SAP is a $300 billion company today. So how is it that SaaS and cloud was going to be incredibly disruptive to the incumbents and yet those companies are all bigger today? It's because it actually was ultimately a sustaining innovation for the incumbent. The incumbent was fine to move to the cloud once they were able to.
I think there's other AI strategies that are not 'we have to build a foundation model.' And so you have to nuance to that. And then the other AI strategies, how are you doing that smarter or less smart? And if you're like, 'Oh, this won't really work.' So, for example, a lot of startups, I think, go, 'Oh, well, I'm going to be building better coding agents.' And you're like, 'Well, okay. There's a lot of work in building good coding agents. What's your particular differentiation that might be harder?'
Aaron, I assume the startups in the room shouldn't presume stupidity on the part of Box.
Yeah. For all other incumbents, that's great advice. Yes, exactly.
You should assume we're going to do the absolute best thing.
Well, good afternoon everybody. I'm Ben Kestnoka from Village Global. Welcome to another Village Global Luminary event featuring our chairman Reed Hoffman, Aaron Levie from Box. We're going to have an hour conversation. I'll ask a bunch of questions. I'll also call on you if you have questions and some of you have submitted questions in advance so we'll try to keep this interactive. Let's get right into it. Reed, you often advise founders to think about what's true and contrarian.
The first question though is going to be for Aaron along those lines.
You're like, 'This is my chance.'
Aaron, you have, speaking of contrarian and maybe true, you have a thesis that agents will actually make SaaS more valuable not less. Wall Street right now has a certain narrative.
It's being tested every day on that theory right now. [laughter] But I started that theory like when the market was like 40% higher than where it is right now. So we'll see if this one's true.
Give us the bull case on SaaS companies in the world in the era of AI and agents.
Yeah. No, my confidence is being tested very aggressively. So the point when I put that out there was more of a philosophical architectural point as opposed to like all of SaaS should be worth more than it is now. And so I think the market does need to decipher which software reduces in value and which one grows in value. And so I think the market has decided not to do that and it sort of lumped everything together. So forgetting what's happened in the past maybe three months in software, I think the general point I was making was there's a lot of software where if you had a hundred times more users on the software than people previously were using it or a thousand times, the sort of value of the guardrails, the workflows, the business logic, the security of how that workflow worked and the data that was involved in it would go up because you actually have even more risk of something going wrong because now agents are roaming through these systems and pounding them at again 100x more than people ever did. And so in some areas the value of that business logic now is actually multiplied because we've actually had a lot of categories of software that were artificially constrained by just the fact that the total number of users you sold into a company was constrained by the headcount of that company or that particular department for that use case. And so there's a lot of categories of software where like everybody knows Harvey at this point but the market size of the contract management market previously was like the total spend on all software ever in time in the contract space is a couple billion dollars a year. That's like how big the TAM is of contract software. Well, now in a world of agents where those agents are doing a substantial portion of the work in the software, the value of that category could be tens of billions of dollars. There's sort of not a particular limit that is out there because now you just deploy these agents 24/7 and they're operating within this system. So I think the challenge is that we haven't as an industry and Wall Street hasn't been able to figure out who looks like that where the agents doing work within the system actually grows the value of the platform versus there's going to be definitely software where the agent is now doing so much more of the value proposition of what the software used to do that it compresses the value and there will be cases where that happens and those will be the areas where the platforms are less valuable and I think right now everybody's just trying to figure out which side they end up in.
So, which are the categories that are growing and which are the categories or use cases that are shrinking?
But what are they for you?
Okay. I was like, 'Yeah, that's a good summary of what I said. I probably should have said it that fast actually.' I mean, there's like, have you seen these market maps of software categories? Like the logos become this small. There's like 9,000 SaaS products in the world. Not even the world, probably per category. So I think it's hard to say. I think my general rule of thumb would be which categories of software again would have had much more demand for that software if the population of users in the company could have just, there was a larger population, like which things structurally add a lot of value but they get underutilized. Those things should go up over time and which things have already reached kind of saturation and they probably have overbuilt their functionality and the functionality kind of gets underutilized because it's just not that, there's too many features, that stuff probably goes down in value.
So take an example like HR tech, is that...?
I think you can't even do it at a category level like that. I think honestly it's like subcategories beneath that because CRM as an example, there's going to be lots of cases where I would spend more on our CRM system if it generated the leads, if it followed up with the leads, if it could connect to LinkedIn and roam around and get all of the updated contact information. We would spend more on CRM systems if it did that whether that was Salesforce or a startup. So in that case I would make the argument that actually a CRM system in a go-to-market world you'll spend more money on because in a world of agents they can be doing so much more work for you to go and actually grow your business. So that's actually an area where I'd say more dollars will go into the category over time. Again, does the incumbent get it? Does the startup get it? That's up to the execution of the teams, but I don't think that less value goes into the CRM or go-to-market technology stack because of agents. I think more value goes in there.
I agree with everything Aaron said, but I think there is one additional question that is partially top of mind for people, which is how much of the additional agentic value goes to the SaaS company, startup or existing, and how much it goes to the AI model agent companies.
Yep. So that's another part of the variable. And so part of it, I completely agree with you that some SaaS will actually become more valuable and some SaaS will become less valuable. But part of it's also how intrinsic is the value of that SaaS ecosystem not as it is today because the rebuild of the tech stack, all deep features, lots of too many features or too few features or whatever, how that plays into essential value that is not just commodified by the agents.
Yes. And I think this is why this is like we could do a four-hour, and we need a whiteboard because you got three to four moving pieces in this and Reed brought up probably the most important one which is does the value of what used to happen in the software go to some horizontal system of intelligence of which there's only three or four major players that could probably capture that versus how much really is actually business logic required where the agent equally needs so much that's in that business logic layer and either the agent is native to that platform or the agent at the horizontal layer is paying some kind of tax to use that system. Then depending on where you are in that continuum will mean more or less value to again that SaaS layer.
And how do you, Reed, think about the startup versus incumbent framework in general? It's obviously top of mind for all the founders here. What's the general framework or specific to agents and AI but where startups should focus that incumbents might be laggards?
Well, the general thing as you know since I'm mostly an entrepreneur and investor is presume incompetence and stupidity from the incumbents. So go for it, right? Because presuming intelligence is not as good of an idea. So for example if you say, 'Oh, what I'm going to do is I'm going to go after Google for desktop or mobile search.' You're like, 'Okay, well what's your particular plan in terms of how to do that.' But generally speaking, incumbents will be slow at getting out of their own way. They have innovator's dilemma problems. They have, none of them are currently built in AI native ways. Question about how the marketplace is changing because of it. So, generally speaking, the default is go for it. Now that being said, if what is not a particularly good thing, and this refers a little bit to one of the things that Aaron was referring to earlier, is that if you said, 'Well, I have to build a beyond the foundation model multi-billion dollar hamster wheel,' then you would better be intelligently on that wheel. I think there's other AI strategies that are not 'we have to build a foundation model.' And so you have to nuance to that. And then the other AI strategies, how are you doing that smarter or less smart? And if you're like, 'Oh, this won't really work.' So, for example, a lot of startups, I think, go, 'Oh, well, I'm going to be building better coding agents.' And you're like, 'Well, okay. There's a lot of work in building good coding agents. What's your particular differentiation that might be harder?' Aaron, I assume the startups in the room shouldn't presume stupidity on the part of Box.
Yeah. For all other incumbents, that's great advice. Yes. Exactly.
You should assume we're going to do the absolute best thing.
But where do you, because you recently said you're trying to, it's a startup reinvention moment for Box. It's kind of a cultural reset for the company. In what ways do you think incumbents are vulnerable to startup disruption either in general or in this moment of AI agents?
Yeah, I think the only, because you gave me a 'but' and so I'll give you a 'but.' So I actually, at least in the way we internally think, we actually force ourselves to think, imagine if the incumbents are actually really good then how would you win? And because if you can develop that strategy then you'll win if they're bad also. But if they're really good, you're going to need to know what is your angle of attack. Now, the part where maybe it meets the middle is, you don't want to get yourself into this loop and people do, which is you over assume what the incumbent will do and then that kind of creates paralysis and then you just freeze and you don't go after the market. So I would still go after the market, but you should actually assume the incumbents are eventually going to figure this out. And again, if they don't, then you're going to be bigger than you thought. And if they are, then you'll at least have had an opening. And I think the only tricky thing is I think innovator's dilemma is very instructive for this whole situation which is you have to identify where is there something that the incumbent doesn't want to have happen from a business model standpoint versus where are they actually totally fine for it to happen. And they actually it's more attractive on the other end of that journey. And that's the classic innovator's dilemma. It tells us exactly what will play out in every given market because incumbents don't want to do things that will make them less money. But they are totally happy to do things that make them more money. And so you have to figure out which parts of AI are actually going to make the incumbent less money or change their business model so dramatically that it's going to be hard for them to execute that or AI has this in some areas or it's so technically different that they just don't have the data structure. They don't have the ability to kind of bring agents into their workflows in a way that is actually going to work. And I would sort of study that for each market. I think the easiest example of all time that's already playing out in the market and so we already have the perfect control for this is what has happened in things like customer support software. You look at the incumbent models you have to sell to the seats of the number of agents, the customer support agents and then conversely you have obviously startup Sierra, Decagon etc that actually just say, 'Okay, we're going to do agents, we don't have a seat model, we have a different way of pricing this.' It's obviously very disruptive to the traditional seat model. That is like innovator's dilemma 101. The incumbent has a hard time getting their business model to transition effectively. The disruptor has a very clear path to going and disrupting that. So that one's really easy and we're going to see that in a number of categories.
There are going to be spaces though that are a little bit tricky, which is like it still requires the same seat to go do the work with the agent and the existing incumbent the seat already contains the data or the workflow that the agent already is going to need to kind of tap into. And so, thus the incumbent actually has this sort of natural data advantage in some cases. And then it's really up to the incumbent to be able to execute flawlessly from a technology standpoint, a business model standpoint etc. The thing I go back to that again people need to be prepared for and then I'll give really good news for startups probably is if you, let's snap the line in 2003 let's say right before SaaS kind of was like fully, we figured out the template for SaaS and then now you look 20 years later at the market the vast majority of on-prem software companies are bigger today than they were back then. At scale, Intuit didn't get disrupted. Autodesk didn't get disrupted. Synopsys didn't get disrupted. Siebel sort of got disrupted, but they got acquired by Oracle. So, we don't really know how that would have played out. So Oracle is 2x, 3x the size of Salesforce. Like all of our predictions were wrong. SAP is a $300 billion company today. So, how is it that SaaS and cloud was going to be incredibly disruptive to the incumbents and yet those companies are all bigger today? It's because it actually was ultimately a sustaining innovation for the incumbent. The incumbent was fine to move to the cloud once they were able to. They got the business model wired up for subscriptions. It actually was a better business model because you had recurring revenue from your customers as opposed to the perpetual license model was this kind of crazy treadmill you had to be on. So, it was actually a better business model. So the question is which parts of the software stack actually have a better business model because of agents versus again which ones are going to be under pressure because of what agents are now going to do. That's kind of the question that you have to do category by category. There's no way to do it in any kind of broad brush strokes.
Who knew SAP was a $300 billion company. That is mind-blowing.
And Oracle's 450. So these are gigantic companies and they're bigger than any SaaS company that's ever been produced. And so the agents will have some aspects of that with one big asterisk that there is a new kind of disruption because as you move away from the UI layer into the agent layer using APIs that will change the value proposition. So anyway, here's what I'd say for startups. The cool thing is there's actually so much white space for startups, but it's not going to look like just do the old software product but with agents. It's going to be go and do a completely different kind of software category that previously you couldn't have done in software. And that's what agents actually let you do is go after markets where there was no software incumbent because and this is the lesson of all the great disruptions in the past two decades is go build a company in an area that there wasn't software for. There was no Uber before Uber. There was no Airbnb before Airbnb. There was no incumbent that had those business models. Those are where the biggest startups will emerge. So in agents, you would go after the parts of the market where there is no incumbent piece of software because you previously never had digitized that workflow. So this is where you're, I mean we never had that much software for doing audits of companies. We didn't have that much software for being able to do legal work. You could go after all the services areas that now agents for the first time ever can go and automate. So there's going to be gigantic companies that get built because there's so much new software opportunity but I don't think it'll be that you go head-on with Workday and you're like we're just agents plus Workday.
Yeah. Fantastic. And so we'll do a few more topics and questions on AI then we want to shift to broader entrepreneurship themes. I've heard both of you say, Reed a few months ago and Aaron just a few minutes ago on the green room, that we've been a bit surprised at how slow in some sense AI deployments have been in enterprises. Like if we rolled the clock back to the day ChatGPT was launched and then said how many companies two or three years later would have massive AI implementations we would have thought the number is a lot higher than it is. It's incredible the number of companies you talk to today that do very very little internally. So I'd like to hear from both of you about where we are in the roll out of AI especially in the enterprise and especially as it relates to predictions like Aaron you've said on multiple occasions you think there'll be a thousand times more agents inside of a company than human employees. Like when does this actually happen? So maybe Reed, we can start with you. How would you reflect and you were at OpenAI or helping the OpenAI team at the very beginning, knew the launch was coming of ChatGPT.
Yeah. Well, you didn't tell anybody that.
Like you're a board member.
Why has it been slower in some sense? I mean, so revolutionary on the one hand, but still so many companies haven't really adopted these tools at all. What does that diffusion look like? And then Aaron, I'd love to hear from you.
So one of the things that I think is interesting is all startups are basically, well like if a startup isn't this they're just basically already dead, is essentially coding agent accelerated and is using it. When you go to the smart competent incumbent companies, it's very funny because you can look at like a map of what the token count is across the developers and what you'll find is that's lumpy. There's a few people like usually a much smaller than 50% that's using it a lot and there's a bunch of people who are using it very little. The using a lot is not always indicative of being much more productive. Because for example currently in all the use of coding agents because one of the challenges you have within incumbent companies is working on legacy code bases is not yet a superpower within these agents. So I've talked to multiple very smart teams not just Microsoft, Google and other but many others and they're working on it still. So because it's the question of the cold start of I've got a new idea and I'm gonna start writing something and I got some spectra great. I have 300 million lines of code and I'm doing the maintenance and the upgrade of it much much more uneven. Now I don't think that's going to stay that way. I think everyone's working on that. But I think that's part of the reason why even when you look at the incumbent tech companies with coding development agents which are the current leading edge of the competence of AI acceleration that's still, and it's happening very fast but that's still in process versus started with Opus 4.6 as kind of a driver. And so part of it then gets to what are the things that enterprises care about? They care about reliability of work. They care about security. They care about provisioning. And for example, there really isn't an LLM right now that isn't, if it's generally speaking open to the world, like just about anyone with a brain can figure out how to do a prompt injection attack. It's not perfectly defensive. So the enterprises kind of go, 'Right, we kind of want to use them but we want to make sure it doesn't open up attack services, it's not bleeding information, etc.' So you have all of that as well and then you add in group coordination, a bunch of things. So I do think we will see the fastest revolution of work processes in human history but we in the valley tend to go, 'It's going to happen next month' and it's like that's not how human organizations especially at scale work because of all this. And one of the things that I strongly, and this is a room I don't need to do that in, but when I'm talking to a lot of people I'm like, 'You should be using these coding agents yourself some to see what it is not just because of the magic part of it which is amazing but you start going, oh right it's still got these jagged edges where sometimes the coding agent goes, I'm just going to brute force my way through this problem and it's like cycling and cycling and it's on the wrong approach. It's like no no no you got to not this one this one.'
You can appreciate the magic and also still the stunning shortcomings by actually using the product.
Yes. So anyway so I think we will see a major acceleration. I do think there will be a bunch of changes. I think most of the things that are like for example, 'Oh, hiring college interns and so forth it's an AI problem.' This is actually a problem of instability of a business environment and correction from COVID, right? So, of course, companies when they go, 'We don't know what we're doing in hiring,' they don't hire interns, right? Because it's like we don't know what to do. It needs a stable basis. And stable basis is not just a question around AI. There are some questions there, but it's like, 'Oh, are we going to have growing number of wars? What is that going to mean? Tariffs, what is that going to mean? Etc.' So Aaron, I want to hear from you in a sec, but just we'll take some questions on AI and agents in a minute. So if you have a question, think about it. You can raise your hand. But Aaron, next 3 to 5 years, what are you seeing across all the companies you're working with? How fast do you see the development of these agents?
Yeah, I would kind of just underscore what Reed said. I think this is going to take a lot longer than we in the valley have maybe, at least what the sound bites represent. And if I can just layer in two to three other things. So I don't know if everybody saw the Dario Amodei podcast recently. Dario actually kind of nailed one key point on this that relates to this, but I'm going to expand it a little bit. And he kind of hypothesized that coding agents are working really well because the codebase has a lot of the context that the agent needs to be effective. Maybe a 300 million line codebase a lot harder but let's just say starting in a smaller environment. But if you just think, so we all like our measure of AI progress mostly in our industry is related to AI coding progress and we do this thing where we try and extrapolate that to other areas of knowledge work but let's just do a ledger for a second. So in coding first of all these models have had a two to three year head start on being trained specifically for coding so all the RL has been done for coding related tasks. It's a type of work that essentially is just text in text out. So the thing that it needs to produce is just text. It doesn't need to be particularly pretty text. It just needs to be text. The outcome is effectively verifiable. Like did the feature work? Did the product run? So instantly within about 10 seconds like did the thing actually do what we wanted. The user base prompting the agents are the most technical users on the planet and they have no problem changing their tool chain literally in a week if there's new alpha that has happened. So you take kind of four or five aspects of that work and then this other key point, there's so much green field stuff that you do where you could just prompt something for the first time and it's generate something very useful if you wanted. So you have four or five kind of huge benefits in the world of engineering. Now go to the rest of the world. Let's just take like law or you're doing accounting or payrolls or something like that or you're an investment banker, you're a consultant. Your context, 70% of it is probably completely undigitized. Like it's in meetings, it's in conversations, it's the lunch you just went to with the client. Your data is not in a singular codebase. It's fragmented across three or five or 10 different places of where your existing context is. There's a lot of legacy infrastructure. I didn't even, I forgot that one in engineering. We've all standardized on GitHub over the past 10 years. Like I mean 20 years like it's like 100% of all work is done in GitHub or GitLab or whatever. But like in the real world of work there's 40 different places where the knowledge is that any kind of agent would ever need to be able to automate the workflow. The users are not technical. They don't want to do any of the things that we're all doing here. They don't want to install OpenClaw. They literally, if they do they don't know to do it on a fresh Mac mini.
Yeah. Sure. I mean like and if they did then they've given up all of their information to the internet and then you have this other problem which is the work isn't verifiable until it hits real life in most of these jobs. So if you go and generate a contract your ability to verify whether that contract was good doesn't really happen until either somebody else redlined it or there's litigation. So you could have like weeks before you know that the thing that was output is actually a thing that is going to work and be sturdy enough for you to put in production in whatever that type of work is. So just compare that ledger and we've got like every perfect condition for AI coding to take off and rapidly self-improve and us be able to adopt these systems and the best practices to the rest of knowledge work. Imagine going to a knowledge worker and tell them about agents.mmd. Like you just wouldn't be able to explain that. This is the delta that we're going to have. So it's going to take years and years for this to play out. Now there's one interesting X factor which is you'll have AI native companies in a bunch of these categories that get built. I'm actually pretty excited by that. Like the AI native law firm will totally outrun the traditional law firm. They'll do 5x the amount of legal work. The AI native ad agency will produce campaigns for five times the number of clients. So, that'll be this really interesting pressure.
And that's the entrepreneurial opportunity.
Yeah. So, either you're building the tools for them or you're building the agency themselves, that'll put some really compelling pressure in the market. But even just again a little bit of pouring some water on reality here even when we've had that kind of disruption we had this in the digital non-digital world in the kind of early 2010s where again Uber etc. If even if you look at today we've had Robinhood, we've had all these things JP Morgan is a 700, 800 billion dollar company so like so even when you will have pressure coming from these new startups the incumbents again eventually sort of rewire their practices to support that. So I think it's going to take a while. This is actually why I think there's so much opportunity is because the opportunity is which companies can actually bridge these incredible breakthroughs that we see in Silicon Valley with real life work environments. One thing that is not going to be surprising is the headcount on professional services whether you call it FTE or something else like this is a multi-tens of thousands of person problem as we enter the real world you have to rewire your workflows in an enterprise if you're really going to get the full power of agents. This is why Deloitte is still going to exist. This is why Accenture is still going to exist.
But they'll be smaller right in terms of headcount.
I don't know, I mean who knows.
Do you think in five years Deloitte will employ the same number of humans? I can argue both sides.
By the way I can argue more same and less.
Yes. Exactly. Exactly. Exactly. Yeah.
Well, I think we can all argue for less. What's the argument for more?
It's Jevons paradox. It's actually in fact at a certain price point given the work is so much more amplified there's a huge amount more demand for it.
Almost all of the mistakes presume that there's fixed demand. And there's many things. There's not fixed demand. And if you imagine, let's say every company in the economy we agree needs to will eventually need agents running around doing all this work. So who's in a really good position to help companies do that? Is it each company one by one themselves learning how to do that or is it five or 10 companies that have built best practices that can go and diffuse them across the economy? This is what systems integrators do. They get really good. They're doing their 10,000th Salesforce implementation. So when you are in line, you get the benefit of that of being 10,001. So when they go and re-engineer your banking workflow with agents, you're going to want to hire the one that just did it across the street as opposed to you have to go learn all the mistakes from scratch by yourself. So there's going to be so much business for all the layers of professional services in this environment.
Questions on AI or agents? Is Lavana here by the way? Okay, let's go down here then.
My question is about how do we size this market opportunity because you said something Aaron at the very beginning which is it's about we can think about how much are companies spending on contractors for a given work and I don't think that's right because imagine that I'm a farm owner back in the day and I'm spending a lot of money on farmers and now I get to buy a tractor that spend is going to go way down and I'm not going to go look at that money and say, okay, how else do I spend it? It's more about I think the type of work that is going to become more or less valuable. And I agree with you that there's going to be new types of work that no one was spending money on at all because it was just too out there, too expensive, and now all of a sudden they're going to spend money on. And so my question is, could you shed light on how do we think about these different types of work? Because I also agree with you that it's not per-vertical. It's not going to be sales this is going to happen, marketing that's going to happen. It's going to be different types of work and I don't think we have a good thesis or philosophy yet on how do we categorize the different kinds of work that get commoditized that OpenAI or Anthropic completely take it over versus other kinds that don't and companies actually spend a lot of money on.
Yeah. I mean this is the tens of trillions of dollar question because it represents all of work. I mean, back to Reed introduced Jevons paradox into this. I think a lot of it will come down to which work actually did we have an artificial sort of demand constraint on or supply constraint on where there's actually way more demand. And I think it's why you kind of look category by category and you have to kind of look at what parts of the economy are either don't change or going to change in different ways. So as an example we are hiring more sales reps than ever before because there's nothing you can really automate other than like we can automate the pitch that they generate, we can automate the outbound email scripts, but if we can be in more conversations with customers, then we will hire more sales reps because that's just even more deals we want to go close. So, there's almost nothing we've been able to we haven't automated away the job. We've just automated various tasks that make them more productive. So, there's lots of parts of the economy that will actually have that aspect.
And Aaron, do you think in enterprise sales it will still be as human to human?
Okay. Yeah. So when agents, my agent is not going to talk to your agent to do a deal,
It'll talk to them as much as it can to then eventually just get us back into a meeting to discuss the same set of things we were discussing before because at the end of the day, you still want to look at somebody and you want to say, you're going to be responsible for making this project successful. I can't hold an agent accountable to anything. I can't me turning off an agent and pulling the plug, it doesn't care. So like I can't that is not accountability in any real sense. I have to look at somebody and say your job is on the line for this thing to be delivered. That's going to be humans.
And Reed, do you buy that?
It's certainly true for longer than most technologists think, right? Precisely because business is built on trust. It's built on a trust of you're actually going to deliver on the product. I'm going to look good for it. I have a person to call in order to make it happen. That is years. It may eventually as hey we have this trust protocol that's actually really working for agents and that works and that can build up but that won't be like this year. Right.
Right. And it will take some time.
Now the other thing I would add
But Reed sorry before you transition but like it's interesting you think about people in high school today and you think about the jobs market 20 years from now like in 20 or 30 years are we hiring human sales reps in the way we are today? I actually would guess we're not hiring the way that we are today. But I would also I could look this is all looking through a glass darkly. We might be hiring sales reps. They look somewhat different. They'll be differently empowered.
Like the agents done 99% of it. Two humans show up to shake hands and say good.
Well, but also maybe like, hey, this is the things that I want to make sure that in the agent dialogue you're really backing up, etc. Let me Okay, I got off a customer call two hours ago and they can't decide if they should go with Anthropic or OpenAI for something. Okay, so what do you think they're going to do next? They're going to have a ton of meetings with both of those vendors because there's no like the agent could have made any recommendation on how to determine that. They're still going to then go and be like, 'Well, shoot, like I did hear this one thing from this one peer and I think they deployed it this one way and but this other direction that's all can only be adjudicated by people talking and working through that.'
Yeah. So I think the human in the loop component will be much longer maybe transformed than people think. Now the other thing about market sizing as a general principle is it is exactly right markets change like for example part of a bunch of people were like silly about Uber because they were like oh it's just limos right and market changes. Similarly, when I invested in Airbnb, I probably could have spent the week calling everyone who was actually doing a stay that week, right? If I had spent my whole week doing it because it was so few. But what you're trying to do is you try to go, okay, where do I think there will be natural strong economic trends and that there's something that you can build a useful asset in there? And that's actually part of the entrepreneurial bet. Now, sometimes it's a pure substitution effect. I have a much cheaper product that's much better and I'm going to replace a lot of it. And as per Aaron's earlier examples, frequently what happens is like for example, people have been going PCs totally dead. It's all going to be mobile. The growth is in mobile. PCs continue. And I don't just mean Microsoft PCs. I mean, desktops and so what it is is you got this big area where all the growth's going, but a bunch of things also still continue. So that's part of what you're doing when you're trying to figure out like what happens here.
One more question on AI or agents from anyone. Let's go. We'll do if you can make him quick, we'll also get some me done, but let's do a co-founder of many of you who built agents for recruiting.
Is it unreasonable or a waste of time to think that there will come a point where the intelligence of models of GPT 10 or Opus 9 is so high that it makes no sense for Anthropic or OpenAI to make that publicly available because they can generate they can capture all the value theoretically or practically a lot of it and not give it away to us. Is that unreasonable or like who cares or
