Aaron Levie on AI Agents, Enterprise Data & the Future of Work

Thank you for your hospitality, too.

Yeah, it's been a wild ride. We started Box with a pretty simple premise: we wanted to make it easy to access, share, and work with your data from anywhere. And it started to take off and we got pulled into an enterprise context where enterprises wanted to start to store, share, collaborate, and manage their enterprise content at scale. That was really the journey we started on, which was we want to be the platform that helps companies manage all of their most critical and important enterprise information and content. That was the journey we had been on, providing a secure platform for automating workflows around content and embedding into business processes. And then really the biggest shift in the past three years, which is pretty obvious, is just the amount of change that AI is going to cause in how you work with your data. And our business in particular, if you think about this from the customer side, think about how much unstructured data you used to have that nobody really knew what was inside there. You couldn't reuse it over and over again. Lots of untapped value inside of all that unstructured data. And the power of AI agents is that they can just go and be deployed against all that information. You can ask the agent, go find me this resource that I'm looking for, or go process all of these documents, these insurance documents, and extract the critical data from them and put them into a database to automate a workflow, or have an agent understand my business and provide recommendations for what to do differently in my operations. So all of a sudden, AI agents let us now tap into all of the value that's inside of our enterprise information. So now the journey really is about how do we transform the way that companies work with all of their unstructured data and enterprise information with the power of AI agents.

Definitely not. Interestingly though, we were thinking AI in the way it exists today three years prior. And what had happened was three or four years prior there was this sort of brief AI wave that was happening. If you go back to like when Alexa was blowing up, Siri was getting better, Google had launched these vision models that were really good at doing things like processing images and scans of OCR documents. And so actually at that point we thought AI across all of your data would let you now tap into this information in this spectacular way. The problem was every single type of data you worked with generally had to have a different AI model trained just on that set of data. And so we had no way to generalize AI across all of the work that people were doing with their enterprise information. So in like 2016, 2017 we were all in. Unfortunately, we bumped into this wall which was it was just so unscalable because the models just could not work across enough use cases. So by 2020 that had sort of gone down a bit and we were on to COVID was happening, more of the remote distributed work dynamic, and so we didn't have the same level of focus and intuition that this would happen on this time scale.

The models were able to span everything from you could ask them legal questions, you could ask them marketing questions, they could do coding, they could do research. So that level of intelligence packed into the model opened up a new set of use cases that we could finally go after.

Very different. And we didn't even, the pre-training processes, the post-training processes, certainly the compute that was available, you just couldn't have had all of that come together. It needed some time to bake. But I think that ChatGPT moment was a very big deal because it sort of showed here's the power of taking this LLM and turning it into a conversational interface and letting you go back and forth with the agent and the agent goes off and does something and then responds to you. That basically became the foundational architecture for effectively everything that we see today in agents.

Yeah, I mean I think you kind of just set it up which is there's so much telemetry out there in the real world and there's a lot of telemetry out there with specific customers. The problem is that previously it didn't necessarily serve you that well to have massive amounts of unstructured data before because it was very hard to query that data. It's very hard to analyze it, very hard to ask questions of that data. So you could do this in very structured data forms, right? So if you go to like a cousin industry like financial services and banking, you know, I can go to a Bloomberg terminal and query it and analyze all of the stock information. I can do analysis inside of spreadsheets. Now go back to insurance and just think about all of the unstructured masses of data. There's policy information to your point. There's weather information. And there's these sort of market signals coming in from a pricing standpoint. And so you have to have some way to be able to run some type of machine across all of that data. And before we could never do that because it's unstructured. And so if it's not structured, then software has been classically very hard at being able to process that data. AI kind of gives us the first opportunity where we can do that, right? Because you can now take the intelligence that's inside the AI model and instead of using that context window for chatting back and forth about a general purpose consumer query, you use that context window to give the agent context about now the insurance claim or the market information that you need the agent to go and work with. And so now all of a sudden you have, especially with context windows that are now in the hundreds of thousands of tokens, you could fit entire books worth of information into that context window that you can now interact with. And so again, that could be an entire policy for a client. It could be an insurance claim process, all the images coming in, all of the documentation coming in, analyze all that, generate a packet for me of that information. You could take market data, be able to have the agent analyze all that information. So I think the setup of we have more data than ever, yet to some extent we're information poor in terms of our ability to get real value from that, that's going to go away in the coming years. And AI agents will be the reason for that because now they can just be cranking and combing through information 24/7, processing anything, analyzing anything, generating new information on the fly for you. And that's just this incredible breakthrough that we have.

Yeah. I think AI is a tale of two cities right now. You have the normal kind of early adopter cohort and in any market in history that's sort of like 5 to 10% of the market more or less. And basically that's cutting across every industry because engineers are often early adopters and they're using things like AI coding agents in a very effective manner and then you'll have these pilots across all these industries and or very applied use cases. But so you have on one hand I think a group that is seeing these incredible breakthroughs from AI and the other hand there's sort of everybody else. And what I think the biggest distinction is where AI right now does so well is where the agent has almost complete access to all of the context that it needs to be able to automate whatever it's doing. And in things like coding for instance the agent generally has access to the whole codebase and then thus can help accelerate the workflows in coding. And then you look at the rest of knowledge work and you think about how messy the rest of knowledge work is where context is coming in from three or five or 10 different places. Sometimes it's a physical environment, sometimes it's a Zoom meeting, sometimes it's a Slack conversation, sometimes it's your documents. So you have lots of context that's fragmented. You have in many cases a less technical end-user audience than obviously in coding where coding by definition you're going to have the most technical audience. You have models that haven't necessarily been trained on just these use cases, right? Because we're only now starting to have models get trained on a broader set of knowledge worker tasks. So I think you have kind of four or five slight impediments relative to again the group that is seeing the most gains from AI. But what I see in our purview is each of those are just going to get knocked down in the coming one to two years. So I think it's realistic that insurance is early. I think in some areas of life sciences we're early, in even many parts of banking we're still early. So I wouldn't say that I'm seeing any one industry sort of behind. I think it's more of you have the engineering kind of automation use cases and then everybody else. And everybody else is in the category of trying to upgrade their workflows, their data sets, how they organize their information, and then the models themselves are going to get better and better. And so I think we are early across knowledge work but it will come in the coming years.

I'm very optimistic on this front and this puts me in a slightly different category than maybe some of the AI labs. I tend to see that AI mostly is effective at augmenting our work, as opposed to replacing how we do work. And even when I look at AI coding, for instance, again this is the area that AI can do the most. And even so, it really takes an expert engineer to figure out how to use AI coding agents, to figure out where to deploy them effectively, to figure out how do you nudge them when they're going off and going in the wrong direction, to be able to review their work to ensure that actually you've built good, high quality code. And so I think that companies are actually in many areas going to be hiring even more engineers because what's going to happen is AI is going to make it so you can do way more with your engineering capacity. That will actually cause you to sign up for even bigger projects than you would have before, right? Which then will mean that you end up having to need to hire experts to go do those things. So even in the maybe the most disrupted category of work which is coding right now because of again the quality of the agents, I actually see this as a growth area of headcount across the economy. Maybe not in companies that are fully staffed with engineers but in all the companies that previously couldn't have hired engineers. Now all of a sudden I think the demand for that will go up over time. So I look across the rest of knowledge work and I kind of see similar analogies maybe holding for lots of areas of work where if you can lower the cost of doing some kind of task, so now an individual can do more of those tasks per day, right? Then the question is what kinds of companies would want to do more of those tasks as opposed to the same number of those tasks at a lower price. And that'll be the big question for the economy is where in the economy is there no more demand for that particular work and so thus if I lower the cost of that work there'll be less demand for it, versus where in the economy is there actually unmet demand whereby if you could lower the cost of doing it the demand would actually go up thus meaning that actually there's going to be even more jobs or at least the same number of jobs in that kind of environment. And I think there's going to be a lot of areas that will surprise us that will be in the latter category. I think we're going to have more lawyers in the future because I think what's going to happen is AI is going to go and find the most esoteric legal clauses in our contracts and they're going to start to fight over those which will mean that they need to be adjudicated and understood by an actual lawyer to understand what do we do in this negotiation about this one thing. You're going to see this in healthcare where AI will help us process our healthcare far faster and at a greater scale than we ever have which means we're then going to want to go see doctors to help us understand what analysis just happened and check on that more novel thing that we were looking into. I think you're going to in many areas of marketing even hire more people because when AI lets you generate amazing videos on demand, well, guess what? More companies are going to be media companies. They're going to start producing more content. So, in all of these areas, I think you're going to see that actually demand for work will go up. There'll be some areas where it gets compressed. There's no question. But I think there's going to be more areas than we realize where because agents can let us do more, the demand for those things that we're doing now goes up, which means that we have more people doing those things.