LIVE: Box CEO Aaron Levie Discusses AI Trends At Reuters Momentum Conference

It depends if we've talked about token maxing.

The whole panel.

Okay. Weird. Okay.

Yeah, probably more and more. I think that's probably what we're all experiencing. And it's interesting because I think we all live this duality which is run as aggressively as possible, kind of adjacent to the token maxing ethos, and then immediately, but what is the ROI? And oh my god, I can't believe you just spent that much money on that thing. And so we have this sort of schizophrenia, I think, as probably we all face this if you're a CIO or on a board or CEO or anywhere you're kind of implementing this technology. And that's just as true at Box where we're like on one hand insanely excited when we go and deploy an awesome new agentic workflow internally, and on the other hand we instantly sort of see like why is that one user using like two-thirds of the team's budget in one day? And then you go and try and dive into that. So I think we're all sort of facing that. But maybe the biggest way to characterize the journey that we've all been on as an industry collectively is we're very clearly moving from this world where AI is about interacting with a chat system and answering questions and ultimately I would sort of characterize it as being rate limited by how much the human can do going back and forth with the AI, to more of a model where we now have agentic systems and you're no longer as rate limited by that human's role in the workflow. The agent can go off and do real work and actually, we're sort of untethered now. The agent's untethered by our ability to go back and forth with it and it can go and generate much more meaningful output whether that is in our world answering an RFP from thousands of past RFP documents or doing research on a market opportunity or obviously doing software development inside of a software workflow. That's clearly where things are moving. And so we're constantly looking around the organization to say where are the next big opportunities where we can think about this new augmented productivity and intelligence that we have and where are the workflows in the business that have historically been gated by intelligence. And that could be in the form of researching something or connecting data between systems or being able to look up information in the background. And we're trying to find all of those workflows in the company that we can go and unleash to ultimately then enhance the human side of the productivity.

It's increased, it's now firmly in the majority, but I would say we're not, we're by being an enterprise software company serving B2B customers, some of the largest institutions in the world, I would say we can't do the full like the things you see on Twitter from the startups. We can't do that. Which is not reviewing the code or having everybody has 50 agents running in parallel. We're not on that end of the continuum. Simply because the stuff that we work on is obviously incredibly mission critical for our customers. I don't know if anybody saw the recent example this weekend, but somebody had a Cursor agent delete his database and now he had to do a three-month rollback for his system. Well, we can't afford to do that type of execution. And so, how do you enable nearly a thousand engineers to be able to go and use agents responsibly both balancing the maximum amount of productivity we can get with the reliability and sustainability of the codebase and the security and compliance that we have to take on.

I think it nets out to for us, we happen to be in a growth side of the equation because, so it's a little bit meta, but because our product is growing faster because of AI we are on the side where we want to hire more people and then obviously use AI to make them more productive. So I think it mostly nets out where the growth ends up solving the cost side of the equation because we're in a pretty dynamic part of the space. So I'm not as worried about how do you pay the token bills simply because there's so much upside in the productivity that we can go and drive. But we look across the entire organization and obviously there's some degree of have and have-nots on who gets the most tokens and what parts of the workflows have the most leverage from AI. But overall we want to try and have every part of the business be as AI forward as possible.

Yeah. So, I'm 100% convinced that this is at a minimum a role that will emerge right now. Whether it fades out over time, I don't have a, I mean, I'm sure it'll be a fun debate for everybody, but for at least the foreseeable future, there's definitely a role. We have a couple people that have become this role and then we're going to be hiring for this role as well. Is effectively these kind of agent deployers or operators or managers in the organization. And I think despite maybe what Silicon Valley's intention would be or what would be ideal in an ideal world where the agent just kind of gets deployed inside of your system whether it's ChatGPT or Claude or a homegrown system and it just all of a sudden is off and running and automating your workflows. I think what we can all feel is that it's actually much harder than that. And the reality is we've seen these insane productivity gains from coding and I think most of the audience is relatively technical so you've probably done this yourself and it's just incredible what the AI can now do from a coding standpoint but if you look at coding and then compare it to other workflows in the enterprise coding actually ends up looking more, basically different, much more different than we ever realized before. And agents are almost emphasizing these differences. So with coding, basically you're inside of a text editor most of the day. The entire input information in and the thing it produces is text. So it's effectively all a text-based medium for the most part. The models are obviously insanely well trained on coding because so much of the code is already online in the form of open source and repos across GitHub. And then the users are obviously by definition the most technical users in your organization. So when Claude says I'm about to install this npm package, they can just say yes and move on with their life because they understand spatially what that's going to do. And when the agent needs to access an MCP server or talk to some other data system via CLI, they already have a fluency in what all that meant and how do I make sure that these systems integrate well to be able to produce the sort of ultimate product that I'm trying to do. So that's the coding world. And so then obviously you have this incredible escape velocity for AI coding. Now go to the rest of work where you're in marketing or finance or legal or supply chain and you're trying to get those same productivity gains. Well, you have a little bit of a different set of factors at play. One, a lot of the context that you work with is not text in and out. It's information that's analog. It's undigitized information. It's something that you just learned in a meeting down the hall that you have to figure out how do you get to the agent. Your systems are not as often cloud forward as a lot of the coding systems have been. So, you have a lot of data that's in legacy environments. And in our world, we see this because it's in network file shares or it's on-premises document management systems or it's in some ERP system that isn't playing nicely with agents. And then you have the real kicker, which is just the messy world of access controls and entitlements. And so the agent sort of only has partial access to the information it needs to do the work because the human only had partial access and so but the human could actually go and talk to the colleague and say, 'Hey, I'm missing this context. Can you open that up to me?' But the agent has no way of actually doing that inside that workflow. Or even worse, you have the opposite problem where the human has access to too much and now the agent has access to too much. And so now the agent is a security vector for the organization. And so you can't just give it free range access to everything. So that's the real world of the complexities of how do you get agents to start to automate the workflows across the rest of knowledge work. And so this is sort of the big opportunity. Now, tying it back and I'm surprised that I'm able to actually bring this back to the question because I was like going somewhere else with this entire thing. Yeah, there was a 90% chance I wasn't going to pull it off. But so to your question, I think the reason why you're going to need these AI embedded roles and they might come from the business is because of all the complexity I just mentioned. Who's the person that's going to go look at the workflow and say actually this is how we're going to redesign access to context? Here's how we're going to make sure the agent has access to the right data sources in the workflow. Here's the moment where a human does have to review the output because we know that there's still going to be a 3% error rate and so you want the human to review at this stage but it's not going to bottleneck the agent because they were able to do this much of the work. Like all of that has to be designed and it's basically doing systems thinking to figure out where does the agent need to show up in the business process? Where does the human show up in the business process? How do you get your systems ready for this? How do you start to document your workflows and the processes with the right context that it's not going to work if that's somebody's part-time side job? And you can see it. There's always a person, everybody has these people on their team right now. They're doing this at 10 p.m. Which is fantastic and God bless the free labor you're getting. But at some point, if you want to go and industrialize this, you probably want the person who's like, I am the AI transformation person. I'm the agent operator in go-to-market or maybe even the sublevel of sales or in marketing. And I think most organizations will end up needing that kind of talent. And my instinct is it's again it's like the future IT role most likely. But where it comes from and where it ends up I think is still open. And the only other final point is this is also why I'm super optimistic on the future of engineering because I think engineers are going to make the best, they're going to have the best characteristics for a lot of these roles. So maybe 10 years ago that engineer was building an internal application or they were doing something with SAP and doing an upgrade that same engineer or the future engineer coming into the organization will be well fit for now they are the person deploying agents in your business processes and I think if anything again back to despite maybe where the industry would like to go I think these jobs are getting more technical not less because of the capability of the agents are getting more powerful so now that's like you're handing a bazooka to possibly a non-technical person that could blow up the whole thing. And you might want somebody who really understands their way around these systems. So I see this as a fantastic future engineering job as well.