In our latest episode of Deployed, we sat down with Ben Kus, CTO of Box, to discuss how they're transforming enterprise content management with AI. As a public company serving over 100K+ customers, Box brings a valuable perspective on what it takes to build and deploy AI at scale in enterprise environments.

A few key themes emerged from our conversation that should be particularly relevant for teams building enterprise AI products:

• How generative AI is transforming the way companies work with unstructured data (with Box as the hub for much of that data — petabytes of it!)

• What enterprise AI agents are likely to look like in 2025, at least at Box

• The importance of "System 2 thinking" in agent design

• Practical lessons from rolling out AI capabilities within a well-established enterprise product

Dig into some highlights below, or check out the full episode on Spotify, Apple Podcasts, or YouTube.

Generative AI: The Perfect Tool for Unstructured Data

One of the most compelling parts of our conversation was Ben's perspective on how generative AI is unlocking new possibilities for working with unstructured data. 

"When Generative AI came out, it was kind of like a gift," Ben explains. "We'd been watching the space develop, and then it was this kind of magical moment when all of a sudden it got really good... Not only does this thing chat well with you, it can actually read or look through your content and then almost act like a personal assistant who just read the thing for you."

This capability has transformed what's possible with enterprise content management. While structured data has seen decades of innovation in how companies can analyze and work with it, unstructured data has typically required human review to extract insights. Generative AI has changed this completely in the last two years.

Box customers can now:

• Ask questions across their entire document corpus

• Extract structured metadata from unstructured documents (cool demo video here)

• Get intelligent summaries and insights from content

• Process complex documents like contracts more efficiently

Bringing AI Agents to the Enterprise

Perhaps most interesting is Box's bet on AI agents for 2025. A year ago, most enterprises were wary of the non-deterministic decision making in agentic flows. Ben notes how quickly enterprise attitudes have shifted:

"It used to be like a lot of people were like, 'well, I'm just trying to get my first use case.' You're telling me about some crazy new thing. But now... I think that the thing that is really powerful about the concept of AI agents is that they're able to do more complex tasks in a more specific way than was possible up until now."

Box's vision for agents goes beyond simple chatbots. They're building what Ben calls "agentic workflows" – AI systems that can handle complex, multi-step tasks while maintaining reliability and trust. This includes capabilities like:

• Self-checking and correction of responses

• Intelligent citation of sources

• Breaking complex tasks into manageable steps

• Maintaining context across multiple interactions

Moving Beyond Chatbots to Agent UX

A key insight from Ben is how agent design requires rethinking traditional AI interfaces, and especially chat. 

"The user interface is different... the user expectation is different... what's possible is actually weirdly different," he explains. This is particularly true when it comes to giving agents time to think.

He illustrates this with a story about a customer doing loan processing. Initially, customers were excited just to have AI help process individual documents faster. But the real potential of agents goes far beyond document-by-document automation. 

"Why even bother to make the person go open up each of those themselves when you could actually have the agent go through the process and give the loan processor a full summary of all of this in some structured way?"

This evolution in thinking is critical for builders – moving from chat interfaces to autonomous workflows that can handle complex multi-step processes and just get things done for people. It requires new ways of thinking about user interfaces that go beyond conversation, focusing instead on how to help users monitor and guide autonomous processes that can run independently.

Rolling Out AI in an Enterprise Environment

One of the most valuable parts of our conversation was Ben's practical advice on implementing AI in an established enterprise product company. Our conversation covers several aspects including how they’ve set up teams and how they’ve decided which AI projects to work on. We’ll call out some highlights here especially for Freeplay customers on tools & testing, but watch the whole thing to hear all Ben’s advice:

Strong Platform Foundations

"The thing that I would definitely do again is solidifying how AI works in your product," Ben advises. Box built what they call their AI Studio, which provides:

• Consistent ways to access and process files

• The ability to switch between different LLM providers

• Tools for customizing and testing prompts

• Enterprise-grade security and compliance controls

Quality Control at Scale on Sensitive Data

Box’s has figured out effective ways to maintain quality despite enterprise data privacy constraints, especially by focusing on LLMs as judges:

• Using LLMs as graders to evaluate responses and produce scores allows them to evaluate quality without a human ever looking at customer data

• They’ve also built representative test datasets from their own data that employees can look at for QA

Huge thanks to Ben and the Box team for sharing these insights with a wide audience. For more information about Box's enterprise AI capabilities, visit box.com/ai.

Want to hear more conversations like this one? Subscribe to Deployed on Spotify, Apple Podcasts, or YouTube.