Box's Aaron Levie: 90% of My AI Work Wasn't Possible Two Years Ago

Aaron Levie estimates that 90% of what he does with AI today was not possible two years ago, and he thinks most knowledge workers have no idea where on that curve their own jobs sit. The Box cofounder and CEO laid out his reasoning in a conversation with Fast Company's global technology editor Harry McCracken, published in the outlet's weekly Plugged In newsletter — and the through-line is simple: model capability has outrun real-world adoption, and a whole layer of companies will be built to close that gap.

What did Levie say about how Box actually uses AI?

The hardest internal number in the interview concerns code. "The significant majority of all code that we write has an agent involved in it," Levie said. He doesn't track exact line counts, but assumes agents now write the majority of lines produced at the company he cofounded in 2005.

His verdict is emphatic. "I'm 1,000% a believer in agent coding. It's allowing us to ship way faster. We're building more features than ever before. We're solving customer problems way faster."

Beyond engineering, Box uses AI on its own content repository. Because nearly all company data lives in Box, the system functions as an automatic knowledge base covering project status, onboarding, sales training, product architecture, and decisions made years earlier that still matter today. Employees supplement this with ChatGPT and Claude for other workflows.

How does Levie handle AI costs?

Not with hard caps. Box has avoided both "tokenmaxxing" culture and mandatory justification of AI spend. The logic: engineering runs disciplined product road maps, so if the roadmap is agreed, the rational move is to spend what you can afford to ship the product as fast as possible — safely and robustly.

Does AI mean fewer jobs at Box?

Levie says the opposite, and claims the company is acting on it. He's inclined to hire more people in most areas, on the theory that higher productivity plus no shortage of ambitions equals more output, competitiveness, and revenue. "Other than a couple of areas, we are firmly on the side of AI as a job creator, and we're putting our money where our mouth is," he said. Box is still hiring across most job functions and expects that to continue.

The pattern, he argues, generalizes. Customers aren't cutting 20% of staff after a 20% productivity gain; they're redeploying the capacity into market competitiveness. A life sciences firm wants to accelerate drug discovery and clinical trials. A bank wants faster client onboarding. Saved dollars usually get spent on something that also requires labor.

What is the 'capability overhang'?

Levie's central analytical claim is that model capabilities are "quite advanced relative to what we've ultimately been able to use them for." The bottleneck isn't intelligence — it's connection. A drug discovery team might have access to the world's most capable model, yet that model still can't reach proprietary research, clinical trial workflows, or scattered data systems.

His forecast: years of diffusion through an "applied AI layer" of products bridging model capability and business process, vertical by vertical. He names Harvey and Legora in legal as evidence, and predicts the same for financial services, life sciences, and manufacturing.

Why won't frontier labs absorb that layer?

Levie answers with a historical comparison. Look back 15 years at the full feature set of AWS and Google Cloud, and you'd never predict Snowflake at $100 billion, Databricks at $175 billion, ServiceNow at $100 billion, or Palo Alto Networks at $250 billion. The giant clouds covered roughly 80% of use cases; specialized companies went deep on the remaining 20%.

The relationship isn't zero-sum. For every dollar Snowflake or Databricks earns, he estimates, the underlying cloud providers capture roughly 80 cents — and the same economics will hold for AI, with frontier labs profiting from workloads the applied layer drives through their models.

How fast is adoption really moving?

Unevenly, in ways Levie finds unpredictable. Coding adoption has gone "fully vertical" while other knowledge work diffuses more slowly. Engineering converges fastest because roughly 5% of any engineering team stays hyper-online, pulls down new best practices, and spreads them internally. Big banks now run agentic practices with Cognition, Factory, Cursor, and Claude Code that resemble leading Silicon Valley setups.

Startups adopt faster than enterprises by a meaningful margin, since the person next to you can implement a tool — or you're technical yourself. Large enterprises drag under compliance, regulatory, and security constraints.

What should individual workers do?

Levie's advice is blunt: start using the tools, expect wildly varying returns — 10% to 20% productivity gains in some functions, 200% to 400% in engineering — and keep retesting. He checks back on dismissed use cases every six months, because a single model breakthrough can flip a verdict. Levie also estimates about two-thirds of his own AI commentary targets Silicon Valley rather than businesses, translating field observations into what builders need to fix.

The implication for the next few years is his closing wager: the applied layer, not the labs, is where the next Snowflake-scale companies get built.