Uncapped with Jack AltmanAgents in the Enterprise | Aaron Levie, CEO of Box | Ep. 3
CHAPTERS
Why AI is a breakthrough for Box: unlocking the value trapped in enterprise content
Aaron frames AI as the missing layer that turns decades of stored enterprise files into usable, queryable organizational knowledge. He explains how most company data goes “cold” after initial collaboration, even though it contains insights that could drive product, sales, and onboarding outcomes.
- •Enterprises store huge value in contracts, financial docs, marketing assets, HR records—but largely don’t utilize it
- •Traditional file lifecycle: create/share/collaborate, then archive for compliance or future retrieval
- •AI enables extracting insights and answering questions from existing content
- •Box’s scale (large customer base and Fortune 500 penetration) makes it well-positioned to deploy AI broadly
- •Mindset shift: operate as if Box were founded today in an AI-native era to avoid slow-incumbent failure modes
What “agents” do inside Box: automating content-heavy workflows
Aaron describes agents as specialized assistants that execute content-oriented tasks within Box, often in the background. He gives concrete examples across legal, procurement, marketing, and finance, emphasizing that agents will become plentiful and highly tailored to workflows.
- •Agents focus on content workflows: review, extraction, summarization, routing, and automation
- •Examples: legal clause review; procurement invoice/payment term checks; marketing asset analysis and campaign workflow automation
- •Financial analyst agent example: synthesize internal documents + notes into an industry trends report
- •Agents can connect to tools and external data sources (e.g., web/tool use) to enhance outputs
- •Box AI Studio provides early “primitives” for customers to build these agents
Cross-app agents: making enterprise systems talk to each other
Beyond Box, Aaron argues the real future requires agents that orchestrate work across many SaaS systems. He outlines an emerging ecosystem of protocols and agent-to-agent communication so a single query or workflow can pull from Box, Salesforce, ServiceNow, and beyond.
- •Work is distributed across Salesforce, ServiceNow, Slack, Workday, and hundreds of tools
- •Future agents must aggregate data from multiple systems to create a complete picture
- •Early signs of how agents interoperate: platforms, SDKs, and emerging “protocol” concepts
- •Multiple entry points: a user could initiate in ChatGPT or inside a SaaS like Salesforce (e.g., Agentforce)
- •Box expects its agents to participate in this broader agent network rather than act alone
Startups vs incumbents in the agent era: where new companies can still win
Jack challenges whether incumbents with data and integrations will dominate. Aaron outlines three lanes for startups: incumbents that miss the shift, incumbents constrained by business-model inertia, and entirely new AI-native use cases where no incumbent exists.
- •Lane 1: “asleep at the wheel” incumbents create massive disruption opportunities (Netflix/Blockbuster analogy)
- •Lane 2: innovator’s dilemma is primarily a business-model problem (not just technology)
- •Seat-based SaaS models may resist moving to consumption/automation economics
- •Lane 3: net-new AI use cases where there isn’t a clear traditional software incumbent
- •Early to know who’s truly adapting; some large platforms are clearly leaning in (e.g., Salesforce, ServiceNow)
AI isn’t only cost-cutting: net-new spend and expanding what companies choose to do
Aaron pushes back on purely labor-replacement models and argues AI will fund new work companies previously couldn’t afford. He uses code generation tools as an example of spend that largely didn’t exist before—augmenting rather than replacing teams.
- •Economist “replace a % of labor” framing can miss new demand created by cheaper capability
- •Companies have many valuable tasks they don’t staff today due to cost constraints
- •AI can lead to incremental budgets (net-new spend), not just headcount reduction
- •Codegen tools (Cursor/Replit/etc.) illustrate new spend without immediate displacement
- •Over time there will be disruptions, but talent also migrates to higher-leverage roles
Customer support and job shifts: efficiency gains get reinvested upstream
Jack notes some functions have finite work (e.g., support tickets), making them more automatable. Aaron responds that savings often get reinvested into more proactive, higher-value human work—like customer success—changing ratios rather than simply eliminating functions.
- •Some support categories (password resets, basic troubleshooting) are prime for AI automation
- •Savings can be reallocated to proactive customer success rather than removed from the org
- •SaaS functions are governed by “ratio roles” (CSMs per customer, SDRs per rep) and are often constrained by cost
- •People can transition over time (support reps becoming CSMs) as the work mix changes
- •Acknowledges real transition risk while arguing the long-term labor allocation is more dynamic than assumed
Pricing AI agents: labor-based pricing vs software-margin convergence (and TAM expansion)
They explore why agents can be priced against human labor today and whether that persists. Aaron predicts competitive pressure will push pricing toward software-like margins—unless a company has a true “cornered resource”—while expanding total spend because AI agents aren’t capped by headcount.
- •Near-term: agents can justify high pricing because they’re benchmarked against labor cost
- •Long-term bet: competition drives pricing down toward compute + software margin economics
- •Exception: durable labor-comp pricing requires proprietary “cornered resources” (unique data/access)
- •Big upside: seat-based SaaS is capped by employee count; AI allows many agents per employee
- •A 20-person company could effectively “hire” dozens of AI roles, dramatically expanding software TAM
Is AI overhyped? Valuations, outcomes, and why the framing can be misleading
Jack asks if this is a near-term bubble like 1999. Aaron argues the “overvalued short-term but undervalued long-term” framing is awkward; instead, the reality is dispersion—many companies will fail, and the winners will make today’s prices look cheap.
- •Value should theoretically incorporate long-term expectations, complicating the usual bubble framing
- •Hype cycles increase the number of mispriced assets and failed companies
- •Winners will validate high valuations; losers will make them look absurd
- •Often takes multiple attempts in a category before the true breakout becomes clear
- •Practical takeaway: broad excitement is inevitable; focus on picking/being the outlier that executes
Lessons from being early to cloud: architectural stubbornness and moving faster
Aaron recounts launching just before AWS and the path dependency of building infrastructure in-house. He highlights key early choices—like refusing on-prem deployments—that were painful short-term but crucial to becoming a scalable, multi-tenant platform that can roll out AI instantly to all customers.
- •Box launched months before AWS, forcing early infrastructure competency and later migration effort
- •Regret theme: not moving faster on initiatives that ultimately proved correct
- •Critical decision: staying SaaS/multi-tenant and refusing on-prem despite customer pressure
- •Multi-tenant architecture now enables instant AI feature rollout across the entire customer base
- •Disciplined platform integration (including acquisitions) avoids fragmented architectures
Neutrality as strategy: avoiding lock-in to one cloud or one AI model
Aaron explains why Box’s cross-cloud, model-agnostic stance has become a competitive advantage—especially as AI model quality shifts rapidly. Neutrality means customers can adopt the best model (Gemini, OpenAI, Anthropic, etc.) without relocating content or rebuilding workflows.
- •Customers benefit from flexibility as the “best model” changes over time
- •Avoids vertical-stack lock-in where a cloud provider dictates model access and pace of innovation
- •Neutral-platform positioning parallels other cross-stack players (e.g., Databricks analogy)
- •Strategic maxim: fully exploit your positional advantage; partial neutrality is fragile
- •Operational implication: be first to support new models/tech if you aren’t training your own
How to stay motivated for 20 years: platform variety, building energy, and AI-fueled momentum
Jack asks how Aaron sustains founder energy through decades and public markets. Aaron attributes it to enjoying building, having a platform with broad use cases, and a renewed surge of excitement from AI—plus the advantage of not starting from zero again.
- •Motivation comes from enjoying creation, problem-solving, and new technology cycles
- •Box’s breadth of use cases (NASA, studios, drug research) keeps the work varied and stimulating
- •AI has sharply increased the pace of “wow” moments through frequent internal demos
- •Established platform and customer base reduces existential stress versus starting from scratch
- •Acknowledges startup “chewing glass” reality and the need for sustained conviction and energy
Tech and politics: party shifts, policy tradeoffs, and building-oriented governance
In a closing political segment, Aaron discusses perceived shifts in U.S. party alignment and why some tech figures moved right. He argues Democrats have policy failures—especially around building and affordability in California—and shares cautious optimism about pro-innovation, pro-tech signals from the new administration while opposing tariffs.
- •Perception: Democrats moved left on some dimensions; individuals prioritize issues differently
- •California as case study: unmatched advantages undermined by bureaucracy and housing unaffordability
- •Democrats can’t “message around” bad policy; need a reset focused on building and execution
- •Post-election debate: centrist vs progressive interpretations and coalition-building challenges
- •Optimism on pro-tech/innovation posture in the administration; concern about tariffs; support for deregulation where it enables building
