Aaron Levie (CEO, Box) on AI agents and the future of work · AGI House, SF Tech Week

And I think you can basically correlate that to where the diffusion is most happening within the enterprise. And so how that shows up is, obviously, coding has gone fully vertical and taken off completely. And then you've got customer service and customer support fully vertical and taken off. I think we can see what that's going to look like over the coming years. And then as you go through the job functions, once things get a bit sort of lower at each bar in the histogram, there's probably legal and then there's investment banking and then there's accounting. And you kind of go through the list. And I think this is actually, to me, a very exciting thing because it's the reason why there's still so much opportunity in AI.

If everybody was using AI across the economy like we are here in the Valley, basically, you could just, you'd say there's no more opportunity left. There's nothing left to build. There's no more diffusion work. And the model intelligence is sort of all that is needed. But the reality is that's not the case. These models have to be brought into workflows. There has to be a significant amount of change management. There's going to be an entire applied layer of AI that has to get built out. And that's what is creating, I think, so much opportunity right now.

Yeah, I think it's, you know, it's interesting. We kind of, we always have this interesting thing of like, is it lagging? Is it, and it's like lagging, what is sort of the question? Like what, yeah, but like, I think that's just sort of always the rate of diffusion. When, you know, I have the fortune of I got to see the very early stages of cloud. And, you know, we all in Silicon Valley were instantly using S3 and EC2 within like three days of it launching. And within a year or two, it was basically like everybody in Silicon Valley and every startup was obviously, you would never rent a server again or build it out yourself, you were in the cloud. But it still took two decades for the rest of the economy to even get to the point where they even agreed that the cloud was the future. And we're still in the stages where you basically have, you know, large amounts of workloads that are still on-prem that are gonna be moved in the cloud. So I think we have to be prepared for this is a two decade long journey and rollout, which is sort of a little bit, you know, sometimes kind of hard because the closer you are often to the research and the acceleration of AI, we are seeing the progress of these models. And, you know, it's this interesting cognitive dissonance of how is it that we're seeing this incredible AI capability, you know, in our own workflows and in our own lives and yet the rest of the world still is, you know, using basic chat bot interactions as their main modality for AI.

I think it's mostly, we're, you know, I think maybe the only real maybe surprises is, is probably, you can't really predict who has adopted what because it is actually so, it's so honestly dependent on some early adopter being in the organization. And so it's kind of this weird thing which is you would think like, okay, so the big companies are really slow, right? And the small startups are really fast. And you know, it's directly correct that the startups are obviously gonna be working in a totally different way. But even at the big companies, you find one person that is sort of like super AI-pilled and they could be at Goldman Sachs or at Citigroup. And it kind of like, it doesn't correspond to the market cap of the company, doesn't really even correspond to the org that they're in or the function. It's honestly just corresponds to their personality. How wired into this ecosystem are they? You know, how addicted to the news streams are they? Like, you know, anybody who is here, you know, is gonna be obviously very wired in. And so it's honestly like a, it's a completely like, I can't ever know going into a meeting. Am I about to talk to the most AI-pilled person of all time? Or is this gonna be the slowest moving company because it's just 100% dependent on like, almost like up to like, what is their X feed? And that almost is like, if I could read, if I could see their X handle before every meeting, I would mostly be able to tell you like, where is the company on its AI journey? But without that information, I almost have no other predictor of how far along that person is or that organization is on its AI adoption.

I think the ones that are getting it right are, you know, unfortunately, it's a very long list. So, there's no panacea to getting it right. I mean, it's some mix of who is sort of going off of who has a good sort of sense of the current of the industry versus who's just sort of listening to the vendor that came and showed up and pitched them on one particular thing. Who has a good pulse for, I mean, just like literally what best practices are emerging versus what are they encumbered by with their existing legacy infrastructure and thus stuck on that particular paradigm for what they're gonna be able to do with AI. You know, there's a lot of, you have to somewhat have a kind of a spidey sense of like what things really are taking off and what products or services have a really, really positive flywheel that are gonna be in kind of almost a self-improving loop, and then which things are gonna be left behind. And that's a decision about what vendors to work with. That's a decision about what models to choose. It's a decision about what evals to trust. Sometimes even just like it's as simple as like, do you have, does a company have evals for their own workflows versus not? And that's a major determining factor on like, do you actually, are you really gonna be able to automate that workflow if you don't even know how AI is impacting that workflow to begin with? So there's just a lot of blocking and tackling that has to be done by most organizations. So yeah.

Well, I have pretty high conviction. It's also somewhat of a low stakes claim because if I'm wrong, also things are gonna be way worse than me being wrong about this. So like, we have way bigger problems if I'm wrong on this prediction. So, but I sort of, my instinct on this is driven from one, just like a body of economic research and sort of industrial revolution research that we've just seen over kind of 100, 150 years, as well as just today what you're actually seeing on the ground of the companies that are the most AI-pilled or the organizations that have rolled out AI most effectively and what does it end up impacting versus in terms of what is our hunch versus what's reality. Now there's a couple of X factors which some of my views are somewhat dependent on, not necessarily model progress but the entire ecosystem of progress and how much can AI successfully automate between all of the systems that AI has to interact with. So there's a couple kind of parts that are hazy about maybe my long-term views but for the most part, what we tend to see is that for most of the areas where AI has been able to deliver impact, let's say engineering, even marketing, even sales, the companies that are using AI the most not only are continuing to hire in those areas, they are often actually growing faster than the other companies. They're getting more work done. They're getting more work done, they're delivering a greater value proposition to their customers and that's causing a positive flywheel to hire even more. And so we have not yet reached a point where it's obvious what the upper limit of that is. So I'm sure most people here are at startups but you can kind of feel it. Like when your company is doing the best, which will often be correlated to the companies that also use AI, they're not doing the best because they use AI but certainly the best companies are using AI, that is accelerating their product innovation, that's accelerating their product velocity, that's often enabling, let's say, their sales team or their go-to-market team to serve even more customers. Well, what is the very next thing that those companies do if they're accelerating? They eventually find a new bottleneck in their organization that then they have to basically hire for, even if the person they're hiring is just to then go and use agents to go and accelerate their work. So there's sort of a recursive self-improvement even between the human and the agent and the need for more humans to deploy more agents and then the growth that that then leads to. And there's a lot of industries where that just doesn't have a particular sign of slowing down. Now there are industries where there will be headwinds but if I look at life sciences, life sciences, we have basically been constrained as a general kind of economy at the rate at which human scientists can do research, our ability to go test that research, our ability to go and work through the FDA process from that research, and then the ability to bring drugs to market. There is no set of first principles that says that the rate of biology and discovery in medicine has to be at today's current rate. There's no law that says it can't literally be 10 times greater. Obviously the FDA eventually will have to then sort of obviously change how it works through the system and our ability to go test new drugs has to obviously accelerate, but there's nothing stopping us from delivering 10 times more drugs for all of the next set of specialized ailments that we deal with. So what would happen if we could accelerate research using AI by 10-fold or 100-fold? I guarantee you life sciences is going to hire two to five to 10 times more people because we're gonna ultimately then be constrained by who can run the tests, who's running the labs, who's then delivering the drugs to market, who's doing the go-to-market, who's doing the supply chain, all of that, our entire new fields of jobs that will get created. You can kind of go industry by industry and realize there are many industries that are highly constrained based on our ability to go and deliver value in that industry. So whether it's software engineering, whether it's life sciences research, I actually have taken the other side of the bet on will there be more lawyers or fewer lawyers in let's say five or 10 years from now? And the reason for that is that I think AI is gonna get us into so many more lawsuits than ever before because just for all the obvious reasons, like AI is gonna create lots of new problems that we're gonna have to go adjudicate. AI is gonna cause even more negotiation in our contract processes because one team is gonna use AI to generate the contract, the other team is gonna use AI to reduce the size of the contract, and then eventually you're gonna get on a call with lawyers and say, whose AI should we actually trust on this? And all of this is just more lawyers. You just go through the list of all of the different roles. Think about cyber. Are we gonna have more security professionals in the future or fewer? Yes, there's gonna be security, all of the security practices will be run by agents, or agents will be kind of executing most of the actual work, but you're gonna need cyber professionals up and down the stack figuring out where do we go deploy the agents, how do we triage what they found, how do we go and incorporate that across the workflows of the organization. So I'm not seeing a sign of that slowing down in a lot of areas I'm seeing it accelerate. Now there will be jobs that have to change. If you were doing frontline customer support for handling tickets of a consumer internet product that the tickets you were handling was like a refund claim, or I can't get into my account, or I have some issue with my shipping, we can basically assume that's gonna be 100% automated. Now, does that mean that a similar job won't just open up or exist and then the let's say AI forward company that says I'm gonna go deploy people to go and proactively go work with our customers more? So the job of doing some form of customer success doesn't obviously go away. That's sort of unchanged. It's the specific tasks that you were doing previously in that job function that obviously have been increasingly automated. So I think there will be jobs that change dramatically but on the whole, I think we have many, many more jobs left that we're gonna both create and accelerate because of AI than reduce or destroy.

Yeah, but I mean, think about the other side of medical that I just brought up. If you could increase the efficiency of the healthcare system by 10, 20, 30, 50%, we absolutely have the need for three times more doctors. So if you could lower the cost of the delivery of healthcare, the next immediate thing is how could we actually go and see more patients more efficiently at a lower cost in more places. There are so many parts of the economy that are entirely constrained upstream by maybe it's regulation, maybe it's the cost of delivering the underlying service to the client, maybe it's the operational efficiency. Operational efficiency, maybe it's the unions. There's lots of these things where you could go very unconstrained and see a lot more growth in those spaces. Here's one final one. It was a funny one that I just ran into. We were trying to get an air conditioning person to fix our air conditioning unit at our house. And one of the biggest constraints was all the AC, HVAC people now have been called in to work on data centers. So we have a shortage of just AC experts in the Bay Area. Like there are jobs across so many different areas of the economy. Once you have a new, once you have a way of being able to actually kind of change the direction of labor and where we land a lot of this labor.

Most of my day is, you know, I think probably the way to think about it is my day often, I'm able to often compress a week or a month of work in a single day that I would have had to have a whole bunch of sort of side async processes and meetings around. And so I'm sort of able to, you know, compress a lot of time into my day due to the use of AI. And often that is some mix of, because of where I sit in the organization, I spend a lot of time on product. So my, you know, usually my go-to best use cases are how do I use AI to accelerate something in the product realm? And so the way that shows up is instead of the normal one to three month long brainstorm process where you have a meeting on Monday and then you do a little bit of progress and then you meet again on Thursday and then you have to skip a week because some people were out. And so then you're like two and a half weeks later and then you got a little bit more progress and then you have to pull the designers in. All of that can be compressed into a single session with agents. And so now when we get into the room with the 10 or 15 people that would work on that domain, you're just coming with the product and you're saying, hey, should we build this thing? Should it be this shape? Should we get the feedback from customers on this idea? And you've compressed again, like weeks and weeks of work to be able to like actually move that forward. So that's probably the single biggest change that has happened in my work. But that's because I'm very close to product and the engineering side. I think if I wasn't doing that, I think I might actually feel more like the jagged intelligence sort of scenario, which is like, okay. One thing I'm overconfident in is I'm not bad at email. So I haven't needed to automate email. I like the Whisper guys because sometimes I will just talk into the computer and that just makes it a lot faster. But I don't need a lot of that kind of daily productivity stuff. The idea that AI is gonna go look through my calendar and have some incredible insights is like, okay, that might be like, help me by 3%. It's usually stuff that's more transformational. Like, can I go and accelerate a product by three or six months because we can use AI in the ideation process? Can I do some amount of customer research that produces either major insights to accelerate our go-to-market engine? That tends to be how I've used AI in a way that is just different from all the chat threads I have on just basic research stuff.

Yeah, I think that there's actually no contradiction there. If you wanted to learn how to code up until literally four years ago, it was all basically the same process as it has been for 30 years. When I learned how to code, you would install some libraries, maybe read some online forums, run a server, put some stuff on the server, iterate, and you kind of trial and error your way through it. In like three or six months, either you have some kind of escape velocity or you're getting better and better and better or not. In my case, it was the or not. And I kind of tapped out very easily, very early in that journey. And then I had friends that were just 100 times better and I was like, I probably should just work with those friends and then really get the true burst in productivity. But it was like a pretty high barrier to entry because you didn't get that much fulfillment for the first few months of effort being put into the domain. And then your escape velocity was basically dependent on is your brain wired for this and how many hours can you apply this. AI lets you sort of compress all of that which is within an hour of your first time ever, you could get the benefit of, oh, that's actually why people write code and why they build software. Now, your ability to actually understand what's actually happening is very limited. So you'd have to have the agency and sort of resourcefulness to go and figure out why is it doing what just happened and how do I learn how that's working and how do I actually go and inspect the code. But I promise you, you will learn how to code in my opinion 10 times faster if I could tell an agent to generate the software project for me and then I can go and look at that versus the trial and error it took to do that exact same software project again in the way that I learned 25 years ago. It's usually guaranteed you would learn literally 10 to 100 times faster because you would just see, okay, that's actually how it works. That's how it's calling these functions. That's how it's calling the database. You would be able to just see that the pattern's far faster. So much, much lower barrier to entry. Anybody now can participate in many of these fields. But now you wanna go and do something really, really deep in that domain. You're at Box and we actually do need to store hundreds of billions of files and we need a highly distributed system to go and do that and it's gotta be incredibly secure. Do you want the person that just learned how to vibe code six months ago or a year ago on how to do that? Or do you want the person that has been building products and knows engineering for five, 10, 15 years that now has the power of agents at their disposal, knows exactly how to veer them down the right paths, knows when the judgment of the agent is wrong, actually has an efficient way to review the code in that process. Obviously every single day you're gonna bet on that latter person. So then the only thing is can the person who is entering the field get to be that latter person far faster as a result of AI? My conviction is yes. We will have more people be able to enter these fields if they so choose. We'll have a lot of people that still bounce out because they're like, yeah, it still is not for me even though I got the dopamine hit of building that software for the first time. I still don't really enjoy managing the system and learning how it all worked. That's not enjoyable for me but more people will discover that they enjoy it. But the depth of these spaces don't change. In fact, they go deeper. So who is the person that most likely is gonna accelerate RSI and research? Is it Andrej Karpathy using agents or is it me vibe coding an RSI agent? I'm gonna bet on him. So I think that still applies to bio discoveries, that applies to cyber, that's gonna apply to software development, that's gonna apply to financial investment research. So I think having that deep judgment with the ability to get a force multiplier with agents is still the best spot to land on. The great thing about AI is it lets more people get to that level if they choose.

Yeah, so this is actually the one part where there is a little bit of like a game theory thing that we have to do as I think as a fairly kind of cooperative game which is we could all sort of optimize for just our organizations only and basically you might end up landing on okay let's get just mid-level people that use agents and that is how we're gonna get the biggest boost of productivity in the near term. But what you would obviously do as an economy and as an ecosystem is you would have this increasing sort of gap then in your next generation set of individuals and leaders because you basically decided to forgo the kind of training environments and just go to the people that can go and use AI super effectively on day one. So I think it requires a little bit of kind of mass coordination to say no, we're not gonna do that, we're gonna still, we are gonna build these training programs, we're all gonna bring in people from college or wherever they come from and have them enter the workforce and yes, day one productivity might be lower than the mid-level person but the long term is we have the next generation of talent that has been trained and learned these jobs and I think that's just a collective action type of issue that we all have to decide to go and do as organizations. I'm generally pretty optimistic that that is what is happening. I don't see a lot of organizations like I fortunately get to spend time with like a lot of CEOs of Fortune 500 companies so I get to kind of see what's going on in the real world. I guarantee you to a T nobody is sitting around being like we're just going to profit off of this AI and we're just gonna, we're gonna use AI to wipe out our workforce and we're just gonna automate everything. I have not met a CEO of a Fortune 500 company that has that as their mindset whatsoever and I think there's a caricature of probably, you know, capitalism, there's a caricature of automation that leads people to think that's what's happening in these kind of closed doors. Most of the CEOs that I meet are deeply sort of trying to be thoughtful about this which is we do have to train the next generation. We do have to be, as an institution, we have to exist in 20 years and in 50 years from now and if we don't bring in a workforce that's excited to work at our companies and able to be productive in the future, then we're not able to do that. That is much more the case from organizations than I think what probably is perceived and so I'm pretty optimistic that we will be able to drive this now. There's a harder problem than the company side and it's probably what you do at universities. This one, I don't know how to solve because I do think we're gonna end up in a weird spot which is the university is gonna say, how do you charge $150,000, $200,000 to teach people things that are available on ChatGPT? And how do we start to make sure that that training ground is as, on some dimensions, modernized as possible and on other dimensions, kind of back to basics as possible which is there's other things that you need to get trained up on in those environments. It's preparedness for how to learn. It's a preparedness for how do you adapt to new environments. It's prepared for the rigor of logic and critical thinking and analytical thinking and in some cases, the least AGI thing or I guess in some ways, the most AGI thing of it's just offline learning of deep and hard topics and it's gonna feel like the least efficient way to learn those things for a year or two but you're getting boot camp for your brain to prepare for the next 50 years of you working. And so how do we work through the university system so it doesn't just turn into some weird, either dystopian or not necessarily modernized environment, I think that's a much more difficult trend to navigate that I have a lot of sympathy for people running universities right now when they're getting yelled at from every dimension on they're doing it wrong or they need to go this way or they shouldn't even exist. That's a very hard problem.

Yeah, and I'm very excited for all of the new ideas. At the end of the day, we still have millions and millions of students every year. I don't know that we've got the systems for everybody to participate in the new versions of this. There are only so many fellowships and Thiel fellowships and all these kind of things that you can go do. At the end of the day, we need institutions that can bring 10,000 people in a year and at scale find a way to, everyone gives universities shit and I think we're quite unfair about it. It's gotta be insanely hard to run these institutions. Now, I don't know who, I don't know if you did full root cause analysis, who's responsible for the cost, massive increases. That's probably the only issue we have is the cost of education is obviously way too high. It should be like a fifth the price of what it is, but short of that, which has a whole bunch of dimensions to it, I think you still need large scale institutions. I think when I was 18, I wasn't, we started and dropped out of college at 19 and 20, but there's a lot of reason why it's like, you're still kind of in this weird phase of life. Do we want every 18 year old to have to go into an investment bank and all of a sudden be grinding in banking deals? I don't know, maybe give people a chance for a few years to develop their wings and that's why I think the university, the university serves multiple purposes. So I'm more hopeful we land in a kind of a different outcome than probably what some are predicting in the AI space on this front. And we just do, we find a way to merge more ways of making, using AI to accelerate as much of the learning as humanly possible, but still pairing that with the real world grind of going to college and just that a way that doesn't put you in debt. So I'm probably more Bernie Sanders-esque on this dimension. So I think it's a thing that we probably just need to fund as a collective thing as a society.