Aaron Levie: Why Startups Win In The AI Era

0:00You'll read a lot of press that basically says AI is coming for our jobs. is most of the press is not inside of big companies seeing how much time we spend on useless activities that are necessary but not strategic. There's a very, very long list of things that software never did before that AI agents are perfectly primed to go do now. And that's basically the opportunity set. Now is the moment this window will end. In this window between a year ago and three years or so, plus or minus from now, this is when the next hundreds of great companies will get started

0:40so Aaron and I go way back I don't know if you remember this I went and searched my old bump so bump was my startup I searched my old bump email for Aaron at box.net yes and I found very old emails where we would coordinate in Mountain View times to meet up with three founders yes you me and this guy named Sam Altman yeah and we would coordinate to try to like do brunch or lunch or whatever and then I saw emails between us saying like oh yeah well Sam's gonna probably like not show up again so that's how we got to know each other. Did he show up? No he would always just not show up. I don't know if you remember this.

1:11You started Box probably before many of the people in the audience were alive. Okay. Let me just show of hands. Who's below 20 years old? Wow. Okay, great. Yes, then accurate. Very accurate. Maybe to start, we're going to talk about AI a lot, but to start, you went through another major transformation. See, I used the word transformation. Yes, thank you. Digital transformation around cloud, going to cloud. Maybe just walk us through what that looked like at a very high level, and maybe what is different or similar about the transition to AI now. We started the company in 2005. And this was a time where you have to literally go back 20 years and think about a world where the internet was much slower, browsers were way worse, we didn't have the iPhone, we didn't have Android, Chrome didn't exist.

2:04Everything was just way worse on every dimension. We basically had an initial idea that as the internet got faster, as you worked on more mobile devices, you'd want to be able to access your data from anywhere. And that was the original idea of Box, where we said you'd go between different computers, you'd access your files, you'd share them, you'd collaborate. So we launched the company. It was initially focused on the consumer market or consumers slash just kind of prosumer, anybody that wanted to sign up. We started to get a little bit of traction. And by a little bit, we're talking like 10 people signed up in the first week or something.

2:41So it was just very, very slow, very slow and steady growth. And what happened was we got a little bit more growth. We got some early funding from Mark Cuban and some angel investors. We then dropped out of college. We got, you know, sort of more of an upswing. We had a freemium business model. So we let people sign up for free and start to use the product. And then one day we kind of ran into this fork in the road, which was, do we stay the consumer, go down the consumer path or do we pivot to the enterprise? And the calculus was, we felt like it was going to be way too hard to compete with all of the consumer technology platforms that would give away storage for free.

3:24They'd sort of embed it into their operating system or their social network or whatnot. It would be way too hard to go in and monetize this. So we decided to pivot to the enterprise where we could be cheaper, faster, easier than a lot of the big incumbents at the time. So we pivoted to the enterprise and we got extremely lucky on the timing because we rode this growth wave of mobile and cloud that were sort of working in tandem to effectively create a new IT architecture within enterprises. And so for us, we got to ride this wave where once we had better security, more functionality than a lot of the incumbent services, as companies moved to the cloud, they needed a way to share their data, access their information.

4:06And so we became an increasingly obvious choice. So that was the cloud wave. And that kind of propelled us to where we're at today. And there's a lot of similarities to the early days of cloud and the early days of AI with maybe one big difference, which is the early days of cloud, we were having to go convince people that the cloud was going to be this big deal. We had to go tell everybody that the future is going to be cloud computing. It's totally safe to trust us with your data. A lot of people didn't believe us. And so that meant we just couldn't win deals in entire segments of customers.

4:37So conversely with AI, you're no longer really having to convince people that AI is the future. Everybody tends to be bought in in the enterprise segment. There's a lot of still slowness in adoption in large enterprises, but it's not because people aren't convinced that AI is the future. It's just because there's lots of natural sort of pace of change that an enterprise has to go through. Why are they convinced? Is it just that they themselves personally have used ChatGPT. Yeah. Is that the main driver? Because I'm not aware of a lot of AI solutions that are deployed to enterprises that have really made a difference.

5:10Yeah, I think it's maybe unlike cloud. Cloud didn't have decades and decades of societal level conversation about cloud. It just emerged one day and it was like, this is like, it seems kind of cool and efficient. But if you're in an IT department, the cloud was actually very scary because you're taking your servers that you manage, you can see them, you manage all the software for, and you're relying on AWS or Microsoft or Google to manage that infrastructure. And so there was a real big shift on IT and the CEO or the head of marketing, the head of sales, they didn't really care how the infrastructure was delivered.

5:48So you didn't have anybody kind of pushing on the IT org saying we have to go to the cloud. Like nobody really cared. AI, totally different situation. We've had science fiction for probably 100 years that has basically said you're going to have robots, you're going to have artificial intelligence. Over the last 20, 30 years, it's been in the zeitgeist, self-driving cars, watching Watson on Jeopardy, using early products like Siri and Alexa. So it's sort of been pervasive that, okay, at some point, AI is going to get good enough that it's going to be this helpful aid for us. and now that you have the ChatGPT moment where the head of marketing can go and play with ChatGPT and be like, wow, this seems to write marketing copy maybe better than even my own marketing people.

6:32You don't need to sell them anymore that AI is like clearly the future. Now it's actually just about like, how can you go implement something that's going to be safe, reliable, works with your data, you can trust it, which is now the new set of changes that all these companies have to go through. Got it, cool. So Box started as basically like a folder in the cloud, effectively, and then you added a bunch more stuff to that, but that is still kind of the core of it. AI seems to be able to completely change what you can do. Maybe just help us understand what are those cool things that you can now do for big companies.

7:03Yeah, so for us, the exciting thing is that AI agents basically thrive on unstructured data. So if you think about it, there's basically two data types that really matter in the world. There's structured data. This is what goes into a database. If you launch an app tomorrow, you're going to start with a database and the stuff that's going to go in the database are like customer names and IDs and user IDs and all that. If you go to a big company, the stuff that goes into a database is all of the invoice numbers and the client record numbers and the amount of revenue they generate and their distribution partner names.

7:35That's what's in their database. Then they have a lot of unstructured data and that's all of their documents. It's their contracts. It's their invoices. It's their marketing assets. It's their presentations, all of that data. The vast majority of data in the enterprise is that content. It's all of this unstructured data. And it's called unstructured because basically it can be totally freeform text. There's no inherent kind of computer structure to it. And so the problem is, is all of the data that goes into something like Box, historically, you've never been able to really automate anything about it.

8:05If you just think about two years ago, you can't go to all of your files and ask them a question. You can ask a question in your database. You can say, please find me all of the records above the following value. You can't do that in your files because the computer doesn't know how to read all those documents and understand what's in them. AI agents basically changes this. So all of a sudden, all the data that's inside those folders becomes immensely valuable to companies because now they can ask all that data questions. You can begin to automate workflows around that data. Our whole vision is basically what if you turned all of this information into this new kind of corporate asset or set of knowledge that companies can operate off of.

8:45And that's where, you know, I think there's going to be immense startup opportunity is a world of how do you have AI agents for almost every task or job function in the enterprise. Let's talk about that then. Let's say this world emerges and we have AI agents that do a bunch of jobs. I think a lot of people are worried like, oh, that means that we don't need the humans to do those jobs anymore. And I know you have a strong perspective that like, no, actually it will go the exact other way. Yeah. Tell us why you believe that. I think basically if you go to most companies and you sort of say, tell us everything that you do all day long across the company, and you sort of assess how valuable is all that work that's getting done, how valuable is every email you send and all of the time you spend going and finding information or all the manual work it takes to read data, kind of look at that document, extract information from it, versus the time that really is the high impact stuff.

9:41You're with a customer. You're coming up with a product breakthrough. You're supporting a customer to use more of your product. And you kind of did a ratio of that time. The vast majority of time inside of a company is on the stuff that really is not strategic. It's sort of necessary work, but it's not strategic to get done. So when you think about that ratio, if you could free up a company to work on the stuff that's strategic and not the basically unstrategic stuff that doesn't differentiate them, most companies actually have a large set of things they would go do with their time. They would spend more time on breakthrough innovation.

10:15They would spend more time with customers. They would launch more marketing campaigns. They would proactively support their customers instead of just being reactive. The reason why I think the press gets this wrong, and you'll read a lot of press that basically says AI is coming for our jobs, is most of the press is not inside of big companies seeing how much time we spend on useless activities that are necessary but not strategic. And so when I go talk to companies and I say, what if you had AI agents do all of this kind of work? They instantly, their eyes light up because they realize, well, now I can actually free up my time and my employees' time to go do much more interesting things.

10:52Or they start to have this list of all of this work that would be much more strategic if it got done, if AI agents could go and do it, as opposed to the work that never gets done. because it's too unaffordable and it's just economically not viable to go into. Yeah, this is like the backlog of stuff in your company that you're like, oh, if I had more people, I could go do those things. Exactly. But I can't. And basically, there's an entire category of work where if you just did like pure microeconomics, I could pay for the labor to do that work if I knew that it would produce enough value to pay for that labor.

11:23Okay. But the threshold of starting that work is too high. I can never even try and see if it's useful. Okay. So I would argue that literally, like, if we go 10 years into the future, the vast majority of work that gets done in 10 years from now will be work that today is in that category. It's the work that, like right now, we can't even attack because we're like, I'm not going to hire somebody, pay them$120 ,000 a year to just see if that thing produces value. So I'm never going to get around to it. And then in 10 years from now, when you just deploy AI agents everywhere to go do those things, we will be doing so much more as a company.

12:03When we launch an ad campaign internally, we translate it into like three to five languages. Our top markets, that's about all we have time for because it's just too expensive. It sort of just hurts your brain to think about doing it across every segment of the market in every region. When an AI agent just takes an ad copy, translates it into 100 languages, our company will just grow more. We will just be in more markets. We will serve more customers. And agents will be the reason that we were able to do that. Where previously we were bound by people time and we would never have been able to justify getting that work done previously.

12:39Makes a lot of sense. And yet today, Amazon announces that you should expect that they have fewer headcount over the next few years because of AI. I totally agree with everything you just said. but then the press sees these announcements. Yes. What are they to make of that? To be fair, I only saw that snippet literally one hour ago. So I didn't see the full memo. I'm sure Andy Jassy had some other thoughtful points. This is why startups are in such an incredible position. You know, I think if you're at the point where, you know, I don't know the corporate headcount of Amazon, but let's say the total headcount is in the hundreds of thousands to low millions, just across all, like every possible, you know, every delivery function, etc.

13:18I could totally see the scenario where for them, they're like, okay, given the markets that we're in, given the things we do, you know, if we can't get this done with hundreds of thousands of people and AI agents don't just augment that, like we're probably running the company wrong. I'm just picturing that's the internal kind of corporate meeting. But now imagine the 50 person company where all of a sudden they can act like a 500 person company. Then you just have to ask yourself if the 50 person company can act like a 500 person company because of AI, will that company become a hundred person company more quickly than pre-AI?

13:53And then that basically tells you, does this thing create jobs or not? And my argument would be that the 50 person company that is in more markets, serving their customers better, doing better research on their customers, they're more armed with the next feature they should build. They can build that feature faster because of, you know, cursor, windsurf, et cetera, replet. Will that company grow more quickly in a post-AI world on the human side? I would argue yes, because they need to get themselves into more markets, they get more done. So I think it's more of a case of you're going to read headlines about the biggest companies, Amazon, et cetera.

14:26And I think there's a case we made where AI is an efficiency gain for them. But now the hundreds of thousands of startups and small businesses or millions of startups and small businesses, I think it becomes an economy where they can get so much more leverage than ever before. Tell you about startups. Like, I think maybe a lot of folks in this room look at the like B2B SaaS companies or the enterprise SaaS companies and just think like, oh, every problem has been solved. Like there is a big company incumbent. Like you are one of those big company incumbents. How should they think about like starting a company that could one day take down a company like yours?

15:01Not yours specifically, like the other guys. I'm not going to give you any advice on taking me down, but I'll give you advice on everybody else. So interestingly, it's a very fascinating proposition in question. So starting with consumer for a second, three years ago, I was having these kind of like, not like existential questions, but like deeply like deep philosophical questions. What year did you join YC? 2022. Okay. Actually, so great timing. So around 2022, I kind of made this list of like nouns and verbs of just as like a, just a fun kind of mental thought experiment of like, think about all the nouns and verbs of like what we do in our life.

15:43We eat, we sleep, we, we travel, we watch something, we are entertained or whatever. And I went through that list and, you know, the list is not a thousand words, right? It's like 50 or something. And basically down that entire list, I tried to plug, has that problem been solved relative to like 15 years ago, just choosing an arbitrary point in time. If we had this conversation in 2008 and we said, you know, music, travel, entertainment, hospitality, you know, food, all of these things, we would basically sit around and be like, all of these things kind of suck. It's just like, it kind of sucks to get food.

16:23It kind of sucks to get, you know, listen to music. Like you had to download illegal music. Everything was painful. Fast forward to 2022, we've solved a lot of problems. Like when I want food, it comes in 20 minutes from DoorDash. When I want to listen to music, it's on Spotify. When I want to watch a movie, I got Netflix or YouTube or whatever. So it was a tough environment for startups because you're kind of like, wow, now we're really only able to do derivative things because the core nouns and verbs like have been solved. And, you know, coincidentally, YC basically created like half of them and or more.

16:57And so basically we had this period of 2008 to 2014 where every noun, every verb just got solved. The same was largely true in the enterprise. So now do the nouns and verbs in the enterprise. Payroll, CRM, email, calendar. You just go through all of those things. And basically every problem had some kind of incumbent or at-scale startup, which is very bad for startups. because you basically had this era of companies that knew how to build a modern technology and they were solving these problems. Like you don't really want to compete with Gusto because Gusto is still a modern, really good payroll system.

17:39There's not like a lot of vectors into competing with Gusto. So that was three years ago. Today, it's the first period in probably about a decade where I'm extremely confident that there's now a new set of nouns and verbs where startups are in the right position to go and create the next set of solutions for because AI has created enough of a change in the landscape to create those opportunities. They're not going to be always the most obvious things that you start out with. Like it won't just be like, oh, it's CRM, but with AI because Mark Benioff in Salesforce is going to do CRM with AI. Like he's going to figure out a way to do that.

18:18They're very good at executing, like that will happen. But there's a very, very long list of things that software never did before that AI agents are perfectly primed to go do now. And that's basically the opportunity set, which is what categories of professional services or work is there no incumbent technology for that AI agents are basically finally able to go and solve? and there will be 100 startups that get created between last year and in three years from now that all become five, 10,$20 billion companies because they're able to find the next set of nouns and verbs or a mixture of nouns and verbs that are like, okay, it is legal work for this thing.

19:02And there's an agent. And for the first time in history, you can go and deliver that by a software as opposed to it used to only be able to be delivered by people. And that's the opportunity that I think everybody has. Very cool. YC's next batch is now taking applications. Got a startup in you? Apply at ycombinator.com slash apply. It's never too early and filling out the app will level up your idea. Okay, back to the video. When you create one of these new nouns or verbs, a lot of them don't look like software in the sense that we think about software today, where you like sell a company access to some number of seats of the software and the humans click the buttons and type the keys.

19:44How will business models need to change or will they change? How will companies charge for these things? Yeah, so if you were building a SaaS company, again, prior to literally this year or last year, your only real monetization strategy was how many humans are there that need licenses to my software? And in the SaaS world, we call those seats and basically how many people need a seat of that software? and you were maxed out based on the demographic size of that particular category. So if I sold software for lawyers and I go to a company, I can only sell the amount of licenses as that company has lawyers, which is like a huge limiter to the addressable market size of your company.

20:31And so agents basically completely blows that up because all of a sudden you can have AI agents that effectively contain the labor of that job function in the software itself. So you can go to a company and you can say, I know you only have three lawyers, but my agents could do the amount of work of basically unlimited lawyers, which means you're obviously no longer going to sell based on the number of, you know, humans in that company related to legal work. You're going to sell based on some approximation of the amount of volume of work that has to get done related to legal work. And that's the new monetization strategy that I think we all have, which is, let's make the example of you're doing some kind of legal review of some set of contracts.

21:17And you basically say, okay, previously a human would cost$5 or$10 per contract to review based on human time. AI agents, you don't tell them this, but AI agents, let's say, can do this for 10 cents. So then you charge that customer$2. And all of a sudden they're like, wow, this is incredible. You've just saved me 80%. And, you know, you're now extracting, obviously, a very meaningful profit from that. And there's no particular limiter to how much they're going to pay you. It's just going to be how many contracts do they have to go through the system. And so, you know, every company, I think every space is going to have a slightly different version of that business model.

21:53But the new business model is some form of consumption. The only, you know, concern that you have to have of going, of overly veering on consumption is the recurring nature of the revenue. You generally want to be in a position where you have some kind of subscription fee for your software as opposed to only being paid, you know, the moment that it happens. Because then you run into this problem where the customer, you know, plows through your system. And then next year they don't show up because they use they use they're done. They've reviewed all their contracts. You have to figure out how you basically keep some kind of ongoing recurring revenue stream.

22:26But besides that, you're going to see more of a consumption orientation with AI. And I guess what you're saying is you think the prices these AI companies will be able to charge per unit of work, let's call it, or outcome, will be a fraction of the human cost, as opposed to what we would maybe consider more of like a software cost, right? Because if it costs 10 cents to do the job, are people really going to be willing to pay$2 for a thing that they know costs 10 cents? Well, the question is basically, how much software do you have to build on top of the AI tokens? And, you know, it's like very obvious, like imagine a continuum where there's like almost like no software.

23:05You will get price compressed down to 2x max of the token cost versus a world where there's like a tremendous amount of software. you could probably support 80 or 90 % gross margins, which means a 5x plus, an 8 or 9, 10x plus increase over the tokens. I'm not going to tell you guys the number because it's proprietary. But if I told you the amount that we spend at Box on storage of storing files, you would be surprised. Because you would say, well, I thought you were in the storage business. But the reality is what customers are paying for is all of the software above the storage. So eventually we're going to get to a point where customers are no longer going to just be paying for the intelligence tokens.

23:53They're going to be paying for the workflow software that goes on top of the tokens themselves. They're going to be paying for your ability to build AI agents that have a unique set of context and connections and capabilities and access to data that can command a meaningful premium on the underlying AI token. Totally. Yeah, I mean, I saw this at Google Photos, right? And when we proposed that we should build Google Photos, a lot of people at Google were like, why would you do that? You can never make money. It's a commoditized market. It's just storage. Like, there's Amazon. You can just store your photos there.

24:27and it turns out it's like a 90 plus percent margin. Yes. It's really great. And this is what's amazing about, we are all incredibly lucky to be in, effectively, I'm sure there's other industries, but let's just say one of the top slash the top industry that has deflationary economics on the supply side. The reason why that matters is because it means that over time, your raw materials will get cheaper. You don't have to basically raise your prices in perpetuity like many other industries, you can actually just get more efficiency gains over time. And so, you know, I don't know the latest price on Google Photos, but let's just say it's 10 or 20 bucks a month, right?

25:06If you told anybody, we'll just store all of your photos ever created for 10 or$20 a month or whatever the number is. Like you'd be like, yeah, that's fine. That customer doesn't need to show up and say, no, I'm only going to do it for$8 because I know your costs are going down. They're like fine to pay 10 bucks for all their photos to be stored. Yet every year, you know, your underlying costs are going down. that's what's going to happen in AI. As long as you can find not how to not be so greedy that your pricing is sort of like, is kind of like offensive, you know, look at like, again, Windsurf, Replit, Cursor, et cetera.