Aaron Levie on why AI is 'total upside' for SaaS and Box's agent roadmap

Summary

Aaron Levie argues that AI agents represent total upside for enterprise SaaS rather than disruption. Agents need systems of record to function. They require infrastructure that defines what data they can access, how workflows are structured, and how results surface to users. That infrastructure doesn't disappear—it becomes more valuable.

Levie runs a small agent team at Box, roughly 20 engineers across three rows, pushing what's possible when agents access enterprise content. Two years ago, he says, the architectural approaches they're now deploying "wouldn't have been in the sphere of possibility." The excitement comes from long-running agents that can use any number of tools, work with any amount of data, and no longer hit the context limits that constrained earlier work. Claude Cowork, Codex, and Claude Code serve as proof points.

On the broader SaaS market, Levie accepts Ben Thompson's framing that model makers will be arms dealers but pushes back on the zero-sum read. The pie expands because agents unlock spend that software vendors couldn't tap before. A customer storing contracts in Box gains the ability to run agents through those contracts and open new revenue streams. That's TAM expansion, not market share theft. He sees this dynamic within his customer base already.

Levie acknowledges the hostage critique but argues it's overblown for high-stakes systems. Ford's supply chain ERP processes billions of transactions. You cannot vibe code that. Enterprises have finite IT resources and must choose where to spend them. Rebuilding commodity software that the market has seen thousands of times makes no sense. Building custom experiences that drive revenue does. On the margin, he expects companies to keep buying from incumbents for core workflows while deploying agents on top.

He lands in an unusual spot: "100% bullish on vibe coding" and that we'll have 100 times more software, but still not willing to build Box's own CRM. The ROI doesn't clear relative to other priorities.

For software buying, Box will stay locked into a core set of vendors, deploy agents on those platforms while biasing toward competitive agent offerings from vendors themselves, and build agents in-house for large knowledge work use cases. Agents need data, workflows, access controls, and user-facing interfaces. All of these are software problems. Incumbents that stay competitive and engaged will retain their positions. Gaps will emerge where incumbents are asleep, creating new startup opportunities.

Levie describes model selection as a continuum from Gemini Flash on one end to Opus on the other, with capabilities moving up over time. His calculus is not about lowering cost but sustaining current costs while adding incremental value. Something that ran on Gemini 2.5 Flash a year ago might now run on Gemini 3 Flash, delivering two or three percentage points of improvement in tasks like transcription or data extraction at the same price. He expects bifurcation over the next couple of years. Solved problems get cheaper. Frontier work, making a $100 per hour knowledge worker two to three times more productive, commands premium models. Pharma researchers will use best-in-class models. Back-end transaction processing will use cheaper models. That split sustains for the next decade as models keep improving.

Information-heavy businesses are most excited about agents, Levie says. Those with 10 million documents see obvious ROI in AI agents that unlock employee productivity. Professional services, consulting, financial services, and law firms are prime examples, businesses where teams repeatedly redo past work and see concrete value in agents that can mine institutional knowledge. CEOs call with specific use cases: "I have 10,000 contracts. I want to find which ones let me sell a new capability." Those companies are extremely excited.

Storage is having a moment. Levie notes Western Digital and Seagate were forgotten, now suddenly cool. Box is in the storage business, long treated as commodity. In the AI era, data becomes the most strategic asset. Getting the right data to agents is the highest-leverage move in the stack. Box has locked-in public cloud infrastructure contracts, so near-term pricing spikes don't concern him. More broadly, he wants abundance through solar data centers and space data centers because cheaper AI means more usage and more value for his customers.