Are SaaS Companies Cooked: Which Thrive & Which Die | Aaron Levie
Aaron Levie, CEO of Box, argues that enterprise AI adoption will be far slower and more labor-intensive than Silicon Valley assumes, creating massive demand for a new class of workers—"agent operators"—who redesign workflows around AI rather than replacing human expertise. Rather than decimating professional jobs, AI will force organizations to tackle bottlenecks they didn't know existed, ultimately creating more lawyers, engineers, and specialized roles than today. Builders should focus on becoming the best platform for agents to work with data, not compete on the AI model itself.
Key takeaways
- • AI won't remove humans from workflows—it will shift where they enter the process, creating a need for "agent operators" (a new job title) who understand MCPs, CLIs, and workflow redesign to make agents effective in regulated enterprise environments.
- • The biggest constraint in enterprise AI isn't compute or models; it's data fragmentation and legacy systems—most Fortune 500 companies have contracts and documents scattered across 10+ incompatible systems, making agents fail at finding the right information without significant change management work.
- • Budget allocation for AI tokens must move from IT spend into operational OPEX budgets (like marketing or sales budgets), not compete against Salesforce licenses—this unlocks new revenue pools for software vendors and doubles total enterprise tech spend from ~10% to ~20% of OPEX.
- • Token budgeting strategies for large enterprises should stratify users: give unlimited access to top 5-10% performing teams on best models, limit the middle 20%, and give cheaper models to general productivity users to stay within budget constraints.
- • Headless-first software (APIs over UI) is now non-negotiable—agent accuracy at tool-calling and document retrieval has improved dramatically in the past year, making traditional user interfaces irrelevant for agent-driven workflows.
- • Professional services firms like Accenture will dominate the next decade because enterprise AI deployment requires 10+ years of change management—data organization, system integration, workflow design, and compliance setup can't be automated away, only done by specialized implementation teams.
- • Agent observability and evaluations (monitoring whether agents degrade over time) is emerging as critical infrastructure that no single AI lab will own—enterprises need evals to work across OpenAI, Anthropic, and open-source models.
Recommendations (3)
"I'll give a shout out to BrainTrust as an example not an investor where agent builders were going to need eval"
Aaron Levie · ▶ 52:28
Mentioned (6)
More from these creators
Cursor Acquired for $60BN | Anthropic Hits $1TRN in Secondary Markets & Figma, Adobe, Canva Dead?
Jake Paul: Traditional VC is Toast & Attention is More Valuable than Cash
SpaceX's Financials Leaked: Is it Worth $2TN | Meta Debuts Muse Spark: Are They Back in the AI Race?
The Early Days of Anthropic & How 21 of 22 VCs Rejected It | The Four Bottlenecks in AI | Anj Midha
ElevenLabs: Building an AI Sales Machine & Why We Set a 20x Sales Quota
OpenAI Buys TBPN & Their Management Team Reboot | Mercor Hack & Why Now is the Time for Cyber