Aaron Levie: enterprise software is accelerating, not dying — agents are creating more demand for platforms like Box

Aaron Levie: agents need platforms, and Box is one of them

Aaron Levie pushes back on the idea that AI is an existential threat to enterprise software. The real story at Box, he argues, is reacceleration — growth running "far past our internal plans" — driven by something counterintuitive: the more agents get deployed inside enterprises, the more valuable the underlying platforms become.

Why agents strengthen platforms, not replace them

The bear case on enterprise software was that agents would simply route around legacy systems and talk directly to databases. Levie's rebuttal is structural. Imagine 10,000 employees and tens of thousands of agents all hitting the same data environment simultaneously. Suddenly you need access controls, canonical records, authoritative document versions, security guardrails, and a traffic cop for agent activity. That is what Box does. "Agents need to be able to work with that unstructured data to make decisions or move information through a workflow," he says — contracts, research materials, marketing assets, financial documents. None of that simplifies when agents multiply. It gets harder to govern.

He points to Cloudflare, Databricks, and Snowflake as examples of the same dynamic playing out across infrastructure and data. Agents need sandboxes, compute, network gateways, and large-scale data retrieval systems. The businesses supplying those layers are, in his words, "totally on fire."

“We have reaccelerated growth far past our internal plans because it turns out that enterprises need core systems to be able to manage their unstructured data... I would say, right now, my view would be it is unequivocally been a net positive for software, even with some of the spaces being under pressure, simply because how much more work agents are doing on these systems.”

What Box is building for an agentic world

Box has upgraded its retrieval layer significantly. Compared to search two or three years ago, the system now captures document embeddings, runs re-ranking on results, and parallelizes queries — running multiple versions of a search simultaneously to improve both speed and accuracy. An external agent requesting a specific record gets a much more reliable result than it would have under the old stack.

Security is the sharper near-term priority. Levie flags the OpenAI-Hugging Face incident as a signal of what happens when agent swarms interact with sensitive data environments. Box is rethinking its security architecture from the ground up, asking what data protection looks like when the threat model is "a 100 times more agents than people."

Personal agents: genuine excitement, honest skepticism

On consumer agents, Levie is notably more bullish than he expected to be. He admits he "did not have the imagination on the consumer front" and credits products like Muse and Instinct with packaging the technology in a way that works. His consumer thesis isn't productivity — it's friction removal and consumption expansion. He uses the example of his own family: agents let him add activities (Japanese tea gardens, camps, local services) that he simply wouldn't have organized manually. The unit economics of the local economy could expand materially if agents make low-friction discovery and booking the default.

The adoption risk he acknowledges is the chasm problem. Early adopters using agents to handle complex tasks is one thing; mass consumer behavior change is another. His conditional optimism is that if the technology executes well enough and the interface stays as familiar as a mobile app — no VR headsets, no CLI — the behavioral ask is low enough that it might actually cross over.

AI governance: the narrow line

Levie's p(doom) is low, partly because serious safety research is happening and the ecosystem has strong incentives to attack alignment problems. The risk he worries about more is political. Safety language gets co-opted by parties with different goals, and there is a narrow line between "elevating risk appropriately" and triggering a voter backlash that simply halts data center construction. He expects AI to be the dominant issue in the 2028 presidential election — jobs, GDP, safety — and thinks the tech industry itself doesn't yet know where it wants things to land.

The tension he names is real: the same week someone signs an open-source pledge, someone else signs a pacing letter. Both can be rationalized, but they pull in opposite directions, and there is no clear industry consensus on how to thread them.