How Will AI Agents Change Companies and Work? | Tetsuro Miyatake x Aaron Levie

Yeah. So obviously an incredibly exciting moment in tech right now. We're finally at a point where, because of the total capability of AI agents, you can increasingly automate more and more work because these agents can actually go and leverage the underlying APIs or interfaces of lots of software that people want to be able to work with. And if you think about what are people doing at their job often, what's consuming a lot of hours in our daily work, it's navigating websites, it's doing research, moving information from one place to another, it's obviously using judgment, but executing tasks inside of these different systems. So if AI agents can now effectively use those tools and they can do it obviously in a way that is in parallel, that means that we can now farm out work to these AI agents that can go and execute lots of tasks for us, right? That's incredibly powerful because now we can be dramatically accelerating the work output that we have in our organizations.

And this could range from everything from you're a sales rep and you want to go and do research on a particular set of customers that you're going after, well, agents can go and fan out and do a bunch of that research for you. Or if you wanted to more quickly onboard a client to your company, agents can go review all of that documentation and process that information. If you wanted to be able to have agents go and read every customer call to create a transcript and then analyze it and look and send it, agents can now go do that. So that's the now power of what AI agents are able to do on our data and inside of our systems. And so, but it's all driven by the fact that the capability level of AI has gotten to a point where they can use APIs, they can use tools, and they can execute these tasks very effectively. And now we're in a world where, you know, for internet traffic, AI agents and bots have now surpassed human traffic, and as you mentioned in many podcasts before, we're now at a point where AI agents are going to start to be far more populous than humans using apps.

Yeah. Well, this is actually an environment that we were more or less effectively built for because, in the very early days of building Box, like within the first three months, we actually had a public API. And what we realized was while you might want to store and access and share your files from the interface of Box, you might need access to that data from a bunch of other places. Those could be a mobile app, it could be another website. So we were API-first basically on nearly day one of the company. And so what do agents need to interact with? They need a set of APIs and they need a programmatic interface that they can interact with to effectively, you know, do lots of the tasks that people have always done with data. So because of our investment in our APIs, agents end up kind of riding along that infrastructure, I think, quite nicely. Now there's lots of things we have to improve and build for agents. So for instance, agents don't read, you know, documents in the same way that people did. It's much better if you can sort of get them chunks of data from a variety of sources and then they can parse that data and make decisions about, you know, do they want more insight from a particular resource or another. So you can actually feed much more data to agents that way, and they can effectively parse that data and use their judgment on what they want to look at. So we have to add and build APIs that work really, really well for agents. This could be our search infrastructure, this can be context retrieval infrastructure, this could be how do you navigate our file system as efficiently as possible. So we built a bunch of tech just for agents. But we also have to keep in mind the core user experience of the product as well because people are obviously going to need to be able to still navigate this data and work with this information. So it's really about building a system that both people and agents can use simultaneously, and doing that with a very strong user experience with really robust APIs for agents.

Yeah, so this super exciting thing is that enterprises are unfortunately reluctantly used to the idea that when you create and share files and data, it'll be relevant for the first couple of weeks. You know, we'll do this podcast and there'll be a video and some audio files and some, you know, maybe documents associated with it, and a year from now it'll always be entirely forgotten. And the question is, what if all of that data was still valuable in a year from now, or in three years from now, or in five years from now, where instantly an agent can go and pull from a document from five years ago and then connect the dots to a document a year ago and then produce some kind of new piece of content or answer? That's what is so powerful about now the relevance of all of this unstructured data. Now this comes with a little bit of a caveat. So the data is insanely valuable to us. The problem is you also have to worry about, well, is the agent going to pull from the right data? The more information that it has access to, right, because there's been an issue around when you see too much data, it doesn't produce good answers, right? It has to understand the context and what is relevant, what is not.

Correct. So you have two specific issues. One is that you have this thing called context rot, and that is if you use up more tokens available in the context window, that the accuracy of what the agent does gets less and less the more you kind of put into the context. So it's not as simple as, well, what if we just gave the agent the 100 documents for it to look at, because what'll happen is it'll start to kind of lose track of what it's doing the more of that data it has to be able to work through. So that's one problem. Another problem is actually a problem that even humans have. So agents are no better than humans on this dimension. But if I sent you into a new company tomorrow and I said, 'I want you to go figure out the safety policy of this company,' and you do a search and you say safety policy and you find a document from 2025 or 2024 that has safety information, how do you know that that is the most up-to-date document? Right now, you have to probably type in that same search query and type in 2025, and then you have to type in 2026, and then you have to look, are there other maybe names of that document that were produced more recently? What changes have been there? What changes have been made? But you have this fun counterfactual. You don't know that any changes have been made. So how do you search for things that you don't know have happened? So there's this really interesting, like, how does the agent know about the unknown unknowns in the data environment, and when should the agent give up on trying to keep working through that? Right, with people, what you do is you have one extra bit of benefit because I can go and I can eventually, as a person, I can go and be like, oh, Bob runs the safety team at this company, I'm going to go confirm with Bob that this is the most up-to-date information. But what if only Bob knows that this is truly the most up-to-date information? Well, the agent can't really go talk to Bob every time somebody does that query, right?

So what that means is that companies have to, first of all, you need insanely intelligent agents. You need a tool set that the agents can use to get really, really good at finding what's verifiable and what's accurate and what's the most up-to-date information. But then you need eventually really good tools at your capability level for data curation, for managing the information and for organizing it. So that becomes another area of work that we have. But what's exciting is that it's a time for enterprises to be much more thoughtful about the information that they're holding on to. How do they organize it? How do they govern it? How do they label it? How do they name it? What's the taxonomy and the data structure? Right? All of those things are now really interesting problems.

Yeah. So we have a whole bunch of features for better context management, better curation of data. How do you have kind of clusters of content that relate to specific topics that you want to give agents access to? All of those things are things that we're building out.

Yeah. I would say access controls and permissions might be maybe one of the most important things to get right of all the things in AI, because if you just imagine, you know, the problem is the moment an AI agent has too much access, that you can guarantee they're going to leak data. They just, they will guaranteed leak data, right? Somebody will go ask an HR question and get HR answers back that they shouldn't have access to, or there'll be a chatbot that will be tied to an external customer environment that has access to internal data that can reveal that data in the chatbot.

Internal leaking can be just as bad. You could have information barriers that get compromised. Think about a law firm that deals with two different companies that are on opposite ends of a particular case, right? These can have massive ramifications to the business or entire kind of industries. Then you have the external kind of risk of I can have an agent go in and maybe an agent goes and tries to find information that is overexposed in a data environment. Both of those things kind of create major challenges for enterprises to deal with. So that's the overexposure of data. Now imagine the underexposure of data. In the underexposure of data, you basically get no automation, right? Because now the agent is always blocked. They're always running into a dead end. They're always running into a wall, or they can't access the resource that they need to make a decision. So then it's kind of not that useful because I couldn't really automate anything when it just kept bumping up against the wall in its workflow. So there's this kind of, you know, perfect balance that you have to get right, which is not underexposed, not overexposed, and until you get that balance correct, you really can't drive automation or you can't drive secure automation in your organization.

Yeah, it's a super interesting question. We actually aren't yet doing much on the proactive front, simply because our surface area is so wide at Box that it's sort of hard to have one perfect kind of heuristic for what to be proactive about, because you know you have everything from a life sciences company to a bank to a media company to a manufacturer. And so we haven't yet gotten to the point where we feel comfortable that we shouldn't sort of be nudging you. Right. The thing that we're thinking a little bit more about is how do you autonomously build these agentic workflows within Box. And so how do you tell us that you'd like to go automate your client onboarding process, right? You'd like to go automate your data intake process, or you'd like to go automate your contract management process, and can we make that more and more autonomous over time? That's an interesting problem that we're thinking a lot about. So it's less that it's proactive, but that it's using the best capabilities of AI to help you go and automate a workflow as much as possible, an existing workflow, because I think a lot of people don't either don't think, or they don't want, or they don't have time to essentially make their workflow autonomous.