The Twenty Minute VCAre SaaS Companies Cooked: Which Thrive & Which Die | Aaron Levie
CHAPTERS
AI “race” framing: commercial competition vs existential panic
Aaron reacts to the Jensen/Dwarkesh debate and argues the US–China AI narrative is often overstated. He frames AI as primarily a commercial and economic race (with safety considerations), not a binary, winner-take-all sprint where a few months decides everything.
- •A month or two of advantage is unlikely to determine the long-run outcome of AI progress
- •National leverage comes from whose technology stack becomes globally adopted
- •Security and safety remain critical, but the debate is too often treated as binary
- •“Upgrading systems” is slow; defensive leaps don’t happen instantly even with early access to breakthroughs
Why “AI will take your job” is the wrong mental model
Aaron argues AI is shifting where humans enter workflows rather than removing humans entirely. He believes fear-driven narratives discourage people from entering fields society still needs, and he expects headcount growth in many professions even as tasks change.
- •Humans aren’t removed; they re-enter the loop at different points (review/approval/oversight)
- •Discouraging careers like engineering or radiology could harm society’s capacity
- •Automation reveals new bottlenecks rather than eliminating all work
- •The key question becomes redesigning workflows around human+AI collaboration
More engineers (and even more lawyers): demand expands outside Big Tech
He predicts more engineers at Box and across the economy because most industries lack automation capacity. As non-tech sectors gain “Silicon Valley-grade” engineering leverage via AI coding tools, demand for technical talent broadens; similarly, legal content generation increases the need for qualified review and throughput.
- •Most of GDP sits outside tech; those sectors are under-engineered and will hire more
- •AI coding tools let banks, pharma, manufacturing, and agriculture build far more software
- •Entry-level roles may shift, creating apprenticeship/mentorship challenges
- •Legal content becomes easier to generate, but approvals, courts, filings, and accountability remain constraints
The emerging role: “agent operator” and enterprise workflow redesign
Aaron outlines a new job category centered on implementing and maintaining agent-driven workflows inside regulated, complex enterprises. This role blends technical fluency with business process redesign and change management, acknowledging that enterprises can’t simply drop in agents without re-architecting processes and data.
- •A new class of “agent operators” will understand MCPs/CLIs, skills, agents.md, and enterprise workflows
- •Enterprises differ from startups: regulation, fragmented data, and entrenched processes drive complexity
- •Workflows must be redesigned for agents (not merely layered on top of human-centric processes)
- •Models change quickly; workflows can break when prompting, indexing, or tool behaviors shift
Are SaaS products just databases in an agentic world? Where value moves
They explore whether agents will reduce SaaS to “valueless databases.” Aaron agrees some button-heavy software may commoditize, but argues durable value persists in APIs plus embedded business logic, governance, and human-in-the-loop collaboration experiences.
- •Value may shift from UI/buttons toward robust, agent-ready APIs
- •“All software has a database,” but differentiation lives in business logic above it (permissions, policy, workflows)
- •ERP/CRM-style systems aren’t reduced to raw storage because of deep operational logic
- •Human+agent collaboration still needs interfaces for review, exceptions, and governance
Agents as massive consumers/creators of unstructured data—and why platforms still matter
Aaron expects an explosion of unstructured content: contracts, marketing assets, reports, and analyses generated and reviewed by agents. He argues this increases the need for a governed backbone to store, secure, and manage that data—positioning systems like Box as force multipliers rather than commoditized repositories.
- •Unstructured data volume will surge as creation/review becomes cheap
- •Enterprises need guardrails: governance, retention, access controls, and workflow coordination
- •Box’s “headless” usage (API calls) already exceeds end-user interactions, making it agent-ready
- •Monetization shifts may occur, but 100–1000x more calls/data can expand opportunity
The cybersecurity tsunami: agents create (and must solve) new risk
Aaron describes a step-change in security risk as AI generates far more code than humans can realistically review. Agents also empower attackers to scan and exploit faster, creating a dynamic environment where agentic security becomes essential.
- •AI-written code scales faster than code review capacity, raising vulnerability risk per release
- •Every new feature can introduce security issues (e.g., unintended ports/configurations)
- •Offense improves via AI-driven scanning and vulnerability discovery
- •Agents are both the cause and partial solution: “agentic security” becomes a major market
Token maxing and enterprise budgeting: from IT spend to line-of-business OpEx
Aaron explains how enterprises will allocate token/compute budgets and why ‘unlimited tokens’ is unrealistic for EPS-driven companies. He predicts token budgets shift from centralized IT budgets to business-unit operating budgets, enabling new ROI-based tradeoffs and potentially expanding total spend.
- •Allocation strategies: “compute Shark Tank” pitches, tiered model access by employee value/role
- •Budget cycles and earnings constraints prevent blanket unlimited token policies
- •Token spend moves from IT tradeoffs (licenses vs tokens) to OpEx tradeoffs (campaigns vs automation)
- •This shift can expand addressable spend beyond traditional enterprise IT caps
Enterprise adoption will be slower and services-heavy: change management and accountability
He argues diffusion will take longer than Silicon Valley assumes because enterprises face compliance, data fragmentation, and liability constraints. Professional services (Accenture-style and new specialists) will thrive by modernizing data estates and redesigning workflows for safe agent deployment.
- •Regulated workflows require human review and controls; agents can’t be fully autonomous in many domains
- •Data is scattered across legacy systems; agents can retrieve wrong documents without curation/context
- •Large-scale upgrades, integrations, and governance are long multi-year programs
- •Accountability matters: enterprises need someone to blame; you can’t “blame Anthropic” for failures
Open-source and Chinese models in the stack: pragmatic usage with ongoing human oversight
They discuss the reality that companies benchmark frontier models and deploy cheaper open-source alternatives (often Chinese) to approach frontier performance. Aaron views this as empirically true but not inherently panic-inducing, while reiterating that even best models still need oversight.
- •Many teams use frontier models to set targets and open-source models for cost/perf deployment
- •Risks exist (e.g., potential backdoors), but the bigger practical issue is model fallibility
- •Even frontier outputs require human review—reliability is not yet sufficient for hands-off use
- •Multi-model strategies reinforce the need for tooling that works across vendors
Why many public-company agent products feel weak—and why this year is brutal execution
Aaron attributes weaker agent rollouts to the speed and complexity of the agent ecosystem: staying current requires deep practitioner-level engagement. He describes the CEO challenge of navigating rapid technical shifts while building a safe bridge for customers.
- •Agent progress moves on a multi-times-per-week cadence, not traditional product cycles
- •Winning requires being wired into practitioner knowledge (not just mainstream summaries)
- •CEOs must balance innovation speed with customer safety, compliance, and adoption pathways
- •“Complete unrelenting execution” is required in software/infrastructure/agents right now
Monetizing agents in enterprise software: pricing, tiers, and Wall Street expectations
They debate whether agent products must re-accelerate revenue growth to matter publicly. Aaron says monetization is real, often via higher tiers that bundle automation/workflow value, but the market will wait to see which companies are structurally helped vs pressured by agents.
- •Agent value can justify higher-tier plans tied to workflow automation outcomes
- •Revenue re-acceleration is possible but depends on category dynamics and execution
- •Markets are currently bucketed; differentiation will emerge as agent impact varies by software type
- •Some companies must pivot; others can ride the wave if they adapt quickly
Frontier labs and the next infrastructure layer: why he’d still bet on the labs
Aaron says he would still invest heavily in frontier rounds, expecting valuations could go higher given market size. He also highlights emerging “picks-and-shovels” categories like agent observability and evals as broadly necessary across enterprises deploying agents.
- •Frontier labs may capture enormous value as markets expand (cloud-like multi-winner outcome)
- •Enterprises will run multi-vendor AI stacks to avoid single points of failure and lock-in
- •Agent observability/evals become mandatory once agents enter regulated production workflows
- •New infrastructure categories can scale beyond ‘Silicon Valley TAM’ into the whole economy
