How to Build an AI Business from Scratch

The more I play with AI agents, I do realize that I need a person at the beginning of the process and the end of the process. So I still end up having more people. >> Some of it will be different roles, but I'm very optimistic that we're going to use this technology to grow more and do more as opposed to just replace. >> This is Aaron Levie, founder of Box. >> Welcome. >> A $4 billion company. 64% of the Fortune 500 uses his platform. He says we have 3 years to build the next generation of AI companies. >> These market windows happen every 10, 20, 30 years in technology. The mainframe, the personal computer, the internet, the cloud/mobile. >> If you were starting today, what would you do to find the right idea, to test it, and to make first money? >> My first thing would be >> Five days ago, you also said this, "We're at a unique moment of history where anyone with high level of ambition and core skills in any area can overcome a lot of historical experience requirements but a role." Can you talk more about that? >> So it's this interesting dynamic where uh a a younger group, not necessarily in age, but maybe in skill or time in that domain, so an earlier group in that domain, um can have as much leverage and in many cases even more because of their mindset differences than somebody that is like super experienced in a field. Now, and interestingly, the the the advice can go in all directions because you can, you know, you could have somebody maybe too early in that field and then um and then use AI in the wrong way and and get the wrong outcomes. Uh equally, you could have somebody extremely experienced that decides to adopt the technology and then they have a total superpower because they understand all of the contours of of whatever they're working on, you know, whether it's writing code or doing healthcare or doing biotech, and they will be actually uh even more capable of of leveraging these tools if they have the kind of right mindset wiring uh to be able to leverage them uh leverage them. So I think that the core idea is that we're just in this amazing moment where if you're super ambitious, you want to go deep in the technology, ideally you're technical or or becoming technical so you can kind of really know your way around these tools, you can make up for again lots and lots of years of of skills that that you would have otherwise had to go and develop. And I think that's an incredible thing for democratizing, um you know, knowledge and skill sets and expertise. Uh I often am am building things or designing things or coming up with things that I have, you know, in any other in any other version of the world I would never been able to go and do, but now I I I know just enough to be dangerous in in in those areas and and it helps me prototype, me generate new ideas, it helps me kind of work with colleagues faster cuz I can kind of like highlight the way I'm thinking about something where normally I wouldn't be able to like draw on paper what I'm coming up with, but but I just say, "Okay, this is the rendering that we're looking to do." Uh and so again, I think that's an incredible technology that's available to everybody for those that want to to adopt it and and lean in right now. >> What would you say to someone uh who's watching this, but they've also heard a lot of news about layoffs and about college graduates graduates not getting enough jobs because they're being replaced by AI? Uh what would you say to those people? >> Yeah, I think there is I think we're at a moment right now where uh and these these happen in history, you know, every couple decades or every, you know, 50 or 100 years where there's a major technology disruption or transformation and there's a lot of questions around, okay, where does that show up? Who are the the people that get enabled by that and they can do even more? Who are the people that may get displaced by that and what do they do next? So, we're in one of those periods where it's a it's a serious topic and a real conversation. Uh I do think that some of the negative kind of commentary and messaging out of, you know, the industry or even, you know, kind of political uh you know, institutions uh probably is over weighting the negative side and under weighting the positive side. Um for instance, I'll give you one example. So, there's a sort of death of the software engineer topic that that comes up. And that comes up because these AI models are really really good at code generation. So, they're really good at writing code and like you you look at them and you're like, "Oh my god, that's incredible how much code it just wrote." And it it wrote that code as as well as another engineer would have. And that's all totally true. But to get that code into production, to make sure that it's secure, to have it maintain an application on an ongoing basis that doesn't get hacked, to make sure it's integrated across all your other data systems and database and infrastructure, that still requires a tremendous amount of knowledge and expertise in the field broadly of coding and in software development. And so, the people that are going to be able to best leverage the technology are actually going to be software engineers using code agents to be able to generate vastly more, you know, code output than they would have been able to before. >> But that's that's today. Do you ever think about like in 5 years AI is going to be able to do that? Uh I don't know. Look at the market, strategize around some problem that the market is not solving yet, build a company, develop software, and that's it. >> a lot of data signal that isn't, you know, digitized in a in a format that the agent can go with. And there's a lot of ways the agent can get confused by accessing the wrong information or doing the wrong thing that you didn't intend. And so, for all of these reasons, it leaves humans in in some kind of supervisory capacity for what these agents need to go do. And so, does it need the same number of humans as we have today for the exact same workflow? No, not not usually. But are there a lot new workflows that businesses will now do because they have access to those agents? That's the sort of bet that I have. And so, the way I kind of think about it is is if you think about that 5-year out scenario. Let me let me paint a a slightly different one. I'm a small business, you know, pre-AI I was three people, we were selling something online. It was a good business. It sort of paid the salaries of these three people. But let's pretend I I had even more ambition and and I wanted to go after a bigger market. What do I do if I'm if I'm those three people? It's like I have to hire a sales team. I have to hire a marketing team. And a lot of people are just like, "That's a really high barrier to entry to grow my business, you know, meaningfully." Now, enter agents. And you're like, "Oh, I want an agent to go and generate this marketing campaign." Or I want this agent to go and build a better website that delivers a better experience for my customers. Well, what happens next if it works? Now, you have more customers. Now, you have more supply chain issues. Now, you have more customer kind of interaction challenges. You have new features they want you to build. Then all of a sudden, because you had agents go and get you some of the way to getting some of the work automated, my hunch is that same three-person business becomes five people or becomes 10 people because they now have automation that's augmenting the prior constraints and limitations that they had. I think that's going to happen as much if not more than the scenarios where you have a company that is sort of saying, "Okay, I have 2,000 engineers today. I'm going to have 1,500 in the future." I think it'll be a much more diffuse set of growth that happens through the economy. Some of it will be different roles, but I'm very optimistic that we're going to use this technology to grow more and do more as opposed to just replace. >> So, Aaron and I have been talking about building an AI-native business. One quick thing before we get to it. Look at your work right now. You maybe have one tab for emails, one tab for research, one tab for decks. They all run inside your company. The problem is none of them really know the others exist. Here's what that looks like for us. We record a podcast episode to turn it into LinkedIn post, somebody on my team opens the transcript, copies it to another tool, prompts a writing agent, then copies the output somewhere else. We do that for every single episode. The agents already exist. The bottleneck is that they can't pass work to each other. A person has to sit in the middle and move files around. That's what Auto Shift by Cisco is here to solve. They call it the Internet of Agents, an open infrastructure where your transcription agent can pass a file directly to your writing agent, which can pass the output to your scheduling agent, and there's no human in the middle. No manual copy-paste. These agents are coming from different vendors or might be built on different frameworks. Doesn't matter. Verify who they're talking to and move the work forward on their own. It runs on existing protocols like A2A and MCP, and it works with whatever you're already building. The open source project is called agency.org. It's a Linux Foundation project Outshift by Cisco was a co-founder with 80 plus members contributing to it today. If you're building with agents, or just watching where this is going, go to outshift.com. Discover the internet of agents and open interoperable internet for agent to agent collaboration. Now, let's get back to Aaron. >> And because we're doing more, we're basically consuming more, right? And solving more problems. So, we're becoming a more abundant world. >> More abundant, and you you you can't escape some ultimate constraint. There's always some constraint in the system. There's a new bottleneck that emerges. Um I have lots of things that I've tried to automate where at the end of the automation, the very next thing you have to do is a human has to do some work. It has to follow up with the customer, cuz I just can't fully automate that entire process. It has to, you know, update data in some system. It has to go into three meetings and and and kind of coordinate with some other kind of, you know, set of set of people. It has to go to the customer site and do some implementation. Um so, there's always constraints in the system. We just haven't identified all of the new ones that happen when agents kind of arise. Um there's a funny article uh about a week ago in the Financial Times where lawyers are now being inundated with questions from their clients, because their clients are going to AI agents and asking questions about legal issues, and they're drafting documents or whatever. But guess what? Like, if you were to go draft a contract right now, the very next thing I predict you would do is you'd go and send it to a lawyer and say, "Can you just make sure this is like going to like, you know, hold up in court?" >> Mhm. >> Because in the 3% chance it's not, which is basically maybe the hit rate of like like what an agent will get right or wrong, that's not worth the risk of of saving $500 of talking to that lawyer. >> Yeah, same with like financial advisors, right? You still want to run something through a human >> like not that interested in automating my tax my personal tax process. Like I am totally fine with with you know, the the one-time fee to just make sure that that is just like a clean process from from somebody that has like done this for 10 years or 20 years or 30 years and and there are just some parts of the economy which is naturally already where you know, dollars tend to flow where you're like I just want this done well. I I want my doctor to be really good. I want my lawyer to be really good. I want my tax advisor to be really good. I want them using AI because if they could somehow like review more of my data or look at more of my patient history or look at more of my my legal history, that would only be a net positive. But I want that person ultimately to have some degree of accountability that's on the line. These agents have no accountability. They're not on the line. They're not on the line for anything. They're going to disappear in in 2 seconds later and then I can't blame Claude. I can't blame Claude's weights. I can't sue Anthropic. Like like all of those things we have a we have a a you know, we have we have rules. We have laws. We have accountability for the rest of the economy. You don't in agents. And so somebody eventually needs to take on that that accountability. And this is more of like the the more like legal related issues. But there's still lots of things where you're like you want to look at your contractor in the eye and say you can you deliver this thing for me? Not in a I'm going to sue you, but just like I want to make sure that you can deliver on that brand campaign and it it's going to go super well. >> I even feel it with social media, right? I could totally generate a lot of posts with AI, but I just don't want to post AI-generated posts. I want a person who knows my taste and my tone of voice to look at them. Yes, maybe generate ideas with AI. >> There's another funny thing. This is like totally random and not and and this is probably more tractable in software over time, but but there's another funny thing which is I do think people will kind of get like they'll probably get prompt fatigued at some point, >> Mhm. >> which is like, man, I have to always prompt this agent the same way every single time just to make sure that it like works or whatever. Like humans don't require that. >> But then it can do cloud project with instructions. >> Yeah, sure, sure, sure. I know how to do it. And some people will get really, really optimized on that. But the nuance is there are some parts of your business where you just want the person to to be able to have that context and you just want like there's a lot of things I could probably automate if I like put my mind to it really, really hard. But like now I am basically doing the work of like five people. And it's just now I have to hold all of that context in my head as opposed to previously that context was in the head of of of those teammates. >> Yeah. >> And I at some point like my brain's going to explode. I'd rather those people hold on to that context. And it's sort of worth it. The value of the the of the of the thing being done well is worth it and worth paying for. Uh and so I again, I want that person to use agents. But I don't want to have to keep track of all their contacts either because I run into a limit. >> for the process and it's in your brain. >> I don't want to be responsible for our company's legal review process. I don't want to be responsible for the invoice process. I don't want to be responsible for the brand creation process. But actually that that's maybe a a kind of a really key point though that you just said, which is which is the more agents you deploy for yourself, the you you take on the role of the equivalent manager in another, you know, kind of organization, the human manager. You basically have to be responsible for whatever the output is. And so the more horizontal you go in what you're giving agents, the more functions you now have to >> brain explodes. >> Yeah, exactly, really. >> Yeah. >> And you see this in the valley. Like people are like totally tired. Like I have never met a founder right now or somebody working on a startup that's like, I'm getting great sleep and >> my 50 agents are running my startup and I'm just sleeping. >> Nobody's doing that. It's the exact opposite. They are managing the 50 agents and they are stressed out of their minds. >> I was talking to a lot of scientists, and they're the ones who tend to be most worried. I talked to Godfather of AI, somebody who has been studying AI for 15 years, and they're the ones painting the picture. >> or who or who? >> Yoshua Bengio. >> Oh, Yoshua. Yeah, yeah, yeah. >> Yoshua, he's like, we have 2 years. It's like, what are they >> He's the same who had 2 years though for probably probably 10 years. >> That's true. Like, but it's But what are they not getting? >> I listen, I have deep respect, obviously. Like, these are these are the best minds in in AI, and we have we are riding on their their work. So, so so obviously a tremendous amount of respect for for what, you know, what all of this kind of category people have contributed and their ideas. I don't know if you've interviewed like Yann LeCun. Okay, well, well, it'll it'll come, and I I kind of, you know, more in Yann's camp, which is there's still just a a fundamental limit to these systems. They they they have to be the work has to be reviewed. Any any error rate above 2%, you know, you still then need some accountability in the process, and everybody kind of says, well, humans are already doing that. It's like, yes, but back to the point, I can fire the human. And so and so there's some accountability at scale in the structure that that that that exists, where the agent just doesn't have any of that. And so somebody has to take on accountability for the output of that agent in your workflow at some point. >> Mhm. >> Because what you're not going to do is be fine when Bank of America says, we lost your money cuz the agent, you know, kind of like made the wrong investment you know, decision. And you're like, okay, but that's not why I hired you. >> Exactly. >> And so that part exists very broadly throughout our organizations and throughout the economy. And so I think I think what some people in the AI ecosystem that lean more to the sort of rapid takeoff, you know, kind of quick takeoff scenario is that they're think they're they're thinking that because the agent can do lots of stuff really well, that that sort of diffuses across the economy in a way that is sort of this destructive scenario. And and and I have I don't know if it's a benefit, but but it's it's certainly a a reality. Like I have the fun pragmatic reality of like I work with enterprises day in and day out. And these are enterprises outside of Silicon Valley. They're in the real world. They're the manufacturers of our products. They're the banks that we we we, you know, kind of bank with. They're the life sciences companies that develop drugs. And and what these really amazing researchers and thinkers don't do is they don't talk to those people who are actually implementing these systems. And so they see this incredible capability take off, but they don't realize the diffusion of that AI across our organizations is ultimately constrained by and bound by 30 other things that doesn't really relate to the super intelligence that's in that model. It relates to how do I implement this thing in a safe way with the right safeguards so it doesn't blow up my my my data structure. I I just think that the time scales are wrong. Uh the way that people imagine the AI being implemented in society is is is generally wrong. Now there's a one real risk that I agree with which is there is cybersecurity risks. There are risks of mis- or disinformation challenges. Those are very real. We need to work through those, but I'm I'm much less uh inclined to believe that this thing takes off, it replaces all white collar work, um and then we're in some really bad scenario on that. >> But we still hear all the news about layoffs happening due to AI. Do you think it's due to AI? >> Some of it is is definitely not due to AI. It's over-hiring during the kind of zero interest rate era, the COVID era. Um so there's some phenomenon that that that kind of relates to that. I would say that some of it it definitely could be related to AI. Like there are some organizations that they're like, "Listen, I had 3,000 people working in engineering before. My product road map is sort of not doesn't need to to triple in in sort of scale. It needs to grow by 50%. And so I think that if each engineer could be, you know, two x more productive and my road map only scales, you know, 50% then uh then then I think that there's some sort of savings there as a result of that. And then they might do a a layoff in that scenario. So, I I think that is real. It's not something that I can I I I can sort of gloss over. But, what I see from customers, and you can go online right now and and and I I I guarantee if you took five random companies in the Fortune 500 just as an example, take five random I guarantee that every one of those five companies is hiring software engineers right now. >> Yeah. >> And so, where are they hiring their software engineers? They're hiring them there's a there's a an interesting posting right now on on Eli Lilly's career website, which is a lab software automation engineer. This is a role to use AI to help sort of automate and in increase review of lab results and automate the lab process in in life sciences discovery. The kind of general idea of like AI is going to displace software jobs, that is not playing out empirically and I I might predict it will not play out ultimately. >> Hearing that story, the thing I keep coming back to is that it wasn't talent. It was the system around her. And it made me think about something very simple. Most people use Claude like a search engine. They type in a question, they get an answer. Most times they're not really satisfied with it and they close the tab. I did the same thing for a month and I was looking at people who were saying AI is changing their life and I'm like Then I spent one afternoon setting it up properly, uploaded a few files about how I think and how I work, and it completely changed. I wrote the whole process up step by step. You get it when you subscribe to my newsletter FutureProof. It's free. The link is in the description. >> I'm just saying. So, you think we're we're going to get more jobs in the next few years. When you're hiring now, how is it different from hiring 5 years ago? What are you looking for in a candidate? >> So, I think right now is a great great time to be going deeper technically. You don't have to like doesn't mean like you have to be vibe coding all the time and building entire products, but you should try and really understand what is the agent doing? How does it work? How does MCP work? How do CLIs work? How do skills work? And and getting really well versed in that. The people that are doing that will have a huge leg up in the next kind of 3 to 5 years because all of these companies will be hiring for people that can do that within their workflows. So we're definitely looking for people that whose technical acumen is is growing, whose AI sort of savviness and fluency is growing. Um you you want to be using these tools, you know, in your in your free time as much as possible so you again understand kind of how they're working and and what's going on. Um uh at the same time, I don't think a lot of the uh I I think it actually still matters that you have like some degree of domain expertise. Um like you're really good at marketing. You understand what customers want. You're really good at selling. Uh you're good at at product management and and interviewing customers and and assessing markets. Like those are the these like timeless sort of, you know, skills that transcend any kind of technology revolution. And so AI is just a way of of augmenting those domain skills. So there in in some respects any role that we're hiring for, marketing, sales, finance, engineering, etc., we need all of those domain skills, but also we need you to now be increasingly kind of AI fluent or a little bit more technical. >> Uh can you recommend top three apps that people should be using? >> Um you know, pro- pro- I mean, it probably won't be much of a surprise. I would I would download Codex. I'd download Claude. >> Even for non-technical people? >> Yeah, 100%. Well, especially uh partly is is because Codex is becoming more uh inclined uh toward knowledge work use cases. >> And what should they be doing with it? Like automate a process within their >> Automate a process. Give it just a crazy problem and see what happens. Like uh go do this research in this market. Um you know, wire up multiple MCP servers to data sources you have. Um so you understand kind of how does it work? Like how is it querying that that other system? How does it How does it accessing my email? Like like oh, scary. Oh. No, actually I understand it now. Like like get get a sense of how that all kind of is working together. So, I think just any one of the top AI tools for productivity, maybe for coding, uh is a good way to get started and and it'll already get you like 90% of the way there. >> So, Codex? >> Uh probably Claude co-work, Perplexity. Like these are some of just easy ones to just get started with and you'll have a good sense of kind of what the market looks like. >> Do you have any examples of workflows that you automated for yourself and you'll never go back to manual? >> The kind of things that I'll never do again is like I'll never do like market research in a traditional way. So, I'm often asking an agent to like go and analyze you know, a hundred different companies worth of trends or information. I'll just never do that again. Like I'll never go to Google and type each company name and and do the research. Like I'm going to have an agent go and and fan out, do all of that and then maybe I'll click like all the underlying sources and verify something or double-check something. So, lots of market analysis. I'll never I'll never open up code editor and like, you know, type code again. And I wasn't for the past, you know, many years anyway, but like the reverse is true, which is now I can actually like get prototypes built when I couldn't have before. Um so, anything coding related, even design is like you just go to ChatGPT and you're like, "Hey, I need this idea done. Could you just like make it like this?" And then it gets you like image the new image rendering model gets you like 75% of the way there. You hand that off to a real designer and then they kind of do the the full thing. Um so, there's a lot in the ideation, the creative process, the market analysis, customer research, all of those domains that um that I am heavily using AI for. >> Is there a certain way you structure memory? Uh like did you cuz I hear some people like upload personal constitution, like their principles of work. Is there anything like that that you've done? >> Um I'm less fancy on that front um and partly because I don't even know what I would write down um because I'm all over the place. So, uh, so I am I don't have a lot of things yet that I would I would know how to really document. It's more process-specific, in which case in which case it's back to the sort of re-prompting issue. I'm more just on the fly just giving it pretty clear instructions of exactly what to go do. So, like I feel like I'm a pretty good prompter. Like like >> So, every time it's a long, long prompt, right? >> Every time it's a long, long prompt, then I'll store those off in in various places. Um, uh, by virtue of Box, like we're like we store lots of data. So, I have lots of documents that have information in them that I'm using, you know, constantly. But, it's it's not as as awesome as like a sole file or a or a personal constitution. >> So, it's not like we're talking right now and 50 agents are replying to emails. >> It's not that. At the moment, um, if you were if you are if you get an email from me that's a that's a very, uh, that's a huge mistake in our system. So, uh, I am not emailing you right now via an agent. >> Got it. Got it. Okay. So, another thing >> Now, 5 years from now, could that be a process that gets automated somewhere? For sure. Like, just as we've always had, you know, automated email systems for sales reps or whatever. But, it probably won't be that that it would be like, "Oh, hey Aaron, like I I have a question about this thing." And then I'm going to have like an agent go do that. Partly because like that's actually just like the kind of context that informs me of what's going well, what's not working well. If I if I automated all of that, we wouldn't know the next thing in the business to go fix. >> So, you don't have an agent that's running your business, basically. Is there like cuz I talked to someone for no reason and they're like, "I share all my business decisions with my AI." >> Yeah. >> And then it looks at all the conversations I have with my team, and it gives me strategies. >> Um, I think, um, first of all, I think that's really cool that use case. I think more startups are doing that. I think if we were at a brand new company and it was like five employees, there's a very real chance I'd be doing more of those types of things because I would be like okay, I probably need to like build out our first marketing engine and I need to build out our sales engine and so I would be kind of I would be documenting way more of that. At our scale, you know, the really interesting important work is being done across the organization. So so that type of of work is more knowledge that like our head of brand design or our head of product design or our best brand design you know, designer needs to know or our product managers in each of the individual domains. Um the stuff that that I do is is sort of look across those areas and try and add, you know, extra nudges in the right direction and kind of course correct and um and you know, an agent could certainly help you know, give me advice for how to do that but but I I'm I'm still at the point in my life where I'm like I'm going to I'm going to see if my brain can do it. >> the founder energy, right? When you were talking to your team, you don't want your agent to be talking to your team. >> I think there's some I I think there's some um there there'll be some like spectacularly hilarious uh examples like that that probably over time sort of like >> seeing them like company data being deleted. >> Yeah, yeah, you're going to do that kind of stuff and and um and I I like um you know, the these agents are like I I I I could be proven wrong about this and maybe five years I'll be like yeah, I was totally wrong. And this is where Jan Jan LeCun I think would agree and some other some you have a you know, really interesting divide in the industry which is are these like probabilistic pattern recognition machines or are they truly able to to kind of go off in their own and think for themselves? And depending on kind of where you land on that continuum, then you have some big judgments that get made. So I kind of think about it as there's lots of business decisions I have to make or that that lots of people have to make where just it's a brand new net new event that happens and and I couldn't have documented what to do in that situation and I I maybe I could have if I I spent like a year writing down every single thing, but it's just it's just a new thing that happened. And so, if I if I try and imagine an agent running around, everybody's asking the agent questions, it's only going to be able to answer the thing that that previously I have in my in my sort of repertoire of answers. Many of the things I'm working on are the brand new net new things in the in the organization. So, me being an agent across the company would kind of be useless because because it would only have, you know, help with the things we already know. >> yeah, previous strategies. That that makes sense. >> Now, now now just just to share the opposite for 1 second. There's a lot of stuff. I'd say 80% of our corporate information is to be reused purposefully. Like you don't need people like making up a new answer to an HR policy. You don't need people making up a new answer to what is Box's security functionality and how should I position it to a customer. So, in those cases actually all of your enterprise information, which is what we do, you know, as a business, is like that enterprise information becomes valuable for agents because they can look at the documentation. They can look at the sales pitch. They can look at the the the meeting, you know, that that you that was recorded. That actually becomes very useful information for that kind of run rate 80% of your company's work. >> Yeah, but it's for specific work, not like a founder's strategy. Yeah. Uh and you said some next uh some next great companies are going to be founded in the next 3 years. And you gave a very specific timeline. Why 3 years? >> Um well, it could be 3 and 1/2 years. >> It's like not 10. It it looks like we have a very limited gap in the market where you can build something great because then it's going to be another like boring 10 years. >> Basic theory is like, you know, these market windows happen every every 10, 20, 30 years in technology. Uh the mainframe, the personal computer, the internet, the cloud {slash} mobile. So, there's already been kind of four of these eras. And if you look at the biggest companies, you know, in in in tech, they generally correspond with when these windows open. There's a couple ones a couple sort of examples that don't. Facebook sort of didn't correspond with any particular window. It was more of a a social change that occurred as opposed to a technological change. But most other other things, Google, Amazon, Microsoft, Apple, uh you know, sort of the the real turbocharging of IBM and and in that era and Intel and and so on, they kind of correspond to a new technology kind of found at the foundation level emerges and then and then you have this opening where a bunch of new companies kind of respond to that. In our era, it was it was Salesforce and Workday and and, you know, sort of enterprise software companies like Box that that sort of were able to capture that moment. And then in mobile, it was like Uber and and, you know, DoorDash and another set of companies. So, we're in a a window right now that has all of the makings of that, which is AI is now emerging. Companies need to Companies are going to want to apply this intelligence in various areas. And so, there're going to be a lot of applied AI companies that that bring that intelligence to to businesses, to society, to consumers in these applied use cases. And the only reason it's not like, you know, 10 years is because there's a lot of network effects remotes that get built. So, if you build one of these companies and you're capturing data from the customer and you're improving the feedback loop of the agent, that will just make your technology better and better over time. Whereby, at least on paper, that that product should become more sort of strengthen in its competitive advantage over time. So, that's why it's like, yeah, it's not like an infinitely long window because, you know, it's very hard to disrupt Walmart today because customers have been using it for decades. >> Yeah. >> Um and uh and so, you kind of want to be in one of those spots as these markets are are emerging. >> Are you seeing any gaps in the market where a startup should be working on right now? >> Um Uh still I mean tons. Uh but um Uh I I think there's still like I think you know everyone sort of knows the example of like Harvey right now for legal. >> Yeah. >> I think there's still lots of of job functions, industries that will have their Harvey. >> Mhm. >> Like I don't think we've heard the end of the the Harvey for X. I think there's going to be new infrastructure that gets built out because these agents are going to need new kinds of tools beneath them. Uh there's uh there's all this new interesting stuff around what when agents are doing work within software, they need more headless technology that they have access to. They might need payments. They might need And so, you know, Stripe and and their their this new company Tempo is is providing payments for agents. Well, now if an agent can pay money, then you can start to think through like well, what would the agent pay money for? And there might be new businesses that emerge >> Mhm. >> that the agent is now going to transact with. Like they're going to need data probably. >> Mhm. >> They're going to need infrastructure. Uh they're going to need to do tasks for you in the economy. Like there's lots of things that you can start to imagine that will become these new business models because of of what happens with with agents doing this work. >> With a whole new layer of active creatures in the market that are agents. >> Uh if you were starting today, can you walk me through a plan of like what would you do to find the right idea, to test it, and to make first money? >> My first thing would be some mix of like, you know, assuming we've got the most intelligent sort of system on the planet. So, that we have this incredible AI intelligence and just imagine that emerges. Then the question is where in the economy would that add the most amount of value? And then try and think through like like are there spots where like an incumbent isn't effectively responding to that? Um uh so, that'd be like one framework. Another framework would be like where in the economy is it hard to deploy agents because there's a lot of other kind of systems that that those agents need access to and that's usually where like there's lots of work to be done to get the agent to work within the environment. I'm pretty excited by a lot of these new um kind of professional services IT integration consulting firms that are emerging because uh when you go to the real world and you're like, "Oh, would you like to automate your work with, you know, Cohere or or Codex or any of these systems?" They're like, "Yeah, that'd be awesome." And then they show you their environment and it's like, "Ooh, like it's going to be a lot harder than you think." >> What markets? >> Anything. >> Anything? >> Everybody. Uh healthcare, law, life sciences, um bank I mean just every industry. Um because if your company's more than five years old, pre-AI, your data is all over the place. You've got 30 different systems you're working with. Your workflows aren't documented to the prior point. So, that's a lot of change management you needed to go do to implement agents. So, what does that spell? That spells opportunity for new services startups. That spells opportunity for the existing Accentures and Deloittes of the world. It's kind of like a a little bit of an up-for-grab market at the moment because of how much work there's going to be. >> How do you decide between like building versus >> maybe the only thing is like Mark Cuban has had this riff and I I fully agree with it. There's going to be like a lot of opportunity both for companies, but even just these will be roles that if you're like graduating right now, you might want to think about is like who's the person that shows up at the 10-person consulting firm in Minneapolis, just to like pick a non-Valley location. Who's the person that shows up that helps them take advantage of AI? >> Yeah. >> Because they don't have like a big IT department. They don't have a way to wire up their agentic workflows very easily. That's going to be like there's going to be tens of billions of dollars, hundreds of billions of dollars to get made between jobs and services firms in just just that over the next decade. >> Also, as an entrepreneur when I'm thinking about that, but what if Claude just makes the process really easy? I don't know, you just deploy an agent they build a specific agent who goes into your email, whatever you have, your box. >> Yeah. >> And creates the whole ecosystem for you. How do you think about that? Cuz these companies are getting more and more powerful, right? >> If I took the your exact scenario and I'm like, okay, um an agent's going to read through my entire email inbox in that in that scenario. And then it's going to access Salesforce. Then it's going to have some kind of like workflow that participates in. Like even me as a I've been building software for 25 years. And I I use every single tool that has ever been produced in AI. Obviously not literally, but but like pretty much. I don't feel comfortable implementing that workflow right now. >> Mhm. >> So the idea that that 10-person company is going to go and set that up and just because Claude became super powerful, I am skeptical that we ever get to that point. >> Interesting. >> Because Because the reason why I'm not comfortable with that is like I don't I have to think through the guardrails of like what happens if somebody emails me and then says, "Hey Aaron, I um you know, you know, you I need you to pull up this Salesforce record for me that that you told me you would you would look you know, look at. And um and you can send me that information." Well, if my agent has access to my email and my Salesforce, then the agent should by design answer that email question and go pull in the Salesforce record and then send it out. >> Mhm. >> That's like a non-starter. >> Mhm. >> Like you can't just like take any untrusted email coming in and then have the agent >> distribute information. >> distribute information that you that your tool has access to. So So even me trying to think through how to implement whatever your scenario was just now, I would have a hard time thinking through how do I set the right guardrails? How do I have the right like alert mechanisms to me? How do I have the right sort of human in the loop of like should I review all of the emails before they go out and have an interface to do that? Should I have some escalation mechanism that like pings me on my cell if like I need to look at something? How does the How does the the the person on the other end of that email inbox not How do they How do they get to me as the real person and like get you know, how do they escape the agentic loop that they're in? >> Mhm. >> There's like 30 questions that I even have thinking through whatever that workflow is. So, it's not a matter of Claude is so powerful. It's a matter of like how the systems talk to each other, the safety mechanisms of those systems. How do you divine the How do you define the actual workflow so it it gets done in a kind of safe and reliable way? That's the work that a technical person generally needs to go do. >> Okay, so but that's like a big shift from, you know, being manual to getting automated as a company. What about niches where like we see Figma's stock go down when Claude releases the design feature, right? And if somebody's working on that type of feature and they're afraid, you know, with the next Claude upgrade it's going to be gone. >> Well, that's >> That's a different issue. So, so so that that's a different category altogether, which is which is, you know, how much will Claude or these AI models eat into the business models of different industries or different, you know, providers. I think that's more of something where you just have to be very thoughtful right now to not just build anything. You have to build things that like what what are you building where even as AI agent progress continues, no matter what no matter how much it continues, it could be infinitely powerful, there's still some other thing that that agent is going to need to do. It's going to need to put its data somewhere. It's going to need to incorporate into a workflow. It's going to need a a human to take the information and put it into the real world. Over time, more and more value will start to look like things where where again, like it's a well-governed process that has lots of security or compliance needs. You have to trust the underlying system. You know, probably just like quick personal productivity tools maybe will be less relevant. At the same time, like in in the Figma example, I think Claude design is is actually very very cool, very powerful. I played with it a bunch and it like, you know, generates, you know, amazing designs. >> Yeah. >> But at the same time, I still want I still want our design team kind of going and doing the last mile of work. And right now, they're doing that last mile still in Figma. Even these things are not as binary as as as I think maybe the like Wall Street for instance would suggest. Um uh so so that's that's why it's still kind of a, you know, we're we're in a pretty dynamic period right now. >> Yeah, and it's also because I feel like the stock reflects what we're thinking about the next 2 years. And because this is evolving so so fast, sometimes as an entrepreneur, I'm like asking myself, okay, I'm building these apps. Why don't a language learner just go into ChatGPT and like build the app with five code in an app themselves? >> I mean, I I think it's a it's a question that every every entrepreneur should should have a very big whiteboard that like thinks through various game theory events that could happen and where will your value get compressed and and where will it where will it not? And um and it's it's hard to, you know, in any kind of generic way have a perfect answer uh cuz it is a very busy, complicated time. Um I but but but again, kind of ironically in like the more macro sense, I think the more that AI is sort of doing in these kind of automated things, you're just going to see new constraints begin to emerge. Like like, you know, a lot of people like have like healthcare classically as this example. And then Jeff Hinton, you know, had uh I I don't have the perfect quote, but I think he you know, he felt like radiology would reduce as an example because, you know, AI will get really good at at looking through radiology images and yeah. >> And and the self-driving >> And self-driving, yeah. >> And now we still have radiologists driving to work every day. >> Oh, sure. Yeah. >> I think it was that. >> Oh, yeah. Yeah. So so um what also happens is these other things that that occur, right? So like, we might have like AI that gets the radiologists 90% of the way there to like look at the right thing or get some suggestions. At the exact same time, what what that's meaning is we're doing vastly more imaging. We're we're doing vastly more scanning. Way more people now can go do it. >> More accessible. >> And it's more accessible. And so so actually now the demands on that role end up increasing as a result of that. So there's a lot of parts of the market where where actually AI facilitates lowering the barrier to doing that work. And by lowering the barrier to doing that work, more people participate in it. And as more people participate in it, a new constraint gets kind of backed up that now real people need to go and kind of, you know, get involved in or or go and work on. >> Mhm. Yeah, I feel like the more I play with AI agents, I do realize that I need a person at the beginning of the process at the end of the process. So I still end up having more people. >> Yeah, and and and again, it's just like a question of like how many roles do you want to play within your company or your team? >> Yeah. >> Like do you want to play designer, developer, marketer, strategist, >> No, really. >> No, you're just like, ah, like I just want to sleep at some point. >> Yeah. >> So yeah, the solo entrepreneur is already used to that cuz was, you know, we or they have been doing that forever. Like when I did uh startups before Box and I was a solo founder, like man, I you had to do 10 things and you're like you're so so tired. And if I could have ever hired somebody to go do half of those things, I would have. So does could AI allow us to get these companies to a little bit more scale to the point where then you can hire that next person? That's that's more of where where I think this would go. >> And do you think it's the best time to start a company now? >> I'm kind of a stickler for this one key point, which is it's only the best time to start a company if you've got a a great idea. >> Mhm. >> Uh I think that great ideas can exist in in any kind of period of time, but I'm not in the camp of just like everybody should start companies um uh because because it's a I mean, you know, it's like it's like really hard work. It's it's extremely stressful. You're working like mad. I don't think that people should feel pressured into into starting something because because it's the it's one of these windows. When I say it's one of these windows, it's it's just to reinforce the point that like this This moment where the best ideas probably will get built. That doesn't mean that you should start one. >> It just means >> you have to rush yourself to starting now. >> Yeah, cuz if you if you rush yourself to starting a bad idea with one of these moments, you're no better off. So, I would say it's a it's a it's a good moment. You have an incredible amount of leverage. That also comes with more competition. >> Yeah. >> Um because because basically by lowering the barrier of of getting ideas out in the market, what do you get? You get more competing ideas. If you get more competing ideas, that's more noise that customers have to deal with. So, interestingly and back to the job thing again, interestingly, it's not so much like now the idea getting the idea out there that's going to matter. It's going to be like, man, do you have like do you have somebody talking to customers? Do you have Are you doing sales? Are you doing marketing? And so there's a new constraint, which is like the constraint isn't code generation. The code this constraint is is are you in front of customers enough? And and are you marketing enough? Which is a new set of dollars. >> Yeah. Uh if you could become 19 again today and start over, would you exchange that uh to what you've currently built? Like to start over again, would you do that? >> Am I starting in 2026 or back in 2005? >> as a 19-year-old. Would you do that or would you just stay stay put with what you've done? >> Oh, I see. I see. Well, um I'm the most excited we've ever been on what we're doing now. So, I would certainly pursue what we're currently doing because part of it is historical, which is, you know, we we've earned the trust of 120,000 customers, which is a good launchpad for the next set of things we want to go do. Um so, and then I I just love the the kind of things that we get to do with customers, we get to work with every industry and every size company, and we get to help uh you know, space launches and medical discoveries and blockbuster films get produced. So, like I'm very excited by what what our platform does with agents. >> Yeah. >> Um at the same time, I have lots of friends that are doing, you know, companies that I'm like, oh, that's a really cool idea right now." And it looks very exciting and and uh and so I I'm just in a period of of like I'm impressed and excited by lots of stuff um while also being again incredibly stressed constantly. >> Any jobs that are going to disappear in the next 5 years? >> I think there's going to be work that gets compressed. And and then and then I think you're going to take those people and often re- re-purpose for for again more of the agent manager escalation path or proactive versions of that work. A very kind of clearly obvious one is is and this is something that that companies always you know, try to sort of automate to some degree. Like like if you're emailing a company you're saying like, "I need you to reset my password." That is probably not going to be a person >> Like customer support, right? >> Cus- customer support, but even that customer support is this funny one which is which is like we we we think about it as a monolithic thing cuz we call it customer support. There's tiers of customer support. There's like the first line of customer support that we will most certainly automate which is >> Change password. >> Change password. And I don't mean to like you know, over minimize that that thing, but like there's a lot of tasks like that which is like, "I need to download this thing. I can't log in. I have this issue." whatever. That we're going to fully automate. But there's a lot of customer support which is like, "I need you to get on this call with me and look at my my specific problem in my computer and why this thing isn't working." >> Yeah. >> And we just have no way to automate that. >> Yeah. >> Like maybe we'll automate like the next line of the of the set of questions. But you can't ever you can't get to the the final thing. I had a I had a friend have a problem with Box 2 weeks ago. He just sent me some screenshots and there's a 0.0 chance that he would be able to have asked the question with an agent. >> So it had to be it had to be you. >> It had Well, in this case it actually it had to be a senior product manager. I had to get the senior product manager for the person, but he couldn't have talked to a chatbot and answered the question. >> But what about like bookkeepers? Something that >> I had I had an issue with with my Mac last week. I spent 10, 20, 30 minutes on AI trying to diagnose it. Never never worked. Had to call IT. They had to come in diagnose it. Couldn't replace that. >> Oh, okay. What about jobs like bookkeepers or >> You know, so some of these jobs like again, they've they've been on the path already to to how do you already automate as much of that away as possible. And so agents kind of are just another sort of layer in that. But again, I'm you know, sort of the same answer. There's still always an escalation path because there's always the exception. There's always the the weird anomaly that occurred. Um, and you can't have like the thing that you can't do is I can't moonlight as a bookkeeper. I can't moonlight as a lawyer. So so at some point there's still is this final path in the escalation which is I may have been able to automate 90% but you just you still have that one part. >> had this I had this >> legal question two months ago and I asked every AI agent the the the same legal question. And it and every single one basically gave me the same answer. And then I called a lawyer and they basically gave me a more a much more uh kind of contextualized answer >> Mhm. >> because they could decide how much risk I needed I wanted to take on or not take on. >> Knowing your personal situation, right? >> personal situation. They also know the fact pattern of like like how does the industry tend to think about this one thing? And all the AI agents were were giving the the the the sort of mean answer of that of that particular topic which is in this case it's it's like it's the more conservative answer. It's the thing that it should be trained against because it it can't give you the it's not going to give you the more liberal risky answer. >> Mhm. >> But but when you talk to a lawyer they're like, well actually yeah, this situation won't actually occur >> Yeah. >> because because of XYZ fact patterns. And so those are the kind of things where like you then are like I want somebody that has seen 20 years of this stuff. >> Yeah. I don't want a model that was just like looking at Reddit and and deciding, you know, how to use it how to use that information. >> Okay, my two last questions. You have a 6-year-old, right? >> Uh almost seven, but yes. >> Okay. >> And four and and seven and seven and a half months. >> Oh, congratulations. Okay, are they going to go to college? >> I shoot, maybe I shouldn't have leaned in so much to to the question. Uh I Okay, I'm here here's the one problem with me. As a B2B enterprise software person, I am like boringly pragmatic. And so I just think change happens more slowly. And and it's funny because I have this like weird duality which is like I adopt every tool. Like I was I was wearing Google Glass like like in week two. I buy every VR headset. Like I I lean into every one of these tools cuz I'm just like I'm excited as a personal user. I love technology breakthroughs. It's amazing. And then I'm like a and then I go in the real world and I'm just like, man, that whole system over 300 500 years that we've we've built up is like is that really going to change because of this one variable? So on the college thing I I I struggle because I'm like on one hand from first principles, it doesn't have to exist. My 7-year-old almost 7-year-old is already way smarter than I was because he can every question that comes to his mind, we're like looking up the answer right away. Whereas like I don't have like a perfect memory of being seven, but like like I didn't have like an instant resource for like every question. But like he wants to know like how fast a peregrine falcon can fly, we get the answer. He wants to know like how big the the Atlantic Ocean is, we get the answer. Like like and so he's just like a sponge for like unlimited information. On one hand you're like, wow, that could probably replace like a lot of the traditional sort of, you know, ways that we think about these institutions but then you're on the other hand you're like well I you know what what is college other than another four years of high school but with more a little bit more vocational kind of kind of orientation a network of people that that you want to be with and learn with and make make connections with a kind of transition period into the real world because you know you're still kind of young at 18 so like so that then I'm like a pragmatist I'm like does that really change in in this like super intelligence world or is the curriculum just changing and the format maybe changes um but like does everybody want to just be at home with their parents talking to an AI bot um like no so so it's like I have these other like you know kind of sort of counter pressures now things that should and must change like out of out of social like society level is like man can we have college cost a fifth of what it does cuz it's insane like like should you really go into debt for 20 years cuz you went to medical school or you went to you know get X degree like that's that's incredibly insane so like should we use abundance to to bring the cost down and and try and do that as much possible absolutely >> Mhm. >> So there's some things that need to change about college but does the very concept change I I I always I always struggle with that one. >> Yeah same same here I feel like as a society we're really slow to just change dramatically when it comes to foundations and college is one of them. Okay last question advice for entrepreneurs who are starting today. >> I I would say just like back to that earlier point lean into the tools like learn the technology see what's what's what's possible with it make sure you're riding the tailwind of of what's happening in technology you don't want to be you know kind of hitting a headwind where you're kind of going against the grain of the AI you want to be like riding the the AI wave out um which can mean a a number of things it might mean do things that actually in a world of AI become more important because people don't want AI to do that thing so it's like it's like this counter intuitive like riding the AI wave might mean do a live events business. Like >> That's what a lot of people are doing. >> Yeah, but like like like do something where we will appreciate this other thing in the economy because AI is sort of so abundant or or um AI makes getting healthcare questions answered so quickly, so you should probably be doing hospitals. Like because now more people are going to be you know go like like actually needing real >> Wellness clinics. >> Wellness clinics. Like like so the just like so sometimes it's a technology thing that you do. Sometimes it's a thing that the technology sort of is is related to an underlying broader societal trend that will become more important as well. Build one of these, you know, consulting businesses that helps deploy the AI. Um I I just think there's going to be like um build a childcare service cuz we're all sort of our brains are exploding and we need help with kids. There's all this kind of stuff that that that is going to need to exist. >> Yeah. Thank you so much. I I like your positivity, especially after talking to some scientists. Thank you. >> Thank you. Thank you so much.