Box CTO Ben Kus talks up the marriage of cloud storage, genAI agents

Box’s heritage as a cloud service provider has helped it succeed  across several computing eras during the last 20 years.

So when the generative AI (genAI) revolution arrived in late 2022, Box was well positioned to take advantage of it because of the amount of corporate data it already had in hand. Now, the company is topping its cloud storage with new AI tools aimed at making employees more productive.

Box is especially focused on AI agents that can provide more context to documents. These agents can solve the decades-old IT problem of extracting intelligence from unstructured data, which otherwise is underutilized in decision-making.

Box’s Chief Technology Officer Ben Kus sat down with Computerworld to talk about the company’s plans to integrate agentic AI with its cloud offerings.

A recent study showed that about 95% of AI projects are failing. How do you tell customers what works in AI and what doesn’t? “AI is very approachable. You can talk to it and get early versions working fairly quickly. In traditional software, you wouldn’t see it working until you’re almost done. Now you get immediate proof of concept, which tricks people into thinking they’re almost finished.

“In generative AI, what used to be the last 10% of work to get something working well is actually the bulk of the work. There’s a ton of work to make it really work well.

“Every customer wants to understand the latest about generative AI. I’ve had up to 10 customer conversations a day on this topic and how we see it evolving in our platform.”

What exactly are you doing with your legacy storage offerings, and what can customers expect? “Our job is providing the foundation of everything we’ve done: security, compliance, storage, collaboration — and adding AI on top.

“Customers say they have interesting ideas for using AI agentically, but they need AI to have context. The analogy is getting a really smart worker who shows up and says ‘I’m here to do whatever you want,’ but you need to give them access to important information to do real work.

“Many customers say, ‘I don’t really have an AI problem, I have a data problem.’ They need to prepare their data. Files here, images there, videos elsewhere — they have these legacy platforms that don’t support unified access.

“The challenge becomes quite complex because most enterprise data is unstructured: contracts, invoices, videos, presentations, and it’s scattered across different systems. The real value comes from bridging unstructured and structured data.

“This used to require armies of people manually reviewing documents. Now AI can watch videos, read contracts, analyze images, and categorize everything intelligently. That’s the fundamental shift, turning the vast amount of unstructured content enterprises have into actionable, information that can be queried.

“Some customers use customized agents. You train them like an employee: ‘Here are your rules.’ Instead of typing long prompts, you pick that agent and say, ‘You’re my agent responsible for reviewing marketing copy.’ You give an agent instructions, then address it as if you had an expert assistant who knows what you want.”

How do you approach deep research when others already offer similar services? “It’s very different doing deep research on the open internet versus your data. On the open internet, everything’s public. But it would be a disaster if you allowed deep research on data users don’t have access to: employee information, confidential material, financial reports, M&A deals.

“When we talk about deep research, we mean on your data, so AI can use the same techniques. It can iterate, think, make plans, and create comprehensive reports like hiring a consultant, but on data you have, not public data.

“To get real enterprise use, you need different flavors. The challenges are security, compliance, and being able to use it in your environment. You can’t drop tens of petabytes and billions of files into ChatGPT.”

How do you handle security concerns that customers have? “We apply enterprise-grade principles. Models don’t keep secrets. If you expose information to them and give them data a person isn’t supposed to have, they’ll tell them about it. It’s fundamental: if you don’t have access to something, the AI can’t have access either.

“We ensure infrastructure security, isolated inference, and never train on customer data. But the hard part is real-time double and triple-checking that they’re only doing AI on data they have access to.

“When they’re using agents, the third layer is to make sure those agents are secure and how they approach things. The more you ask agents to do, the more they resemble people in some forms.”

How do you handle multimodal content in your workflow? “The same techniques we use for text-based documents work for multimodal content. You can extract structured data from construction images, identify safety violations, review digital assets for brand guidelines, and check if images are crisp with proper foreground focus. AI can do all this.

“We handle the tricky technical details and provide customers the ability to customize on top, making their employees more productive or automating background processes.

AI is very text-based right now. Are you running into issues with rich media and multimodal content? “Many top models have multimodal capability. They handle not just text, but images, audio, and video. The technology works similarly. It converts pixels into tokens with positional information, like text. It’s mind-blowing when you see AI look at an image or video and truly understand it, not just label it.

“We have media and entertainment customers with storyboards where AI can create stories from comic strips. It looks at pictures and creates narratives. This is huge for unstructured data going forward — doing AI across modalities.

Could you talk about some upcoming features? “We focus on making more sophisticated agents that can do complex tasks and workflows. Enterprises will automate tasks that were too long or expensive for people to do.

“One example is extracting structured data from unstructured data. Customers realize they can pull data from anything and then query it, filter on it, showing all contracts or digital assets. They ask, ‘Can I automate review checks so we don’t have problems with assets that have legal concerns, brand concerns, or PR concerns?’ [We’re] having AI agents do tons of checks as you go through processes.

“This is what enterprises will do in the future. Our effort is providing agent capabilities and automation capabilities so AI can work for you in the background according to how your company works.”