a16z
Box CEO Aaron Levie: Why Big Companies Keep Blowing Their AI Bets (a16z Podcast Recap)
Aaron Levie and a16z partners dissect why most enterprise AI projects fail, exposing the real obstacles: legacy systems, organizational inertia, and the messy reality of integrating AI agents into sprawling companies.
If you only read one thing
Aaron Levie doesn’t mince words: MIT found that 95% of big-company AI projects fail, yet boards keep demanding more. The real culprit isn’t lack of ambition—it’s the tangled, outdated machinery inside large enterprises.
Aaron Levie argues that most large companies are set up to fail at AI because their structures, legacy systems, and decision-making processes are fundamentally misaligned with what AI needs to succeed. Boards demand 'more AI,' so CEOs hire consultants and launch centralized projects that ignore how work actually gets done. These efforts almost always flop, leaving companies even more risk-averse and paralyzed. Martin Casado reframes the challenge: instead of embedding AI into products, treat AI as a user—an agent that interacts with software like a human. But this exposes new headaches: agents hit the same permission walls as employees, can’t ask for access, and lack the context or relationships to navigate undocumented processes.
Meanwhile, some companies try to boost AI adoption by tracking token usage, which just creates fake productivity and junk output. Levie and Casado agree: for AI agents to work, companies need to modernize their tech, rethink access, and treat agents as digital employees—with onboarding, culture training, and their own licenses. And if every employee gets an agent, most SaaS systems will buckle under the load—they were never built for this kind of automated, parallel usage. The future isn’t just smarter software; it’s a total overhaul of how big companies operate.
Why it lands
The real risk in enterprise AI isn’t technical—it’s organizational. Legacy systems, outdated workflows, and centralized decision-making doom most AI projects before they start.
The agentic model demands a total rethink: companies must modernize, treat agents as digital employees, and prepare for a world where software is used by machines, not just people. Winners will be those who rebuild their foundations, not just talk about AI.
Why Big Companies Keep Failing at AI
Boards demand more AI, but CEOs respond with consultant-driven, top-down projects that don’t fit how the company actually works. The result: a cycle of failed initiatives and growing skepticism.
- 95% of enterprise AI projects fail, according to MIT.
- Centralized, consultant-led AI projects rarely fit operational realities.
- Failure breeds organizational paralysis and reluctance to try again.
The Agentic Model: Treating AI as a User, Not Just Software
Martin Casado says the real shift is to treat AI as a user—an agent that interacts with software like a human, not just a backend feature. This exposes new integration and access headaches.
- AI agents hit the same permission walls as humans in legacy systems.
- Agents need their own identities, licenses, and onboarding processes.
- Fake productivity emerges when companies incentivize AI usage by token count.
Legacy Systems and the Coming SaaS Crunch
Enterprises with sprawling, decade-old systems can’t just bolt on AI. Agents operating at scale could overwhelm SaaS products never designed for automated, parallel usage.
- Legacy integration is a massive unsolved problem—AI doesn’t fix it.
- If every employee gets an agent, SaaS systems could collapse under the load.
- Modernization and new access models are prerequisites for real AI automation.
Worth stealing
- Enterprise AI failures are mostly organizational, not technical.
- Treating AI as an agent/user exposes new integration and access control issues.
- Legacy systems and centralized decision-making are the biggest blockers to AI adoption.
- Fake productivity is a real risk when AI usage is incentivized by metrics like token count.
- SaaS products must be re-architected for agent-driven, high-volume usage.
Lines worth repeating
We need more AI. And what does the CEO said? Oh, okay. I'll get like a consultant to do more AI.
Aaron Levie
MIT had this stat like 95% of AI efforts in big companies fail... AI failures have created some amount of bruising.
Aaron Levie
Instead of viewing AI as software, just view it as a user... take your product make it a CLI tool and then have the AI be an agent that actually uses it.
Martin Casado
Any enterprise of a thousand people or more or that's older than 10 years is just a mass of stuff that’s sitting there waiting to be integrated and AI actually doesn’t help to integrate anything.
Aaron Levie