Aaron Levie Is Betting on Agents, Not Models

Aaron Levie has long been known as one of tech’s most outspoken and insightful CEOs. As the co-founder and CEO of Box, a content management and collaboration platform powering over 100,000 organizations, Levie is now at the center of one of enterprise software’s most pressing questions on how do you build a real AI business without losing focus or going to war over models?

Speaking on the Mad Podcast, Levie opened up about Box’s AI strategy, what makes agent-based architecture valuable, and why he believes the majority of enterprise data still sits underutilized. “I would argue that probably for 95% of our data, people have things they would love to be able to do with it—but they just don’t,” he said. Agents, he argued, could finally shift that equation by making once-unrealistic workflows feasible. “The thing I get excited about with agents is the expansion of what people can now do with software.”

Unlike some of its peers, Box has deliberately chosen not to build its own foundation model. “That is just a war you want to ride the tailwinds of, not play in,” Levie said. “We debated [fine-tuning]—for like 10 minutes. It’s not worth it. Things are changing too fast.” Instead, Box supports models from OpenAI, Anthropic, and Google’s Gemini, with plans to support others like Meta and xAI. “Give us the API key, and we’ll plug it in,” Levie said. The goal is to remain flexible, focus on orchestration, and deliver utility instead of chasing breakthroughs.

What Box Has Actually Built

Box’s AI efforts are anchored in practical deployments layered onto decades of enterprise content infrastructure. The company now offers document conversion, text extraction, embeddings, vector databases, and a model abstraction layer. All of these are exposed through APIs, a product called Box AI Studio for building agents, and Hubs—RAG-based chat interfaces centered on specific business domains like HR, legal, or finance.

“We constrain the problem,” Levie explained. “The user contributing the data to the hub is only going to contribute the most relevant and authoritative content. The user coming in is psychologically focused on a topic. We’ve eliminated 95% of the problems you get, which is data sprawl and unconstrained queries.”

The approach seems to be gaining traction. As of early 2024, Box began rolling out AI capabilities to customers on its Business, Business Plus, and Enterprise plans. New features include Box AI Extract Agents that pull structured data from unstructured content—including images—without requiring manual tagging. These agents are configurable via API, and the extracted information is saved into metadata for downstream automation.

To help customers track usage, Box introduced Box AI Units—a consumption-based model that quantifies each AI interaction (such as querying or metadata extraction) into discrete units. “Traditional pricing models often lack transparency,” VP of Product Management Yashodha Bhavnani wrote in a February blog post. “Box Units aims to provide more clarity.”

IBM Enters the Chat

In April, Box announced a partnership with IBM to integrate its AI features with IBM’s watsonx platform. The collaboration allows customers to apply IBM’s Granite and Meta Llama models to Box-managed content, supporting workflows like data extraction and contract analysis. IBM is using the Box AI and watsonx integration internally, and Box is using IBM watsonx.governance to monitor and manage its own AI lifecycle.

“The impact AI is having on the enterprise is unlike anything we’ve seen since the birth of the internet,” Levie said in the announcement. “By deepening our partnership with IBM and leveraging watsonx, we’re empowering enterprises to harness AI responsibly and at scale.”

Internally, Levie has pushed for a company-wide embrace of AI. “We’ve told everybody in the company that we want to use AI to be as productive as possible and aggressively use it across the business.” Weekly all-hands meetings now include demos from employees showcasing how they’re using AI in their work. “Everyone is involved—sales, support, engineering. It’s now fully integrated. Took a year or so to get there.”

The development team is also hands-on with tools like GitHub Copilot and Cursor, and Levie is bullish on the idea of agents handling routine engineering work like SDK updates or language rewrites.

Levie doesn’t shy away from the current limitations. He called out compound errors in chain-of-agent setups, poor search recall, and the risk of surfacing confidential information due to lax data governance. “The moment that the AI finds the wrong thing, you’re path dependent now on the wrong thing,” he said.

That’s why Box is focusing on use cases where tasks are constrained and auditable. “If you ask it to pull out data from a short document, it’s going to be 100% accurate. If you ask it to pull out data from a 200-page document and you only give it one shot, it’s going to get 75% accuracy.”

Levie’s view on AI monetization is tempered. “We don’t expect a magical AI premium on pricing. But we can grow by solving more problems and entering new markets.” For Box, the focus is expanding use cases, not raising license fees. “It’s about real utility.”

And despite the hype, Box is staying close to what made the company successful in the first place—content, compliance, and enterprise readiness. “We want to be the best place for managing content,” Levie said. Everything else—agents, models, apps—layers on top.