among models in a “model garden.”
- •Aaron Levie says Box runs hundreds of tests on models
- •He describes the Complex Work Eval across domains like life sciences, financial services, public sector, and tech
- •He explains an additional holdback eval based on Box’s own internal usage patterns
- •He mentions tracking incremental improvements and choosing defaults based on cost/accuracy thresholds
- •He says customers can choose models, while Box also provides tuned defaults
- 27:23 – 29:25
The model race through Box’s lens: correlations with coding and a few surprises
Sonya asks Aaron’s view on the model performance race. Aaron says results mostly correlate with coding benchmarks, while noting one exception where Gemini performs disproportionately well for some Box use cases, and he describes the competitive “neck and neck” feel across leading models.
- •Aaron Levie says model performance generally correlates with coding benchmarks
- •He notes Gemini is disproportionately better than expected in some Box document/workflow scenarios
- •He references external evals and says they broadly track coding model performance patterns
- •He characterizes the environment as a tight race with rapid iteration and constant rumors
- 29:25 – 32:35
Enterprise adoption of open-weight models: hedging, reliability gaps, and long-term blending
They discuss how much enterprise demand exists for open-weight models and what drives adoption. Aaron says adoption is higher than many expect but below what enterprises say they want, with “sexiness” and hedging as meaningful drivers today, and a future where workloads are peeled off into cheaper models as they mature.
- •Aaron Levie says open-weight adoption is higher than people think but far below where it will be
- •He attributes some usage to the appeal of trying open source as a hedge
- •He expects future adoption to increase as stable workloads shift to cheaper inference economics
- •He notes current issues like token inefficiency and occasional odd behaviors making enterprises cautious
- •He describes blended systems where orchestration routes tasks between frontier and cheaper models
- 32:35 – 37:29
Memory and continual learning: what belongs in context vs weights (and enterprise permission complexity)
Sonya raises personalization, memory, and continual learning beyond RAG. Aaron is intrigued but emphasizes enterprise constraints like permissions and access control changes, and he focuses on the hard question of deciding what should be baked into weights versus retrieved in context.
- •Aaron Levie highlights permission complexity as a core challenge for continual-learning approaches
- •He contrasts a researcher’s broad access with a lawyer’s strict compartmentalization requirements
- •He questions how models should adapt when user access changes daily across projects
- •He calls out the decision framework needed for what goes in context versus weights
- •He suggests firm-level or domain-specific customization may be more feasible in some verticals
- 37:29 – 39:02
Box Labs and the applied research agenda: hill-climbing accuracy for real automation
Sonya asks what Box Labs does and what research areas matter most. Aaron frames it as applied AI work focused on raising accuracy in real client workflows, including systematic methods to “hill climb” from baseline model performance to automation-grade reliability for tasks like loan document extraction.
- •Aaron Levie describes Box Labs as an applied AI team (not necessarily “capital L” yet)
- •He says a key focus is improving agent accuracy by meaningful margins on specific tasks
- •He describes work on automated approaches to improve answer quality over client data sets
- •He emphasizes the practical difference between ~70% and ~97% accuracy for automating enterprise processes
- 39:02 – 44:55
Systems of record in an agent world: build a great in-product agent and go headless for others
They discuss the role of systems of record amid many competing agents and interfaces. Aaron argues SaaS platforms must do both: build an agent that is significantly better at their product than a generic agent, and also expose APIs or headless capabilities so external agents can operate deterministically via integrations like MCP.
- •Aaron Levie says systems of record must deliver an agent that’s provably better for their domain/workflow
- •He argues companies also must expose APIs and headless integration so external agents can use the system
- •He emphasizes new “white space” use cases that systems of record can uniquely unlock
- •He gives an example use case: background governance agents that detect and prevent policy violations
- •He describes how MCP-style access increases how much he uses systems like Salesforce for queries
- 44:55 – 48:01
Beyond chat as the UI: background agents, dashboards, queues—and most tokens happening automatically
Sonya asks whether chat will be the dominant enterprise AI interface. Aaron says chat will remain a universal entry point, but most enterprise value will come from asynchronous, background agents that monitor, process, and triage work, surfacing results through dashboards, task lists, and workflow queues.
- •Aaron Levie says chat UIs will be ubiquitous like search boxes are today
- •He argues enterprise work is mostly processes running behind the scenes, not one-off chat requests
- •He lists non-chat interfaces enterprises will want: dashboards, workflows, queues, and task lists
- •He predicts the applied layer will thrive because it can tailor UI and controls to specific workflows
- •He bets most enterprise tokens will be spent on tasks users didn’t directly initiate
- 48:01 – 54:32
Why coding diffused fast while knowledge work lags: rate limits, data location, and change management
Aaron explains why AI adoption exploded in coding but is slower elsewhere. He cites coding’s text-based output, strong benchmarking focus, and technical users who can self-troubleshoot, contrasting that with other roles constrained by external dependencies, messy enterprise data environments, and access control realities.
- •Aaron Levie says code output maps directly to value creation and is heavily optimized by labs
- •He notes engineers can troubleshoot and wire tools themselves, accelerating adoption
- •He contrasts other roles (like sales) that are rate-limited by customer behavior and timing
- •He highlights missing “GitHub-like” standardized data access for most knowledge work
- •He argues enterprise access controls, legacy systems, and change management slow diffusion
- 54:32 – 1:05:21
Making a company AI-first and staying current: internal champions, training, and relentless curiosity
Sonya closes by asking how founders and companies reinvent for the AI wave. Aaron attributes his position to Box’s long exposure to unstructured data plus personal fascination, and he describes learning primarily through Twitter and podcasts, while operationally emphasizing data hygiene, internal enablement, and constantly evolving workflows.
- •Aaron Levie says Box was “well primed” for agents due to two decades of unstructured data work
- •He describes learning primarily through Twitter, plus specific podcasts including Sonya’s
- •He explains Box benefits from centralized documentation and strong data hygiene for agent deployment
- •He describes internal structures like an AI center of excellence and internal FDEs
- •He mentions tracking internal usage patterns and running trainings to spread effective workflows
- •He notes enterprise constraints like security and code review limit how far automation can go compared to labs
