How AI Is Changing Enterprise
49:28 Share

Wait a second, if we could use AI to automate more, we can build more. If we could build more, we could lower the cost of things. If we can lower the cost of things, then we can actually lift up anybody's lifestyle right now. I think that we're in the middle of the revolution, and the revolution does not have to be a black mirror. It could be something that is driven by Jevin's paradox, driven by abundance for everyone, and that's certainly the timeline we want to be on, so... That's the future I'm betting on. ♪

Welcome back to another episode of The Light Cone. I'm Gary. This is Jared, Harj, and Diana. We're partners at YC, and collectively we've funded companies worth hundreds of billions of dollars. And today we have a really awesome guest, Aaron Levy of Box. Oh, thank you.

Love that intro. Aaron, you're one of the best product CEOs out there. Public company as well. That's what I write in my Wikipedia. Yeah, that's how I classify you. We're in the middle of the AI revolution, so how are you feeling? Oh, pretty good. It's a good time to be in software right now. So yeah, feeling pretty good today. Something we've been speaking about for a while, which I think we probably agree on, is that the chat GBT wrapper...

Was like a bad meme and that actually there's like lots of value and always has been in building apps on top of these Foundation model companies in fact the opposite might be true. Yeah, which is gonna be more valuable in ten years time Yeah, so so it's interesting So there's there's a like probably 2% truth in the meme and then 90% not not truth And so I mean PG, you know with the sort of wedge theory is like actually you do want something that is sort of a simple product that finds a little wedge and then you expand from there and

In the early days of cloud, if you were to be building software that let you manage documents and data, you would have been like, well, that's a wrapper on Amazon. And it was a total misunderstanding of the entire scale of software you have to build to make the storage bucket be useful in a particular application.

So on the wrapper conversation, the exact same thing is true, which is how much software do you need around the workflow and the proprietary sort of business logic and the data that the customer brings? That's actually the value, not just like what is the set of tokens that are coming out. Where it's a little bit true why startups should at least think a couple steps ahead

is you probably don't want to be something that just ChachiBT would incorporate. So it's less that the model will incorporate your value proposition. It's more if there is, you know, if the model provider also has a consumer scale application, like you don't want to be right in the way of something that ChachiBT will just fold in directly into its functionality. So in that case, I think you have to be sensitive to being kind of, you know, a quote unquote wrapper. How do you separate out stuff that's going to get incorporated into the model from stuff that won't? Because I feel like the hard part

with that often is we don't know what the next models are going to be capable of. And there's this general sense of, oh, well, like anything could theoretically be incorporated into the next model if it's powerful and approaches some version of generalized intelligence. Yeah. I mean, I look at all this through just the B2B lens, which I know that then probably you just lost half the podcast listeners. B2B SaaS is cool these days, actually. On the B2B front, to me, it's a simpler question because

And enterprise doesn't want a model. It wants an outcome. It wants an outcome of customer support, conversations being answered or healthcare transcription going into an EHR system or an automated workflow of reading documents and contracts and plugging that into a contract workflow. So the model getting more intelligence is actually usually a better thing for anybody building software in those use cases because then you're doing less

in terms of hacking your way into the model because it's sort of insufficient at solving that particular problem. But what the customer actually wants to buy is like, I need software that will plug into my ERP system, that will plug into my support system, that will power the workflow, that lets the customer do a password reset. Like that's actually what the customer wants to buy. And what the model is doing is sort of really abstracted from the ultimate customer value proposition. So I think as long as you're building...

software that really can deliver that full outcome to the customer. And two years ago, the initial wave of these use cases started to emerge.

And I think the companies that will do best are ones that realize that you need to abstract the model away as much as possible from the ultimate value proposition. And then you just incorporate all the model updates as quickly as possible for your customer. And again, they just buy the outcome of customer support. But it's just getting better and better every time there's a model improvement. I wonder maybe one good analogy is when you were building...

for your customers. They didn't care what was the underlying database or cloud or what was the networking gear or all the hard drives that were running. It was all about the end user experience at the software level. And the analogy to today is the end users of B2B AI workflows don't care whatever model is or how it does it, but that it ultimately does the workflow. Yes.

Yeah, I think that is absolutely the conclusion. You'll often get some idiosyncrasies in different organizations where they do care, okay, where is your data center hosted or what's your infrastructure provider? But that's a small minority.

In AI, I think we're going to go through a temporary period where you do see differences in the models for anybody that has a discerning set of skills on this front. So you can see this in Cursor with Anthropic. People like that combination and they can sense the differences of the output.

If you fast forward five years, I think you'll see a convergence of basically models and intelligence to the point where you wouldn't really distinguish the quality levels that much for 90% of business use cases. MARK MANDEL: It's definitely interesting to see how especially developers have developed different preferences for different models. I remember at our AI retreat a few weeks ago,

Anthropic and Claude has also emerged as the preferred LLM to orchestrate your agents. If you have multiple agents and you want the LLM to intelligently call the right ones, people seem to prefer Claude for that.

What do you think is going to happen to the model companies themselves in this world, though? Probably everybody needs to update their understanding of what a model company is, just in general. I actually think there's very few model companies. There are sort of AI companies that have frontier model labs, but increasingly they're selling software.

to either consumers or businesses. I don't even know who I would consider to be a pure play model company at this point. Anthropic, if you look at their software revenue, it's effectively an API business for enterprises. I'm sure they have some large scale consumer kind of cloud business, but you're really paying for the security, the compliance, the governance, the privacy, the uptime, the SLAs, talking to somebody that manages your account. And the

The model just continues to sort of switch out underneath all that. If you look at OpenAI's revenue, anything that's been kind of leaked publicly, it's very clearly a software company at this point that has AI models that power its underlying software. Google, obviously, you know, is just GCP. And then Meta doesn't need to monetize it.

because they can just open source it. So maybe XAI is almost the closest thing to now a model company, but that will show up in Grok, et cetera. So what you probably wouldn't want to do right now is start a pure play model company expecting you're going to have licensing revenue by just selling your model to people to go and use.

if you don't have enough other kind of surrounding value proposition that, again, lets you get incorporated into enterprises or has a large scale consumer application where you have some degree of kind of traffic that keeps people within your ecosystem.

I think it'd be very bad to be just a pure play model company at this moment, just because you have enough different business models that have emerged now in AI where it's going to be pretty hard if your pure play business model is just pure AI tokens, because you always have meta tokens.

that will always create a counterbalance by just open sourcing, you know, a frontier model that kind of wipes you out. And now DeepSeek. And now DeepSeek. And so, and that's what's amazing is like, now that like you can basically guarantee meta has to do anything DeepSeek does because it obviously has to stay in the game on the open source front. And so we always have, there's enough dynamism in this industry that basically ensures to Gary's opening point that like the cost of intelligence is going to go to zero. Like it's just like absolutely guaranteed. So extrapolating,

What does it mean for startups as intelligence becomes commodity, basically? Well, the good news is we kind of know the playbook on this with one X factor, which is like AGI and what is the ultimate kind of... X factor. Yeah, exactly. A little bit of an X factor of wiping out all of relevant business models if we don't have money in the future. But if you put that to the side, I think these companies need to look like software companies. And it's sort of back to basics, which is...

We used to have an API into a database. We had an API into storage. We had an API into compute. Now we have an API into intelligence. That intelligence is... It should be... Basically, the cost of that intelligence will go down to the cost of the bare metal. So whatever the underlying cost of the GPU is, that's what you're going to pay. With a little bit of margin from a hyperscaler. But the cost of the actual tokens will converge at zero. And so then it's all about, do you build software that...

takes this complicated technology and delivers it to customers to solve real world problems. And so you guys have talked a lot about vertical AI. I think that's a massive play. I think there's going to be certainly a whole layer of AI software that kind of stitches together different AI systems. So you have horizontal plays.

you have vertical plays. I think the idea of every single industry and every job function probably will have some degree of new startups and agents that they get built out for those, uh, those, those like slots. Um, I don't know if you guys have like a whiteboard of every industry and every job, but like, but like you can just basically play, you know, bingo on that. And then, and then until it's fully filled in, like there's probably still opportunity left with an AI. We figured out the first wave of SAS, uh, of, of how to do this. Um,

YC was obviously a big part of a number of major category killers in SaaS. And I think we'll see the same playbook happen in AI. One of the many interesting things about DeepSeq was specifically the first open source reasoning model. In the short term, do you think there's new ideas in the enterprise in particular that are going to come out? Because now we have...

of open source reasoning models? So what we've seen is, so we do a number of benchmarking kind of exercise internally for the reasoning models versus, you know, kind of let's say non-reasoning models and some things that they're actually better at, some things they're weirdly worse at. And I don't think we've even discovered why they're worse at these things. Maybe they overthink a problem. In general, I would kind of just argue that anything that directionally improves

you will see B2B use cases. You'll see the value of those use cases go up because you can begin to reasonably chain together more agents working together. You can get more agentic workflows happening. Anytime we can get the intelligence factor to go up,

I can now reliably introduce this for a more important business process. And so in the enterprise, you can almost think about it as there's some probably either two by two or chart. I don't know if anybody's made it, but like kind of like mission criticality of the workflow, AI level of intelligence. And kind of there's an element of like, you can't introduce it to closed banking, like banking systems, sort of end of day data.

you know, yet, because it's not particularly deterministic. It's, you know, we don't sort of, you know, know all the answers that it's going to give, but it could write a summary for, you know, a new product launch at a bank.

or it could help answer, you know, banking, you know, product questions if you're a consumer. So there's a continuum there. And as we get every degree of intelligence going up, we get more use cases that we can now implement this for. And then there's another axis, which is like how many of those use cases, you know, can you string together to complete like the full

the full workflow of that business process. And that's yet another, I think, access that, that, that we're early in, but like, you know, I was, I was in New York a couple of weeks ago meeting lots of banks and, and, you know, just generally what you think of like the New York industries and enterprise and,

And I would say we're like 10% of the way into the adoption of, let's say, just like general chat. So like assistance and like 1% of the way into adoption of anything we would all call agents. Like, and that's maybe even inflated numbers. When you're in the room with like the banks, the Fortune 500s, all the people making their decisions in enterprise, do they really have zero interest in the underlying models? Like is Deep Seat coming out just like a total nothing burger for them and they just care about...

what you're pitching them and offering them? Or do they have interest in the actual underlying tech? There are people like us and the people listening at every company on the planet. And so those people care.

By the time you get to the, let's say, line of business, so I'm the head of wealth management at a bank, they don't care. But the CTO cares and the head of AI cares and the IT folks that dabble and they're hanging out on Hacker News, those people, they care because they're using Cursor and they're seeing the differences between Anthropic and OpenAI tokens within that.

But when it goes to talking to an executive in the business or the daily end user, they have no interest. It's all a foreign language to them. And I think that will remain the way forever. I think more of the expectation is that, again, these things will converge. And what's amazing about AI is because of the I don't want to call these the models fungible, but directionally fungible.

Because they're somewhat fungible, you will see characteristics that we've seen in other areas of compute, which is any best-in-class model eventually has to match the price of anybody that beats them in pricing because you can just switch to a slightly inferior model. Even if it's inferior by 1%, the risk is that you could switch to that and find it acceptable for 80% of your use cases, which then by definition means

whoever is at the frontier actually still has to match the pricing of somebody just slightly worse than them because they could just, you know, the end user actually doesn't care. And their business could evaporate if they don't do that, which means ironically, you could actually stay on one of the providers. You could just pick a provider and you kind of know that you're,

your tokens will become as cheap as the second or third, you know, cheapest option because that first, you know, whatever that first provider is, the marginal, the next marginal customer doesn't have to choose them. They could choose the second or third player. So you, which eventually then, you know, you run that experiment out, you know, 10 years, you converge on basically the same pricing, which is what we've seen, which is like,

The difference in pricing of storage buckets between the top three or four hyperscalers are not so different to drive business model, fundamentally different business models in the software stack, similar to compute, et cetera. So really you're making a choice based on

some other set of reasons, like how much data do I have in the system? What are my workflows that I've built in the system? And then I think that, again, the price of the AI eventually becomes largely the same. Actually, I think what you're saying applies to what we're seeing for startups. I've done a number of office hours with AI startups that are selling to enterprises. And a particular story is this company that scaled to 12 million revenue within a year. Yeah.

They actually switched models underneath a number of times. And the end customers, which are these big enterprises, didn't care. What they cared was that ultimately the contract and expectations was that just get the workflow done with this level of accuracy, done. And as the cost per token has been cheaper, they actually have been increasing their margins. I think when they started launching, I think their margins were around like...

30%. The next cycle of iterations middle of last year with all the model releases got to 60% and I think now they're at like 80%. That sounds like file storage. Yeah, exactly. We love that. So that's kind of what happened. I have a complete example for this company that has done office hours on

Exactly. I mean, your analogy is actually quite literal. So right now, publicly, we have 81% gross margin. If you had said, let's say when we started the company in '05, that a business that was perceived to be storing data would have 80% margin, you'd be like, no, that doesn't make any sense because-- Just commoditizing. People are just paying for the storage. And it's like, no, we have nearly 1,000 engineers a box.

1% to 2% of them are working on what we would call file storage. So what is everybody else doing? We're building software that is the abstraction layer of compute and storage and databases to produce workflow and data governance and automation and insights on data. So the storage is now a small fraction of what we overall provide. So similar to tokens as

you know, the tokens go down as a ratio of what you're really delivering as value is that software stack. So I think, you know, probably one way to think about it as a heuristic is like how much software is necessary on top of the output of the tokens for your value proposition to work successfully for the customer. The less software there is there, probably the higher risk you'd have to either more competition or commoditization, the more software there is.

where the tokens are just one contained component of the full thing, then you're probably in a position where you can build a moat, you can get stickier, you can then solve more of that customer problem. But you might get to the point where the customer pays for a discrete outcome.

This is like one of the big open questions. I'm sure you're seeing it in the batches, but like, what is your pricing model? Do you do you pay, you know, let's say you're a startup that does AI lead generation. Do you pay per lead? Right. That's that's a fairly obvious kind of thing that you'd expect. And then and then basically that company now could be, you know,

Or do you even pay for like qualified lead that the customer sort of says is actually successful? So like there's a whole continuum of like, I pay for a successful outcome, I pay for any outcome, or I pay for the underlying kind of resource utilization, which we also see in like coding startups

you know, where it's like, okay, I want to buy some unit of compute measurement that goes into useful work. But the cool thing is we're going to see, you know, a mix of all new business models and software that we haven't seen before. This would be one of the biggest changes in the advice I give to the startups during the batch is I feel like

it was really hardwired into me when a startup comes in, we're getting like a pilot or they're going to pay us as we go or something like that to say, that's not a real customer. Like you have to go back and you have to get them to sign like an annual contract and like lock in revenue. Otherwise you're just kind of wasting your time with someone who's sort of one foot in, one foot out. But over the last year in particular, when I look at the most successful companies, essentially often they're like replacing a BPO or some sort of service like that. The customer actually wants usage-based services

And the revenues just keep going up and up and up. So I'm no longer like, oh, you have to sign my annual contract. Yeah. Yeah. I mean, for a lot of areas where we're sort of, there's a direct relationship between the thing that the company is selling and then let's say labor on the other end, you do need a lot, you need long-term contracts. So you have to hire people. There's a lot of infrastructure you have to build out. The great thing about AI is it's entirely elastic.

So we're going to have to imagine a new world where all of a sudden I have elastic resources

for things that otherwise used to be very operationally intensive. So one day I can just say, hey, I want 10,000 leads generated, go run AI to go do that. And in a traditional way of doing that, that might take months of hiring and staffing and building out teams. Now it's a week later, you're off to the races, you're generating those leads. You can kind of go through any analogy of the business.

And that becomes possible. So totally different relationship between the company and getting outcomes and output. Totally different relationship also then for the software provider and what their business model is with that customer. Aaron, can we go back to your trip to New York City? Yes.

Or just any of my trips? You spend a lot of time talking to senior execs at Fortune 500 companies about their technology and AI strategy, probably more than almost anyone in the world. And I'm just really curious what those people are thinking about AI. Are they focused on it? What do they think it's going to mean for their business? Are they building AI initiatives internally? Are they trying to buy products from other people?

What's happening? Yeah. I mean, definitely all of the above. Did you see this thing that went viral like two weeks ago? David Solomon, CEO of Goldman Sachs. The S1 prep. Yeah, the S1 prep. He basically had this quote at an AI event at Cisco. They're doing projects internally where AI is writing an S1 in like 10 minutes or something. And it used to be a team of six people that work on that, et cetera. The exact same quote, just parallel universe quote,

15 years ago, let's say in the early days of cloud, just as a useful kind of comparison, I'll probably keep coming back to the cloud thing.

it probably would have been a banking CEO saying, we'll never go to the cloud. Like, we don't trust the cloud. And now the exact opposite. Which happened, right? Didn't Jamie Dimon say that? Jamie Dimon did. I think he sort of evolved his thinking, but you had that kind of across the board. And these were these famous moments. It's like, we'll never be a cloud company. We don't trust it. We don't want to move. I mean, Amazon, it's a bookstore. Like, that was the refrain. And it made sense. I mean, I even said that when I saw S3. Like, the bookstore is going to power? Like, what? So,

So think about how different of a world it is that the CEO of Goldman Sachs is basically saying, like, this is now what's possible. He wasn't saying that in like a we shouldn't do it way. He was saying that in like a we need to open our eyes up to all of the potential use cases that AI is going to have in the business. And he was saying it as a way that they're leaning in and starting to try out all these use cases. So for that to happen for a top five bank in the world...

At this early in the cycle, you know, it only goes kind of it only goes more aggressive from there because he's in the most regulated of all all the businesses in the in the most important financial market in the world. And he's already leaning in. Is that because he's like a particular early adopter or are you just seeing this across the board?

Yeah, I mean, also he's like a DJ. So maybe he's hanging out with like EDM people that are just like really into AI music. Okay, so 10 years ago, we'd host these dinners, 15 CIOs from different industries, heavily financial services, let's say if it's New York,

And, and it's like the, the, you know, we're going to try cloud for this one tiny part of our business. We don't really think we can scale. The idea of being cloud first is like, it would be like totally in, you know, an anomaly. Like you would never, like a bank would never say they're cloud first 10 or 15 years ago. Uh,

Today, it is sort of like we're trying this in as many areas as possible. Everybody's still insanely early because you've got privacy councils, compliance councils, regulatory bodies that have to look at everything. But everybody understands how big of a tidal wave this is going to be in their business.

on a few dimensions. One, they know that the workforce is going to completely change. I think there's a recognition. This kind of hit me maybe a year ago in some of these conversations. There's a recognition that basically if you're entering the workforce today, you've had now a couple of years of Chachapiti, of a college. They're native. They're native. It's an AI native era of the workforce. And we could make some jokes about it maybe two years ago, like, oh my God, the writing essays, I can't believe it. But like,

I basically almost don't search the internet anymore. Like I only know how to use AI to find information. And guess what? Like I'm, I'm, I find 10 times more, more information as a result of that. So, so actually many respects, the AI native, you know, people will be smarter on the topics that they decide to go in on than, than, than the prior generation of,

of whatever that is. So what does that mean? Like these Fortune 500 companies are changing how they hire? So I don't necessarily know how they hire, but it'll become clear that if you don't have AI, if you're not an AI first bank or media company. So what's your AI strategy? Literally, what is your AI strategy? Because why would it to the customers or internally? The enterprise will basically realize that they can't actually hire the next generation. You can't go from all of a sudden I have this AI native way of operating in college or high school to going to a company that makes me use, you know,

the equivalent of a fax machine, you know, level of technology. It's just like they won't be able to hire people. And then their competition will have more output. They'll do more investment banking deals. They'll onboard customers faster. They'll get better financial advice to their clients than the company that sort of doesn't do that. And so everybody's sort of realizing this is actually a competitive issue.

Cloud didn't really have that. Cloud was purely an efficiency story. It was like, yeah, I don't really want to have to be building data centers as much. Elastic capacity sounds pretty good. I want to be able to test this new product faster. It was not like my customer is going to experience a different thing

about the output of my company, whether I'm cloud or not cloud. Now, I think we all believe that there is a difference there. It was not tangible to the buying side that cloud was going to make your company five times more competitive in a way that AI is very clearly, I think, right.

resonating that actually your company's competition will now be at risk if you're not in. Yeah, that's a really interesting point. Because with cloud early on, a lot of the benefits were actually to the startups who pushed it forward because they just wanted an easy way to get set up and not have to deal with hosting. But you're saying it might be the opposite where with AI productivity tools, startups don't really need them. But actually, maybe the counterpoint, to me what this sounds like is actually way more opportunities for startups. It's just all the office hours that we're doing

I think it's the fastest I've seen B2B SaaS AI companies get enterprise deals. And I think you painted a very good picture of what's the vibe shift. Yeah, if I can just slip in one funny anecdote. I went to this banking conference eight or nine years ago and I did this little keynote at this banking conference. And it was all about like, we have to be cloud first and enterprises have to modernize how they operate with the cloud.

And I think I've never been, I think I've never bored an audience more than that keynote. I remember getting off the stage and nobody caring. It was just like, 'cause it was like, why are you talking about backend infrastructure? Nobody cares. Yeah, whatever, we could be in the cloud, we could not be in the cloud. At our levels of budget, if we're spending $500 million or a billion dollars a year in IT, who cares if we save 100 million 'cause some of it is elastic or not? It's not that big of a deal.

if you were to do the same thing, but an AI first to non AI first enterprise, people would be like, Oh shit. Like actually I probably can't run my business anymore. Not AI first. Cause you would just show people like, like, do you know the productivity gains of somebody using, let's say cursor? It's like, you will be blown out of the water competitively. If you do not know how to build an AI first company right now. And yeah. Has box invested in any internal AI tools to speed up how you run the company? Yeah. So, uh,

A few categories. So one, we've been rolling out AI on the coding side and we're trying everything. Like internal tools for your engineers to... Yeah. So basically just how do we make the engineering kind of more productive? And that's sort of obviously one of the biggest X factors of our business is can we output more code that is obviously useful and aligned to our product roadmap? So we will be fully AI first in terms of how we develop AI.

know this year it's uh it's the it's sort of the big year for all the change management on that we are you know incrementally rolling out ai for different customer facing things just again can we solve the customer ticket problem can we can we improve the rate of response and then as an ai provider a lot of the knowledge management use cases we sort of now do ourselves with ai so you know if an employee has a as an hr question or benefits question we have a feature that lets you talk to all your hr data um and all the internal knowledge management and so what became

This breakthrough for us was all of a sudden all the things that were inside your documents before become useful for now just interrogating with questions as opposed to reading documents. And so there's a lot of just embedded productivity that we focus on from that standpoint. What things do you think they're going to do internally and what things do you think they're going to buy solutions for?

I would basically guess that, again, it kind of looks pretty similar to kind of historical ways of thinking about this. I think Jeff Moore created this, and if I'm getting it wrong again, please, Jeff Moore, like pop into the comments. We'll just come in the right name. And the idea was sort of context is all this stuff that is sort of like you have to do, it's necessary, but it's not going to make your business successful.

better than your competitor. And so that's your HR system. That's your ERP system. Like you have to have it. It's important. It has to be done extremely well. But like your version of the HR system is

you know, is not going to be radically different than, than checking any of the box. You have to check the box, but it's gotta be a good box and whatnot. And then there's core, which is, this is like literally your value proposition. Like, like you sell, you know, wealth management to, you know, you sell wealth management services to people. And, and that, that is something you, you own. If your thing looks exactly like your competitors, then you have no, there's no sort of, you know, profit, you know, margin that you'd be able to, you know, you wouldn't be able to get reasonable profits. You have to have something that's unique.

And I think companies need to really understand which category is which, partly because if you put the core in the context, then you'll probably be at a long-term disadvantage. And if you put the context in the core, then you're wasting a tremendous amount of time and energy.

And and this is why I really enjoy that the the Klarna announcements or whatever. It's fun to read, but I also think it's sort of misunderstanding the context versus core core thing. Like you don't need to build your own HR system. I'm glad they're doing it. I think it's provocative. Like it's fun to see different approaches. That's the people going. You know, it gets our like we're like our juices are flowing, but like the average bank is just like not it's like not a priority for them to like reinvent their HR system.

Now, so so so I think I think whatever whatever is that for every industry, you know, in life sciences, you probably really want to understand, like, how are you doing drug development? And you should probably have a very strong team working on that problem because that's something that sounds like it's going to be IP for you.

But the automation of the clinical trial process, that's probably context for you because everybody's going to want to be doing that as quickly as possible. It probably doesn't involve a tremendous amount of proprietary data. And then obviously your CRM system, your HR system, and so on. So it actually sounds like

They're going to buy a lot of things externally because most functions in a business are actually contacts. Yes. Yeah. I think most of the way AI will show up to a knowledge worker in 2030 will be from what we would have thought of as an ISV 10 years prior. What's an ISV for the people watching? Yeah, it's just basically a software.

you know, provider. The CRM system will still come from Salesforce or X competitor. It won't be that they built a homegrown AI generated, you know, you know, CRM system. They might talk to that CRM system through also a new vendor or something that they build internally. It's interesting. There's actually a lot of chatbots being built internally by companies right now.

I'm sensitive that I think it might be a temporary phenomenon. I like UX better than chat. Yeah, okay. I think we're going to see a hybrid of these two things. I think the GUI is not as dead as people think. You do see a lot of kind of chat interfaces where you're like, I think you just did way more work

than just going to the dashboard. Like, I'm 90% sure you just probably took all the savings from AI gains and efficiency and then just spent it on figuring out a prompt that like a dashboard would have solved. But I think we'll have a universe of both those things. But I do think ultimately most will come from software as like a, you know, on a percentage of like, like...

time that a knowledge worker spends inside of technology. But some of the most valuable things absolutely will be homegrown. The algorithm you use for discovering the thing or personalizing the medicine or personalizing the wealth data or Netflix's recommendation engine, those will be homegrown things.

Maybe still using a model from a proprietary player, but like the scaffolding there will be, I think, largely built internally. Interesting. So I guess the mental model for people watching might be that there's inside the house and outside the house. Outside is context, right?

which are just check boxes you have to check. And those might be, you know, end-to-end things that just do the thing like Salesforce. Yeah, I think that's right. And I think maybe another way to reverse engineering is if you were the customer of this company, would you care what technology they used for that category of thing? Or is it like, just get the box checked and I'm good? Yeah. As a customer of Netflix, I literally don't care what their ERP system is.

And then on the inside, like the things that are core that you really need to do, maybe the opportunity there is infra and dev tools that allow an internal IT team or internal engineering or AI team to run really fast, but then create exactly what that business needs. Yeah, yeah, yeah, exactly. And then there's still there's a little squishiness in there. It doesn't be very little or about it because like you also probably don't want to do homegrown dev tools like so. So you still might procure those.

And then there's this layer between the dev tool and the customer experience, which is, what am I doing with the output of these tokens? That's something I need proprietary software built on. And so what is in that universe? But I think that, for instance, I think there's going to be amazing dev tool opportunities coming from startups right now because it's a whole new stack that you have to build out for managing AI within your enterprise. That in-between bit is actually where you see open source a lot. Have you seen any...

renewed interest or any trends on open source interest in the enterprise? Selfishly, I love open source because we don't have to pay for what we're using. We should give grants. You don't pay for the hosted versions all the time? Sometimes we do, but I was actually thinking about this. I think a thing that would be cool to see tech

people do more often is just like start large scale open source projects and fund them. I think more open source software for the common good of just like lots of different services, I think is a huge net positive. I'd love to see more open source in general right now. It's such a great symbiotic model because the hacker community or the startup community, they want to move fast. They can't often afford the licensing of whatever that thing is. The large enterprise community,

They want support. They want to have people that are experts in managing this stuff. And so it's this great relationship, which is like you have this great good for society and you still actually can be building a successful commercial business because lots of people want to pay for the commercial version of that. Yeah, Aaron, in the early days of this cycle, there was a lot of concerns from enterprise about the security implications of using hosted models, like a whole bunch of companies like BandChat, GPT, use internally. What's happening now? Have they gotten comfortable with the idea of having all of...

their data go to open AI and anthropic or did they still, are they sort of really worried about that and trying to host open source models and things like that on-prem? Yeah. I think you'll see a different, different categories by industry. So a lot of times they'll go into, you know, a bank and they'll say, and they'll be proud of this. And as they maybe should be, they'll say, you know, we have our own kind of enclave version of, of X model. And then we built out a wrapper on top of that, you know, for, for deploying it to employees. Yeah.

And I think that you'll always have some percentage of the market do that, 10% of the market, because you still have a lot of people that have on-prem systems for a similar reason that will always be there. But the comfort level is absolutely increasing. That tends to always happen in...

if the industry or the company takes it very seriously. So OpenAI has taken security, privacy, compliance, regulatory controls very seriously. And so then that builds trust over time. So then more people can go and can experience those use cases. So anytime a software category matures and becomes more enterprise grade

military grade, you know, you'll see a correlation of the amount of trust in putting data into those systems. And I think AI is no different than, again, software and kind of cloud in that respect. Maybe part of it is just a lot of the behavior is from all these companies

they got in a lot more comfortable because of what you paid with cloud, right? They kind of already have preferences on how to do this, host it, and okay, we know how this looks. So it's a lot harder for you to sell on that first cycle. This cycle, you kind of compounded it for this next generation of founders. So thank you. People need to be really thanking me

I have not. Actually, you know, it's funny that you put it that way. I don't think I've ever been thanked for the hard work. Consider yourself thanked. The years of dinners, all the CIOs that made it possible for your AI startup to get sold in the enterprise. I'd like a big thank you. I have not been respected. That's it.

That has not shown up in a lot of ways, but I appreciate it. Every single white hair is a dinner that you had to do to teach people that you should buy this. This is the only way to get that. I would like a plaque at the office. I don't know why I don't have that yet. Because none of you have gray hair. And I'm responsible for AI working now. Yeah, but it's an interesting question, right? And this is what's so fun about the compounding nature of technology is

AI could not have happened in 2005. Let's just pretend the breakthroughs actually happened. Let's pretend an alternative universe where the exact same somehow, like Jensen had made GPUs as powerful as they were now in 05, but everything else was the same. Okay. Weird world, but let's just say. And you're, imagine going to a company and you're like, okay, we know that all of your infrastructure is on-prem and we know that all of your software is on-prem and we know you're on Siebel and we know you're on PeopleSoft and we know that all of your data is in this system, but

If you just move it to this other thing that you have no control over, it's in a data center you don't know about. You know, you have no ability to kind of go in and look at the data center and look at the hard drives. If you just move everything over to this, think about all the intelligence. I mean, it would be DOA. Like this industry would just be over. It wouldn't matter what the tokens are coming out with. So you actually, it was a requirement as a prereq to have cloud have happened, to have SaaS have happened. And then now AI happens building on that.

And then even the consumerization piece is so interesting. Imagine a world where we were only an enterprise using technology. We didn't have the kind of consumer access to technology that eventually happened with mobile, etc.,

you wouldn't have people seeing Chachaputi and Perplexity and Grok, et cetera, in their personal lives. And then saying, wait a second, why are my IT systems feel so archaic by comparison? That's what's causing all of this pull also. So you needed consumer adoption of technology. And then the fact that now a billion people, billions of people can use AI to then pull that into the enterprises. You need the enterprises infrastructure to be modern enough

So it's an incredible Cambrian explosion of now opportunity because we're in a moment where enough of the kind of core plumbing has been built out. Also, enterprises just became used to using like different software solutions versus just relying on like the- Three vendors, exactly. Yeah, 100%. Like they now, now some of them don't like this, but their IT stack is now hundreds of vendors versus a dozen. And so they can accept, okay, I'll take a meeting with Anthropic. Like this is totally, it's, you know, it's,

only now 1% increase in my total kind of IT, you know, vendor universe versus, you know, 15 years ago, I'd be like, nah, I can't bring in something that's not Oracle. Yeah. Well, keep accelerating. Please just keep driving the startups that cause this to happen. You had this front row seat to the transition from on-prem to cloud.

And now we're at the dawn of the next transition from cloud to AI. How do you think it's going to play out similarly? How do you think it's going to play out differently? And how do you think that relates to the TAM for software going forward? Yeah, so if I could try and merge these, actually, there's a cool connection point. So the probably single biggest bear reason why people didn't invest in SaaS in the mid-2000s

was they thought the market sizes would be basically the same size as the on-prem software company. And so if it was the same size of the on-prem software company, but also the software company that's already there is the incumbent, it's like, how do you squeeze out enough money to kind of make the business really, really, make it really interesting? And what everybody basically got wrong was it turned out that the TAMs were probably about 10 times larger. Why did the TAM grow?

grow so much. Because just to bore everybody, if you wanted to buy a CRM system in 20... Sorry, in 1999...

You had to be like, okay, I'm going to go to the systems integrator. I'm going to get a data center. I'm going to buy a bunch of servers from people. I'm going to install some software. I have to manage the network of that. And like, you know, lo and behold, two years later, you might have a CRM system and you probably spent 5, $10 million on the full project to do that. So think about who is the market that then can implement a best in class CRM system. It's the world's largest enterprises, you know, 5,000 companies, 10,000 companies.

Salesforce comes out and they're like for three seats online with your credit card, you have a CRM system as good as Siebel. Obviously there'll be some nuance because it didn't have as much functionality, but like for that company that was as powerful as Siebel, you know, getting started now, all of a sudden your TAM is basically every business on the planet.

So you go from a market that had maybe 10,000 customers, 20,000 addressable customers to now a market that has 5 million, 10 million potential customers. It is a totally different scale, like, you know, two or three orders of magnitude more scale that you can now go and serve. We had a similar experience, which was, you know, the industry we were disrupting was like legacy enterprise document management, enterprise content management systems. Same exact thing as Siebel in terms of like, we would read the S1 of our biggest incumbent competitor. And they were talking about like,

a thousand customers or a couple thousand customers. And literally, you know, now we have 115,000, but like, but at the time we had, I don't know, five or 10,000 when we started thinking about disrupting this and like the scale was just completely different.

So that meant the market sizes were so much larger. ServiceNow today, I don't know the exact latest market cap, maybe it's 150, $175 billion. Their incumbent competitor, when they were first growing and disrupting the market, today is worth maybe about 5 or 10 billion.

So, so if you had looked at this company 20 years ago, you'd be like, at best service now should be a five or $10 billion company. If it just was like a better version than the current thing. And it turns out it's, it's 15 times larger than, than, than what you would have thought Salesforce did the exact same thing. And so on AI, I think has a similar dynamic.

because you're basically increasing the total spend on software. So it's not so much that a new set of companies will buy software for the first time. It'll be all companies use software to do things that software didn't do before. And that will take from budget that previously was sort of untapped from software. So the budget will be from a variety of things. But often because now the software is doing useful work for you,

you can now afford to spend even more on that software because the alternative was a much more expensive sort of proposition. Here's where I think people kind of get it wrong though. They think about it as zero sum from, well, then all you can do is sort of take from the labor side of that spend. But it actually just turns out most companies aren't even spending on the labor side. They're just not doing the thing.

So, so, you know, most companies are like, just like globally are not like spending time to translate their advertising into a different language. So it's not that, that, oh, the, the, the market for the translation services are this big and we're just going to digitize it. It's like, no, a hundred times more people will do translation. Um, the, you know, in, in our business, like we have, you know, software now that reads your contract and pulls out the critical data from that. So you can automate a contract workflow, but

And like the number of people globally that are reviewing contracts and pulling out that data, maybe it's, you know, 10,000, 50,000 people. I don't know the exact number, but a very small percentage of companies are doing that with their contracts.

So they will now decide to prioritize automating a thing that they didn't automate before. You know, cursor, you know, as just a, you know, back to that example, a Replit or Devin or whatever, there's probably not a single dollar that's being spent on that technology that comes from, takes away from what people are currently doing. It's purely additive because now it's expanding the use cases that software can kind of tie into. So I think we're in for a potential scenario where the size of software, now you have to include AI in that,

could be five times larger in the next decade because it just, it supplies the actual underlying work that you actually bought the software for in the first place. And that just changes everything because now you're going to be paying for work as a paying for a tool that enables other people to do work.

I think that's such a powerful AI white pill, actually. It's not merely, you know, zero sum. We're converting payroll into software revenue. And ha ha, that's it. It's actually we're going to do things that enterprises should be doing, would have been doing. They never got around. And then.

you know, actually the people who are the consumers on the other end, they're going to have better products, better services. Like the thing will actually just be better for. I haven't yet read like the full like economic study on this, but where economists always get this stuff wrong is, you know, they do probably by default tend to have a kind of a zero. I mean, you wouldn't have like, you know, Jevon's paradox, uh,

if economists always knew how to anticipate these things. But what I think they often get wrong is they look at the total amount of market labor in a category and they're like, well, shoot, if AI automates that, that's now gone. Look how many jobs that is. I think we should debate it and we should talk about it because it's a very serious thing. But what they don't actually ever think about is the more microeconomic impact of this. So if I'm a company and it could be Box or it could be a 20-person company,

And I use AI to, let's say, code faster. Okay, well, why am I coding faster? Because I want to build a better product for my customers.

If I'm building a better product for my customers, my revenue should be growing faster. If my revenue is growing faster, I probably am hiring people to go and do things to drive that revenue growth. Maybe it's people selling the software. Maybe it's customer support. Maybe it's HR people to help scale the operation. And eventually I'll get to a point where I say, should I reuse those dollars that helped me grow faster to hire more engineers to grow even faster and to build more of that roadmap?

And if we were in a market that wasn't competitive, maybe you could say, you know what? I actually just want to take the profit and be happy. But we're in a competitive market. So if you're the one company that decides to sit on the profits that AI generated and just live off of higher profit margin, 20%, 30%, 40% profit margin as a company, you'll just have somebody come into the space and say, no, actually, we'll just, I'm fine to have 20% profits and do that same thing.

And then that company will eat into your lunch. So you actually then reinvest those dollars back into the things that are helping you grow faster. And that's actually like the microeconomic outcome of automation is you decide that you take that efficiency gain and you redeploy it into the business in something that will make you more competitive or grow faster or better serve your customers because you're in a competitive ecosystem. Right.

And that's why I think as, you know, over time, yes, you'll have some displacement in different categories. But over time, this is why it generally just looks like an upgrade to just how we tend to work, you know, in the world. Like we just tend to use tools to work faster, to make better decisions, to build better customer products. The customer gets a better result out of that. But we reinvest those dollars back into the businesses because we're in competitive ecosystems. And then the ultimate winner is the consumer. Consumer always wins on this stuff. Right. Right now.

with one X factor, just like bring it home to SF, like, like assuming we have a regular regulatory environment where those winnings can actually turn into surplus. Right. So if those, if it turns out that we take those winnings and then regulate, you know, the ability to build housing, then all of a sudden everything's still just as expensive. But like,

like the total utopian, which is sort of the abundance thing is like, wait a second. If we could use AI to automate more, we can build more. If we could build more, we could lower the cost of things. If we can lower the cost of things, then we can actually lift up anybody, anybody's lifestyle right now. You know, the, the, the, the 10 year old in an underserved community is

is now all of a sudden has access to the world's intelligence in the form of an AI agent. So they now are able to be educated better. If we can lower the cost of delivering services to people, then you get better healthcare. That's like the ultimate utopian state is we use this automation to actually deliver better outcomes for the world. And that will require tons of jobs as a result of that. Yep. We can be a society again because of AI. There you go.

Aaron, thank you so much. Thank you. Thank you so much for being with us. I think that that's a great place to end just because, you know, to be continued. Like, you know, I think that we're in the middle of the revolution and the revolution does not have to be a black mirror. It could be something that is driven by Jevin's paradox, driven by abundance for everyone. And that's certainly the timeline we want to be on. So let's do it. That's the future I'm betting on. Thank you.

♪