Box opened its annual developer conference, BoxWorks, on Thursday with one of its most ambitious sets of product announcements yet. At the center was Box Automate, a new framework that allows companies to break workflows into segments and deploy AI agents at scale across their organizations.
The launch reflects the company’s rapid acceleration into AI. Box rolled out its AI Studio in 2024, added data-extraction agents earlier this year, and followed with search and deep-research agents in May. With Box Automate, CEO Aaron Levie says the company is moving beyond individual features to build an “operating system for AI agents.”
Tapping Into Unstructured Data
Levie argues that the greatest impact of AI won’t be in structured systems like CRMs or ERPs, which have already been heavily automated, but in the messy world of unstructured content: contracts, marketing assets, due diligence documents, and creative files.
“For the first time ever, we can actually tap into all of this unstructured data,” Levie said. By embedding agents into workflows, Box aims to help enterprises handle tasks like legal reviews, M&A deal assessments, and content approvals with far greater speed and consistency.
Guardrails for Business AI
Still, Levie is clear-eyed about AI’s limits. Customers, he notes, are concerned about workflows “going off the rails” if agents make compounding mistakes. To prevent that, Box Automate lets companies split tasks into stages for instance, a submission agent, a review agent, and an approval agent, ensuring that no single system runs unchecked.
This staged approach also addresses a key technical issue: context window limits. Even advanced models can lose coherence when handling very long or complex tasks. Breaking processes into sub-agents allows companies to keep context sharp and results reliable.
“We’re in the era of context within AI,” Levie explained. “What AI models and agents need is context, and that context is sitting inside your unstructured data.”
Balancing Power and Control
The debate between massive frontier models and smaller, specialized ones continues to divide the industry. Levie takes a pragmatic approach: Box doesn’t pick sides. Instead, its architecture is designed to be future-proof and model-agnostic. As models improve, Box customers inherit the benefits without being locked into one vendor.
Critical to this flexibility is Box’s longstanding emphasis on data governance and permissions. Unlike generic AI tools, Box Automate ensures that agents only access data authorized for each user, preventing accidental exposure of sensitive information. “When an agent answers a question, you know deterministically it can’t draw on data that person shouldn’t see,” Levie stressed.
Competing with the Giants
With companies like Anthropic, OpenAI, and Google pushing their own agentic features, Box faces stiff competition. Yet Levie believes the enterprise layer gives Box a defensible edge. “What enterprises need when they deploy AI at scale is security, permissions, control, APIs, and choice of models,” he said. “That’s exactly what we’ve built.”
By combining content storage, vector embedding, workflow automation, and neutral access to multiple AI models, Box is positioning itself as the connective tissue between foundation models and enterprise-grade deployment.
The Bigger Picture
For Levie, the Box Automate launch marks a turning point. The company is betting that the era of context, where the effectiveness of AI depends on how well it is embedded in enterprise content and workflows, will define the next decade of business technology.
In a landscape crowded with powerful models, Box is wagering that trust, governance, and workflow integration will matter more than sheer scale. As Levie put it, “There’s no free lunch in AI. But with the right architecture, we can finally bring automation to the work that has always resisted it.”
