The AI Opportunity that goes beyond Models

So, for everyone I haven’t met before, I’m Alex Rampell on the apps fund. I’ve been at the firm for 10 years, and I stole this from Chris Dixon, who published a post like this probably 12 or 13 years ago. The whole premise is that product cycles drive growth.

The top of the chart here is the Nasdaq from 1977 to the present. It goes up sometimes and goes down sometimes. Over the long run, it has gone up, but there have been some very scary down points.

There have really been 4 major product cycles. There was the PC—and actually, before the PC, there was the semiconductor—but we’ve got to start somewhere. We’ll start with the PC.

There’s always an infrastructure layer of companies that are building the backend, and there’s the application layer of people building things that are actually used. Lotus was one of the first application companies. Adobe, Symantec, and all of these companies grew out of the 1980s. The infrastructure players, if you will, were Apple and Microsoft.

Then you had the internet. That was enormous. There were lots of bubbles along the way, but also some very enduring infrastructure companies, like Cisco and Akamai. There were enduring companies in the application space, like eBay and Amazon, that were built on top of that.

Then you had cloud. AWS accounts for the vast majority of Amazon’s market cap. You’ve got Workday, Shopify, Veeva, and others that were the application layer.

Now, mobile took all of these things that came before and put a supercomputer in everybody’s pocket. The vast majority of humans on planet Earth have a smartphone, which is pretty amazing. That was the mobile era, which is still playing out.

I just bought an Android phone to test things with. It was $40, and it’s more powerful than the ENIAC in 1946, or whenever the ENIAC came out.

Then, 2 years ago, the AI era started coming out as well. The Nasdaq is higher—we know that—but the AI era is really playing out. The cool thing is that this is not a net-new thing. It’s building on everything that came before.

If we didn’t have smartphones and we didn’t have cloud, but we just had the ENIAC, AI would be pretty cool. You could go check it out in a museum. But the fact is that you now have 8 billion humans on planet Earth, the vast majority of whom have smartphones, and the adoption of this new technology is taking off like never before.

The AI layer—the AI era—is here. The vast majority of net-new revenue happening in software is actually coming from AI, both at the application layer and the infrastructure layer.

It’s hard to think back 2 years ago. At that point, of course, GPT-3 had launched. I think GPT-4 had also launched, but it was all just text, imaging, and some basic reasoning. None of the native audio stuff or real-time interaction had happened. It’s hard to imagine how far we’ve come, even in just that 2-year time frame.

It’s really remarkable what these things have done. One of the ways I joke about this is that we have this idea of artificial general intelligence, or the Turing test: When can we tell the difference between a computer and a human if we don’t know who our interlocutor is?

If you were to take a person 10, 20, or 30 years ago and show them this, they would say, “Oh my God, this is fully sentient. This is smarter than any human out there.” We keep changing the goalposts a little bit on what exactly AGI is, but the pace of innovation here is just remarkable.

The important thing is the opportunity set that it unlocks. Whenever you have a bull market and very exciting technology, there’s always somebody saying it’s a bubble, it doesn’t work, or it’s all overhyped.

There was some MIT paper that came out—I want to be clear that this is not a fault of MIT; this is somebody who published a paper—that said, “Most enterprise deployments really aren’t working in terms of AI.”

We’re seeing the exact opposite, and I’ll show 2 things.

There’s a company called Ramp. They offer credit card and expense-management products. You see this giant tick up in January 2025, which is when enterprises started adopting this technology.

These are not necessarily startups. They’re more forward-thinking companies, not necessarily GE—companies with thousands of employees, maybe in the Bay Area or New York, that want to be more tech-forward. They’ve just realized, “Wow, this stuff—generation 2, to your point. GPT-3.5 was pretty good. GPT-4 was amazing. I could write a new episode of Seinfeld with it.”

Those were amazing things that I could do almost to wow my friends, like a magic trick. But now the magic trick has actually gone into the enterprise and is saving people time and money.

One of the things you’ll potentially get out of this presentation from me is that I have this prevailing view of human behavior: Everybody wants 2 things. They want to be richer and lazier. They want to do less work and get more economic value.

This is really what generative AI unlocks, and it’s really starting to happen right now. This has been a little bit of a flat curve, but it has been inflecting a lot. You see this in the expense data and in the growth of all of the companies, both at the infrastructure layer and at the application layer.

Whether they’re overvalued or undervalued is almost not the point. It’s hard to time the market on these things. The amount of value that they’re generating is tremendous, and we’re going to get into this in a second.

If anybody knows Maslow’s hierarchy of needs, this is a philosophical term for what humans need. At the base of the pyramid, people would joke, is Wi-Fi. You have all these things that have been true for hundreds of years, and at the very top of the pyramid is the concept of self-actualization.

But what I really, really need—if you talk to any teenager—is, “Where’s my Wi-Fi? Where’s my Wi-Fi?”

What’s starting to happen now is that the next thing is actually AI. Obviously, you can’t have AI without Wi-Fi, but something like 15% of adults on planet Earth now use ChatGPT every week.

Why are they using it? It’s just part of their daily routine. Whether it’s settling a bet with their friends about how something works, getting directions, or trying to solve a problem, people are using it for all kinds of things.

My wife just used it to complain to the school because her child couldn’t open the door. It’s against the law to open the door. This is a true story.

My wife had ChatGPT scan all the laws of California and the U.S. federal system writ large, even though our government is closed down. No, that was completely made up. She sent a very polite note.

I’m sure the school is going to start adopting ChatGPT to respond to people like my wife, saying, “We apologize on behalf of the bus driver.” They did send an apology: “Sorry, we made that up. Next time, we can open the door for your child if he is on time, when the bus has already closed the door.”

There are a countably infinite number of use cases for these things. The growth of minutes per user in the U.S. is astronomical. As these things work better and unlock more use cases, it’s obvious that the growth in minutes will go up. This is happening at a breakneck speed.

The key paper, which was co-written by the very smart Noam Shazeer in 2017, was “Attention Is All You Need.” It introduced the Transformer model.

I remember that we have a partner here, Frank Chen, who’s been here for a very long time, and he demoed ChatGPT—or GPT-2. It didn’t really work that well. It reminded me of this thing called ELIZA, which was a famous Markov-chain-based system. It was basically a therapist that came out in the 1960s or 1970s. It’s still around; you can try it.

Basically, you say, “Doctor, I’m not feeling well,” and it says, “And why is it, Jen, that you aren’t feeling well?” It takes the words that you say and turns them into a question. It feels kind of sentient until you ask it, “Hey, I want to complain to the school about the bus driver,” and then it says, “And why do you want to complain to the school about the bus driver?”

It doesn’t actually give you an answer or anything that you need.

From 2023 until now, we’ve really entered the golden age of apps. I base that purely on the numbers. The number of companies reaching extraordinary levels of growth is unlike anything I’ve seen before.

I’m used to companies that would grow—I don’t know. We used to talk about “double-double-triple” or “triple-triple-double,” and all these different ways of measuring revenue growth. Normally, if you’re selling a software product to an enterprise for, let’s say, $100,000 a year, you might sell a couple in 1 year, a couple the next year, and a couple the year after that.

Very rarely have we ever seen a software company go from $0 to $100 million in revenue in 1 or 2 years. We’re seeing this right now.

We’re not seeing it because people had too much money and are buying these things. These are companies buying these products because they unlock so much value for them.

They want to be lazier, they want to be richer, and this is unlocking that. I’m going to talk about 3 broader themes that we’re seeing in AI applications. More broadly, these are the types of companies that we’re investing in.

Partially, this is what we ask ourselves: What is defensible? What is it that the labs aren’t going to do? This is a very good question. It’s not like OpenAI just wants to be the back-end layer for everything. They have a leading consumer app. They just launched, arguably, a competitor to TikTok. Microsoft is also getting into the space in a meaningful way.

If you look at the history of software, this firm was started by Marc Andreessen. He started a company called Netscape. Netscape became roadkill due to Microsoft, which ended up in an antitrust case because it made Netscape roadkill, and whatnot. But how do you build an enduring company? What are the areas that potentially have the most enduring growth? There are 3 that I’m going to lay out.

The first is that traditional software is going AI-native. This is no different from saying that if you could build a time machine right now, go back 15 or 20 years, and say, “I’m just going to invest in every single cloud-native company that pops up,” you would have an incredible portfolio. You’d have Shopify, Veeva, and NetSuite. NetSuite’s a little bit older. You’d have Salesforce when it first went public.

It turned out that the incumbents couldn’t really respond to that because they were selling on-premises software or shrink-wrap software for a lot of money up front. They didn’t really know how to go for less money every month as a subscription. Category 1 is traditional software that’s going AI-native.

Category 2 is arguably the biggest. It’s not competing with the software market at all. It’s software that does the job people used to do. This is arguably a much, much bigger market. The laws of business still apply: You have to build real moats. You can’t just build a little widget that somebody underprices by a dollar tomorrow. We’re going to talk about that in a second.

Lastly, I call this the walled garden: really interesting proprietary data models where the value of the business becomes much greater because you’re able to deliver the finished product thanks to AI. I’ll talk about number 1 first: Existing categories are going AI-native.

We actually have a post coming out about this in a couple of days, but I’m sure everybody here has heard of or played bingo. I’m from Florida, so there’s lots of bingo in Florida. There are lots of different names on this list.

One of the key lessons that I’ve had as an investor is that Mercury is a great example of the tortoise that beat, and is still beating, the hare. Mercury built a neobank for startups. They said, “We’re going to be the better source for you when you start your company to deposit your money with us. We’re going to help you pay your bills, track your expenses, and be a basic accounting system.”

Mercury never stole an existing customer from Silicon Valley Bank until the weekend that Silicon Valley Bank failed. It’s what I would call the canonical greenfield opportunity versus brownfield opportunity. Brownfield means you’re selling to an existing market.

Let’s take email marketing as an example. You use Mailchimp, and I want to sell you a competitor to Mailchimp because it has AI. That’s going to be really hard. Or you use NetSuite, and I say, “Ditch NetSuite. I’m going to give you AI NetSuite.” That’s going to be really hard.

If you’re a net-new company—and this is what I mean by greenfield—you have no existing product. You’re not using anything; you’re a brand-new company. Or sometimes you hit an inflection point.

The inflection point I’ll pick on NetSuite for a second. I have 50 employees, and now I have 3 entities in 2 currencies. I’ve been using QuickBooks my entire life. QuickBooks can’t handle multi-entity, multi-currency support very well, for whatever reason. KPMG says, “You have to move to a better ERP system that supports that.” Now I have an opportunity to pick the better product in the market.

NetSuite is one product in the market. Or I can try this thing called Rillet, one of our companies, which is basically like NetSuite but closes the books for you and has 50 AI features built in. That’s a greenfield example.

These things don’t grow like weeds because you have to wait for new-company creation. You’re going entirely for greenfield and not for brownfield. But on every spot on this bingo board, the incumbents are all adopting AI, and they’re going to make their businesses much, much better with AI.

Bill.com is going to be a stronger business. SAP is going to be a stronger business. Adobe is going to be a stronger business because of AI. They’re going to be able to charge for new things. Workday will start charging. I mentioned this in my presentation a couple of months ago: Workday will say, “Do you want us to do reference checks on every new employee that you enter into our system? That’s $500 per reference check.” Why can’t somebody do it for $4.99? Because you’re stuck with Workday.

There’s a saying that I use a lot: The best companies have hostages, not customers. I’ll talk about a couple of examples here. For RPA, there’s an existing company called UiPath, which is a public company. For customer support, there’s an existing company called Zendesk, which is now a private company. For ERP, there’s SAP and NetSuite.

In some cases, like Zendesk, the company charges per seat per month. That is almost an extinct business model for support software. You might say, “Wait a minute. I don’t want to pay per seat per month when 99% of all queries can be answered by the support software. I want to pay per outcome.”

We’ve been aggressively betting on the bingo board. We evaluate every company that we see in this space—payroll, support, ERP. The important thing is that these are systems of record.

The best companies take hostages, not customers. We don’t want to invest in hostage companies. We don’t want to invest in companies that have a negative 100 NPS. We want to invest in companies that still have a very, very strong moat. That’s what I mean when I use that expression.

All of the companies that we’re looking at here are systems of record. What is a system of record? It means that it runs the entire business. Everything on that bingo board—how do you get rid of NetSuite? It’s basically impossible. You can enter in with an AI wedge.

More often than not, a lot of these bingo categories are about building the new system of record. The existing incumbent is doing that as well, but it’s still a no-brainer whenever you’re brand-new in the market or at this inflection point of deciding whether to use the old product or whether you need the new one.

The second theme, which I’m personally most excited about, is where new categories are emerging and labor is software. There’s no bingo board for this at all, because there weren’t software companies that did this before.

The predominant theme is that you have a lot of things where you would hire a person, but you can’t hire that person. Or the person you were going to hire doesn’t speak 21 different foreign languages and won’t work 24 hours a day, while software can do 90% of what that human would do.

Now you will pay for software, not necessarily at the same rate that you would pay for labor. This is not something that you would have hired a software product for before. Obviously, I can mention this ad nauseam, but the labor market is astronomically bigger than the software market.

The governing principle here is that you go look at a job: front-desk receptionist at Plaza Lane Optometry. Plaza Lane Optometry has a bingo board as well, in terms of the software that they spend money on. They probably spend money on Microsoft Office. They probably spend money on Squarespace or Wix. That’s on the order of $500 a year.

If you can deliver them a software product that does 5 out of the 8 things on this job posting, they will hire that software product. What do they pay for that software product? This is the part of the market that is almost unknown.

They’re probably almost definitely not going to pay the $47,000 a year that they’re advertising for this job, or whatever rate they’re paying for the job. They’re probably not going to pay $500 for software. But the creator and developer of this software product—an application software company—might say, “We’re going to charge you $20,000 a year.”

They need to be careful about how they do this. We often want to see them turn into a system of record, so that if they’re doing 5 of these 8 job responsibilities, somebody doesn’t pop up and say, “We’re going to charge $19,999 a year.” We want to make sure that this is a very, very sticky end solution for Plaza Lane Optometry.

You’re going to see, I believe, a lot of market-cap creation on the bingo board of existing software products that have a new, better alternative going after greenfield.

But here, you can go after brownfield. You can go after existing companies. You could probably charge a lot more. There's a path to much, much more explosive revenue growth.

But just to take a step back, you probably heard a ton about what's happening in legal AI. Given how document-intensive the industry is, there are tons of applications for LLMs in the space. Most of what you probably heard is around companies like Harvey, serving the defense and corporate side. Maybe less familiar to you is the plaintiff side, which is really about representing individuals in areas like employment law or personal injury.

We spent a bunch of time looking at the different companies on the plaintiff side, in part because one of the unique characteristics of that side of the market is that these attorneys operate on a contingency basis. Meaning, they only get paid if they win. So they're incredibly aligned with their clients. They don't bill by the hour; they take a percentage of the actual case outcome.

As a result, for every 100 leads that a plaintiff attorney gets, they often take 1 case, because anytime you take a case, it's an investment in your time and your labor. So, just incredible alignment with AI's impact on their core business model. Right? To contrast that, if you're a corporate attorney, and your junior attorney is 50 times more productive, you just eroded some of the revenue that you can actually charge your end client. Again, in this case, if you can make your attorneys 5x more productive, you can potentially increase your revenue by 5x or more.

And so, the Eve guys had a particularly interesting point of view from a product perspective. They really wanted to own the end-to-end workflow from intake all the way to outcomes. And so, to Alex's point earlier around voice, they recently launched a voice agent, which is actually collecting evidence from their prospective clients. It's sifting through mountains of medical records or employment documents and helping these attorneys figure out which cases to take.

Because it's generating this data set of case characteristics, it can say, “Hey, this case is potentially worth $50,000. This case is worth $5 million. You should probably spend time on this case over here.” And then it'll just help step through all the different phases of pre-litigation and litigation for these attorneys. It'll draft a medical chronology. It'll draft the core artifact of these cases, which is known as a demand letter. It'll file complaints.

And ultimately, I think what's so interesting about this business—and it speaks to why moats matter—is that these attorneys are living in this product all day long. One of the core pieces of feedback that we heard when we were diligencing the business was that literally 100% of the cases were flowing through the product. But interestingly, as EvenUp begins to generate data on outcomes, that data isn't public, right? That's not something the large labs can train models against.

And that data is actually informing better intake. Right? So they can then go back and say at intake, “Hey, given the characteristics that we've seen in all the cases that we've prosecuted across the Evee platform, these cases have 3 variables that make this case potentially worth a lot more money.” Or, to Alex's point, it can reduce the cost of taking on a case.

Before, an attorney was only taking a case that, at minimum, could potentially make them $50,000, and suddenly they can afford to take cases at $5,000. The market expands, right? And there's a big supply-and-demand imbalance today on the plaintiff side that Evee’s unlocking. And as a result, the market pull for this product has been, candidly, stronger than we even anticipated.

My hope is that it has a lot of characteristics that we'll be continuously investing in, where AI is just incredibly aligned with the business, both driving revenue and saving these folks money.

Well, thanks, David. Yeah, the reason why I wanted to talk about this is I think it's really cool as EvenUp, but it's a metaphor for the types of businesses that we find compelling. And why 0 to 30, certainly, or 2 to 30, is not normal, but it actually is normal if you're able to move very, very quickly and just deliver, again, this promise of, “I'm going to make you lazier and richer.”

Let's go to the next slide. Actually, before we go to Salient, Alex, why don't we just take some of these questions here because they're relevant in the context of Eve as an example, and also, before we switch to Salient, use them to exemplify why we find these to be particularly compelling.

There's a good question here from Brian. A lot of consumption-based AI apps have found it hard to become mission-critical, but easy to switch on or off as a part of the broader suite. How do you evaluate that in diligence? Maybe, David, if you want to use EvenUp as an example or others that we have in the portfolio. How do you evaluate that in diligence, and what patterns have you seen around which apps actually graduate to being essential?

Yeah, I mean, one of the distinctions that I often draw is this notion of differentiation versus defensibility. And I think AI is an incredible tool, often, for differentiation, right? So, the idea that the voice agent can speak to folks in 50 languages and gather that evidence is highly differentiated versus a human, right? Obviously delivering value, but that capability alone, in my opinion, is not a source of their defensibility.

The source of defensibility for Eve is owning the end-to-end workflow, right? It is actually in building a product that is contextual to all the work that that attorney has to do. And then I think, not unique to EvenUp, but one of the X factors is that the data that that business is generating—which Alex will get into a bit—has some of the characteristics of a walled garden. It isn't public, and it creates a source of compounding competitive advantage for the product itself.

The more cases that Evee can prosecute for all their different clients, the smarter the product becomes, and it actually reinforces that loop. It becomes sort of—you’re showing up to a knife fight with a gun, right? And so soon it's going to become an essential tool for any plaintiff attorney to operate with. And that just becomes very difficult to displace.

So it's not so much the AI-ness, right, in the voice or the ability to summarize documents; it's actually in becoming the system of record, this end-to-end workflow.

For sure. And in fact, there are multiple threads to pull on, but maybe I'll ask this question first, relatedly, around the potential upside of the market size of these companies around the labor-versus-vertical-software bucket, and how do companies in this category build defensible moats and, particularly, earn attractive margins as AI proliferates and costs continue to scale down?

Yeah, well, why don't we come back to that one at the end? I think, hopefully, what you'll get from this is that it's not like we're just investing in companies that do labor and that's the end. Their moats matter, if anything, more than ever, because the one thing that's happened in software is that, once upon a time, there was a company called WordPerfect, and WordPerfect kind of kept growing for a very, very long time.

Or once upon a time, there was a company called VisiCalc. And then whoever had the most distribution said, “I should do that,” and copied it. Obviously, WordPerfect is toast, VisiCalc is toast, and Lotus 1-2-3, which was the one that beat VisiCalc, became toast. But it would normally take 5 years for the bread to become toast. And there was a very, very high level of proliferation speed.

I mean, now Anish, David, Jen, and I can go build a software product. We can vibe-code, if you've heard that term. We can go build software very, very quickly. That actually increases the peril for anybody who's built a software product that has an enormous margin pool: your margin is my opportunity. Well, I can vibe-code against your opportunity.

It has to be very, very sticky. It has to have some unique competitive advantage, and data is often one of those. Actually, why don't we go to the next slide here, and I'll just talk about Salient a little bit. Sorry. So Salient is in the Eve mold.

And I know we also had a question about the societal impact of everybody losing their job. I don't think that's actually going to happen very quickly. 90% of Americans were farmers in 1789, and obviously, the tractor made some of them unemployed and made them do other things.

But most of what we're seeing, candidly, is not about eliminating work. I do think that the 3.5 million people who drive trucks, at some point in time, we have a better solution than the truck-driving human. You have AI doing that. Most of these things are really about having cost here and value here. You would never hire a human where they are producing less value than their cost. It just does not make sense. But if you can now hire AI effectively, you can hire AI where the cost has just gone down and the value has stayed the same, you’re going to hire a lot of AI. You’re not going to get rid of a lot of humans. If anything, you never know—this is so hard to predict—but what will humans do?

There was no job like product manager 75 years ago at a software company, or designer. All of these jobs that exist today wouldn’t have made any sense to somebody in 1800. So it’s hard to pontificate on that, but a lot of the things we’re seeing aren’t displacing people per se. I know it sounds pithy to say software is eating labor, but really, software is augmenting labor.

It’s like all these people I can’t hire—whether there’s a job shortage, a skills shortage, or whatever—I can now deploy people who will answer a phone. I would just never hire somebody to answer the phone for me at 2:00 a.m. I would hire somebody at 4:00 p.m., but not at 2:00 a.m. The value-to-cost equation is inverted.

A great example of this is Salient. They are going after people who collect on auto loans. It’s called auto loan servicing. You go to an auto lender, and they have to make sure they’re collecting on their bills. If a person is in a car accident and the insurance carrier is supposed to pay you, how do I make sure that insurance carrier is paying me on time and writing the check to the right person?

In this case, because I have the lease, they need to write it to me and not the person in their actual name. How do I do all of that? I would hire lots of people. I would train lots of people. A lot of these people hate their jobs because it turns out people yell at them all day and say, “I’m not paying you back for this car,” or the insurance carrier keeps them on hold for 4 hours, and that hold music is just terrible. You’re going to want to kill yourself if you have to listen to that 12 hours a day.

All of these are reasons why humans don’t want to do this, or why you can’t hire humans for this. The key thing with Salient is not that they’re saving you money. The key thing with Salient is that they collect 50% more. That’s the key thing, because Ari, the CEO, kept pitching, “I’m going to save you money. I’m going to save you money. I’m going to save you money.”

People like saving money, but if you go to somebody and say, “I will collect 50% more revenue for you every single month, and I will make sure that you don’t go to jail because none of these people you hire, who aren’t very well trained and have to listen to this horrible hold music for 4 hours a day, will say something they’re not supposed to say. I can make sure that AI doesn’t do any of these things.” That’s why that company is growing so explosively.

It really is much more about the value generation. Yes, the cost is much lower, and this is one of the questions where I’m like, how do they figure out how to charge for the product? They went to their first client, which had a $50 million-a-year call center with, I think, a 40% to 70% annualized churn rate per employee. And not because they’re firing people; it’s just that nobody wants this job.

So they now say, “I will do it for you with software. I will give you a system of record. I will make sure that we’re scraping every single new federal and state statute, because what you say in Missouri is very, very different from what you have to say in California, which is very different from what you say in Iowa. We’re going to do all of these things.”

No human can keep that in their head at the same time. It’s like, “All right, I’m talking to David. Shoot, what do I say? He’s somewhere in California. Oh, wait, but actually, he’s traveling to Kansas. I don’t know what to say.” Salient knows exactly what to say, and it knows how to say it in 21 languages. That’s why the collection rate is 50% higher.

This whole category of “we are going to make you more money, and it’s going to cost you less” is just a very, very hard thing to move away from. The key question for us, which I think is a very good question, is how do we make sure we’re backing the right one? How do we make sure that Salient is not—this was my number one question when Ari came in.

I said, “Imagine there’s a company called Talient and a company called Zalient. Why is Salient going to beat Talient and Zalient?” Ari, the CEO, actually had a very, very good answer to this. It wasn’t that he had looked up on ChatGPT, “How do I answer this difficult question from a VC?” But again, moats matter.

We know exactly what script to say. This is an example of a data moat. Because we’ve done millions of phone calls, we know exactly what to say. We have lower latency on every single statute that comes out. They actually have a very, very good product that ingests every single law, even as it’s proposed as a statute, in all 50 states.

Sometimes it’s at the county level. They’re doing all of these things that make it so much harder to compete, so they will not lose a deal. Moats matter more than ever because you’re able to create software so much more readily.

Actually, maybe this is a good dovetail to this section: Does this then mean software becomes way, way, way more specific in certain categories, and it doesn’t need to win a bunch of different categories to become a huge business? I think that might actually be a good dovetail to this theme that you want to cover here.

Yeah, this is the thing we don’t know. We obviously have many examples of vertical software companies that have become very big. ServiceTitan is a vertical software company. Mindbody is a vertical software company. Toast is a very large vertical software company.

Toast is designed for restaurateurs to run their business, integrate with DoorDash, pay their waitstaff, and do everything around operating a business. It’s a vertical operating system. It’s very, very hard to displace one of those. People would have doubted how big that could become. In fact, a lot of people did.

It was very hard for Toast to raise its Series B round because people would say, “Well, I look at the restaurant space, and half these restaurants go out of business every year. I look at how much software they buy. Well, they don’t buy any software. Therefore, this is a bad company. I’m not going to invest in it.” Fast-forward 10 years.

The reason that happened was that it turned out the business was much bigger in this case because they added financial services. The financial services were, “We’re going to do lending to restaurants. We’re going to do payment processing for restaurants.” We make it very, very sticky because it’s an entire software platform.

There’s no way for First Data or Global Payments, or any of these companies that traditionally do payment processing, to append some kind of software solution. That’s why Toast—you know, people got Toast wrong. It’s a very valuable company and a public company today.

I think the same thing applies when I’m adding in labor. It’s not just that I do labor and then somebody does labor for a penny cheaper. I need to build some kind of system of record for you, some kind of vertical operating system for you, so that you can’t just switch out for the cheaper player.

Maybe this is a good way to go into theme 3 here, which I’m very excited about. I call this the walled garden. This is really important today because if you look at—take a metaphor here—this amazing company called OpenAI shows up and they’re like, “Hey, we’re a vegetable farm, and we’re farming tokens. We’re going to sell tokens. We’re going to charge for tokens to all these people out there building applications.”

It plays out exactly as I talked about. OpenAI is an infrastructure company. We invest in all these application companies. But then OpenAI is like, “You know what? We should put some restaurants on our farm. A lot of people come to our farm. Let’s just have restaurants here.”

All these restaurateurs are like, “Wait a minute. You’re selling me vegetables. Now you’re competing with me. That’s not good.” The reason I bring this up as an example is because it actually is happening, and it’s a blueprint for how to potentially deal with a world where the source of the raw material is actually what’s rare.

Let’s go to the next slide, and I’ll show you. I’ll make this a little clearer. As I mentioned, this is kind of the world’s second-oldest profession. There are lots of cases where I construct some physical property, build a wall around it, and charge you for access to my property. You can do this in the data world as well.

I’ll pick an example on this little bingo board here: FlightAware. I’m not sure how many people have heard of FlightAware. How did they get their data? Their data, by the way—what is their data? There’s nothing proprietary about it. It’s all public.

You can buy an antenna on Amazon to receive what’s called ADS-B transponder data. Every airplane, after the Malaysian plane went missing, has a little transponder on it that shows its height, its speed, and all these different attributes. It beams that data down to planet Earth.

Antennas can pick this up and figure out: this tail number is at this place. I can buy one; it’s free. FlightAware, I think, has something like 100 antennas around the world. They pick up all this information, and they can charge for that—that’s a piece of data. I can ask ChatGPT that. They don’t know that. Only FlightAware knows that.

Or PitchBook does this for funding rounds. Who knew what the Series B price of a company in 1992 was? PitchBook somehow has that. LexisNexis knows this. CoStar knows this for real estate data. Bloomberg knows this for all sorts of exotic financial stuff. In many cases, it’s all free. Ancestry.com built its entire data model by buying genealogical records from the Mormon Church.

All this stuff is not available on ChatGPT. It’s not available on Anthropic. Of course, they can license it. The reason why I mention this is: What do you do with FlightAware data? What do you do with Bloomberg data? Or what do you do with PitchBook data? I’ll tell you what I do with PitchBook data: I hire an analyst and say, “Analyst, go write me a memo about this company called Eve and compare it to every other company in the legal space that had ever done something before.” PitchBook just sells us a subscription: every single Series B of a legal tech company since 1992.

Okay, that’s valuable. What would be more valuable is saying, because they’re the only ones that actually have that piece of information, they should probably charge $2,000 for that. That might mean—maybe this makes you nervous—we might need 1 less analyst because now we have a finished product. What we don’t want is just a subscription to PitchBook data. We want to somehow take that vegetable, if you follow my metaphor, and turn it into a finished meal.

One of my favorite examples here is DomainTools. DomainTools does 1 thing which is very interesting: they run a WHOIS query, which says who owns a particular domain name. This company has been around for a very, very long time. If I want to figure out who owned a domain in 1998, there is 1 place to go, and that’s DomainTools. This model has been around for a very, very long time before AI. Very, very large companies exist in this space. When you add AI, it makes it tremendously more valuable. I’ll give you 3 examples that hopefully hammer the point home.

There’s a company called OpenEvidence which, if you use it, apparently 2/3 of doctors in America use it pretty much every week. OpenEvidence is exactly like ChatGPT. The interface looks exactly like ChatGPT. Except, you know who has an exclusive license to the New England Journal of Medicine and every other medical journal out there? OpenEvidence. If I tore my Achilles and I want to read about what I should do—all of the evidence-based care out there—I can go to ChatGPT. It’s moderately useful. There’s no reason not to do that.

OpenEvidence is so much better because they’re the only ones that actually have it. In this case, they found all the data. They found all the unique vegetables out there. They convinced the vegetable seller not to sell it to any other restaurant, and they have a restaurant that delivers the whole thing.

Or there’s a 26-year-old company called VLex, an incredible company that just got bought. The CEO was telling me that the origin story of this company—he’s from Spain—is that he bought up every single legal record in Spain. Why would you want to buy up legal records? Because, I don’t know, Wilson Sonsini wants to know Spanish case law in case Andreessen Horowitz goes to invest in a company and needs to figure something out.

VLex would aggregate and digitize this information and sell it to law firms and other people that need legal information. Pretty high gross margin, but very, very low scale and predominantly European, in Spain. Then they said, “We should add AI to this,” and apparently it quintupled their revenue. Why would it quintuple their revenue? I might love Harvey. I pay for Harvey. Amazing product. But if I want to have a finished memo for my client at 7:00 a.m., I can’t get a paralegal to go do this. I know that it needs to incorporate some element of Spanish legal data, so VLex is my only solution.

Instead of charging $2 a month, or $2 an article, or $200 a month, or whatever they can charge for the raw material, they can charge for a finished product. Ask Leo is a procurement product. Every employee at every company hates their procurement department because, on the one hand, the procurement department is supposed to save the company money by making sure that some rogue employee doesn’t buy expensive widgets at an overpriced price from an unapproved vendor. But on the other hand, they introduce all sorts of complexity into the process.

Imagine that I’ve got a contract from Deloitte to give me AI and somehow revitalize my company. Who has 50 other contracts from Deloitte so I can understand what to push back on? That is actually very, very useful proprietary information. I wish I could go ask ChatGPT for this, but they don’t have the world’s treasure trove. What is the information they will never get? They’re never going to get 50 old Deloitte contracts. Where would you find them? I guess you could do a FOIA request or something, but you’re not going to find them. Ask Leo has these.

Go back 1 slide here. It’s hard to say where we’re going to find these things, but the most compelling ones that we found are when all the information is free, just like ADS-B flight transponder data. You find something that just wasn’t worth that much before because what do you do with flight data? What do you do with WHOIS record data on the internet?

I actually talked to an entrepreneur recently. He was like, “I like to figure out historical subscriber data on YouTubers.” YouTube doesn’t publish how many subscribers MrBeast had on August 4, 2017. Where would you find that? There’s some company that collates and collects that, and they’re just selling the data. It’s not available anywhere else.

We just published a post on the walled garden—we called it “Fruits of the Walled Garden.” All of these things, like creative archives and logistics, are walled gardens. You go to a county recorder’s office, and you can see who owns what property. You have to go to the county recorder’s office to find it. It’s all free, but you can digitize it, make it available, and then add AI to it.

This sounds like, “Oh, just add AI.” It’s much more valuable. The reason is that you’re saying, “I have something that nobody else has.” There’s a reason why people were buying this before: they were trying to create something that is of higher value at the end, and you can now do this.

So, go to every museum. Actually, I just talked to an entrepreneur who found every old manual. This is a great example: he found every old manual for blenders made in the 1980s and 1990s. You can buy this stuff for pretty much nothing on eBay. Where would you find a manual for an old blender from 1999? I have no idea. But apparently eBay is where you find it. It just shows these walled gardens that you can build with data. You could have built this before. You could build a company 10 or 100 times more valuable today.

So, Alex, maybe can I pause you here, in part because in the last era of investing, you gave the world a great framework for thinking about the battle between startups and incumbents: If startups could figure out distribution before incumbents could figure out innovation, that was their success.

How do you take us through the dynamic when you’re thinking about which companies to invest in—where it’s very clear that they can disrupt the incumbents in the category, versus the examples where it probably doesn’t make a lot of sense for someone to build a company that has a proprietary walled garden that is going to be very difficult to unseat?

Yeah. I think there are 2 ways of thinking about this. In the case of the used blenders on eBay or the manuals, there just wasn’t a company before that charged for access to a subscription like, “I’m going to sell you a data article that I digitized,” or, “I’m going to charge you $20 a month.” Probably not that interesting. But now, if you have this finished product that you can charge $1,000 for versus the raw material that you charge $1 for, maybe now the business is tenable.

One category is you just find a new data source, and there’s a reason why, in venture capital school, we learned to always ask, “Why now?” If this is such a great idea, why didn’t this exist 10 years ago? Uber had a great answer when it came out. There was no iPhone and no GPS sensor in every device. Once you have that, now you can have Uber.

The why now for some of these more esoteric things is a little bit like this: Why isn’t this a $20 million business? vLex, after struggling for 26 years, why is it now a $100 million business? It’s because you can deliver the finished product. Of course, I would argue that a lot of the old things that were out there, like Ancestry.com, are valuable companies.

They digitized LDS data, and a lot of people want to figure out where they came from. There’s an NBC show that says, “What are your roots?” and people like watching that. It’s a valuable company.

That would be one where I’d be hard-pressed to say, “How do you make that dramatically better with AI?” Maybe it’s, “I’m about to die. I want to figure out which one of my heirs to leave all of my money to. Please email them and set up dates with me so I can figure that out.” That’s the value add that you do with this proprietary data.

This is why I’m an investor, not an entrepreneur. Not anymore. I’m out of good ideas.

There is an existing data store. Maybe I license that, like OpenEvidence. They didn’t create new medical journal entries. They were just like, “Hey, let’s go distribute this to doctors. We know that doctors are really interested in this stuff. We know that all of the information is in these old medical journals, and the back catalog is very useful.”

Of all the things Michael Jackson did right and wrong, probably the thing he got most right from an economics perspective was buying the back catalog of the Beatles—or buying a big chunk of it. That ended up being worth a lot because, until the copyright runs out, a lot of people like listening to the Beatles. That’s going to become more valuable.

You can buy existing stuff that is already out there and already has a business, and that’s OpenEvidence. Or you can try to create something net new, which is more of the AskLio opportunity.

I don’t know if that perfectly answers your question, but my view on everything that’s happening in AI right now is that it’s one of these weird situations where it’s very different from cloud. Most on-prem software providers were like, “Cloud is stupid.” Most potential customers were like, “Cloud is stupid. It’s not safe. I don’t trust it. I want to host things.” Your entire IT staff would be like, “I don’t trust that stuff.”

The existing incumbents did not build cloud providers. PeopleSoft did not say, “Let’s go build PeopleSoft cloud.” They have it now, but that’s where Workday came from. Workday was like, “We’re going to build this.” It took a while for the business—and for everything—to catch up.

I’m very bullish on incumbents. I hope I can say that because I think NetSuite is going to figure out 15 different ways to monetize with AI. I think QuickBooks—Intuit—has a gold mine on its hands, where it’s just going to start charging per collection that it makes to all of its existing hostages that use QuickBooks.

That still does not mean that you don’t have these greenfield opportunities. You don’t have these new data opportunities. There are so many new opportunities that have popped up largely because of this value-cost thing. You find something where everybody would want it at $5, but it’s currently only sold for $10. Therefore, nobody wants it. Therefore, it’s not a business. Wait a minute: AI allows me to sell it for $5.

It’s one of these rare situations where it’s good for both. Whereas with mobile, most people thought BlackBerry was great and the iPhone was stupid. That’s why the incumbents didn’t—why didn’t Booking.com build Airbnb? Why didn’t a taxi cab company build Uber? Most people thought this was stupid.

Everybody thinks that this is a good idea because, of course, intelligence—like AGI in everybody’s pocket—is a very good idea. Nobody can argue against that. It’s more about the existing incumbents.

This is why I’m very bearish on the brownfield opportunity on the bingo board. I’m very bullish on the brownfield opportunity for walled gardens and for software that does the job of labor.

By the way, I thought you were going to say the smartest thing Michael Jackson did was let his family use his likeness for the Michael Jackson live show, which, according to Ben, has now generated more revenue from that show than his entire existence as a performer. But anyway.

I give more credit for this. Apparently, what happened was somebody took Michael Jackson aside and said, “You know where the money is? It’s like that movie The Graduate. It’s plastics.” Somebody was like, “You know where the money is? Back catalogs.” It’s a good point: I have a lot of money, so I’m going to go buy the Beatles back catalog, and then I’ll make money from it because CDs are going to come out, streaming is going to come out, and there are so many different ways of monetizing this. Smart move by the man.

Let’s—actually, let’s cover some of the questions. There was a question about the walled-garden metaphor that Daniel had here. The implication is that the new restaurant is direct-to-consumer. Why wouldn’t the company sell to the end user rather than to a business that is ultimately the intermediary?

This is a great question. vLex is a good example of this. vLex could have sold its data to Harvey. Instead, it realized this exact point: it should just be in the business of selling directly. It shouldn’t be selling to Wilson Sonsini anymore. Or, if it is, it should dramatically change the pricing of its products or its pricing strategy.

Instead of saying, “We’re going to charge this tiny subscription fee and allow so much of the value creation to occur elsewhere,” it’s going to do what OpenAI does. OpenAI charges very little per million tokens. We’re just going to consume that, enrich everything that we have that is proprietary to us, and then go sell that directly.

It’s a good question, but I think the point from an investment lens is that a lot of entrepreneurs are now looking for existing companies where the company doesn’t know what’s going on, and they can just buy that data. Those existing companies, if they’re run by an entrepreneurial CEO, realize, “Wow, I can make my business 10 times better.” We’re going to go invest in those.

Lastly, I’m just going to buy an antenna from Amazon and listen to Malaysian Airlines flights or whatever, and then aggregate this information that’s completely free. But it wasn’t free in the past tense, right? The number of subscribers MrBeast had 5 years ago—the number of subscribers today, you just go to YouTube and see exactly what that is. If I wanted to see what that was 10 years ago, that’s what is actually proprietary.

Sometimes the proprietariness, if you will, accrues over time. Everything is free. Anybody can go collect this stuff that’s free. The value only accrues over time.

There are a lot of examples of this. I can go to the Mormon Church and get my genealogical information, and they’ll probably give it to me. I don’t have to pay for an Ancestry.com account. But it’s useful and easier to do it with Ancestry.com than to fly to Utah.

Sometimes it’s just the ease of going to somebody who’s already digitized and put this information into an easier-to-digest form. That’s one of the reasons why people go to LexisNexis. That’s one of the reasons why people go to a lot of these providers, because sometimes they’re the only game in town. Sometimes they’re the best game in town. But increasingly today, they’re the ones that can actually give me a finished product.

It saves the end customer money as well, because I don’t really want to buy LexisNexis data. I just want to know if I should accept or reject this transaction. There’s a lot of enrichment that I do with the data. There’s a lot of workflow, and there are a lot of analysts.

If I’m a financial services company, I hire fraud analysts to tell me what’s going on. The raw material that I need to figure this out is this LexisNexis information. LexisNexis—this would be kind of bullish for an incumbent—can probably do a lot of things if it’s the only one that has that information.

Alex, I feel like you paid Joe to ask this question, but I’m going to take it here, and then I’ll switch gears to Anish and his 2 sections here. What is your view on white-collar-services AI roll-ups—that is, fully verticalized software-plus-services companies that are popping up?

I wrote an article about this 2 years ago. I called it Barbarians at the Gate, but with “barbarians” spelled with an AI, in homage to the RJR Nabisco deal in the 1980s and the book that was written about that.

I think it’s very interesting. What we’re great at is saying, “Here are 2 people that are going to change the world. They don’t know how they’re going to do it. We’re buying an out-of-the-money call option.” There are a lot of private equity firms out there that are like, “We’re good at firing everybody, moving people to the Philippines, and doing this and doing that.” This is a big thing that private equity is looking at.

At the same time, we do have a couple of bets in this space. It’s a very smart entrepreneur, but there’s never a question of, “Can I get more clients as an accountant?” because I can’t hire more CPAs to do tax returns. The hardest part is getting the clients. You have to go to Chamber of Commerce meetings.

It’s just very, very hard to buy one accounting firm and then, by virtue of all sorts of cost synergies, onboard 10,000 more clients. The way you would have to play that game is: you buy one accounting firm, integrate it for 9 months, then go buy another accounting firm, and then buy another accounting firm. Yes, is there value at the end? Absolutely. But you probably have to buy 200 accounting firms, and then you’re left with a pretty interesting business. There’s probably a big competitor called mid-market PE that’s done this 500 times, and they’re going to do a better job of that playbook.

On the other hand, there is a strategy that we think is very interesting, which is: instead of having a sales team, you buy one. Take the example of debt collection. I could buy a publicly traded debt collector that has lots of people, doesn’t do a very good job, and doesn’t follow lots of laws. I want to get started somehow. I built this great tool that I believe in, and I want to dogfood it, but I don’t have any customers right now. I know—I’ll buy a company that has declining revenue but 5 blue-chip clients.

I’ll buy this company for 3 times EBITDA, and now I’ll transform it with AI. I don’t have to buy a second one, a third one, or a fourth one. I can just say, “I have better collection rates. I have 5 blue-chip customers that love me, and I’m cheaper.” So, do you want to be lazier and richer? You’re like, “Yes. I already have the customers to back this up, and I can now onboard 1,000 customers into the existing acquisition that I made.” That’s actually quite interesting.

So, the question is, which one are you doing? I think the strategy of, “We’re going to roll up 100 dental clinics, and we’re going to make it better. We’re going to roll up dermatology”—I just don’t think we’re good at that game. I have a friend who rolls up dermatology clinics. The problem is that dermatology clinics are local: just because I bought one in San Carlos, it doesn’t help me do anything in Florida. I have to go buy more there. It’s the same with accountants, versus debt collection, which is very, very national. You could buy one, and that is your entry point. It’s kind of an opportunity cost.

Do I hire salespeople to go sell, or, if the best companies have hostages, not customers, do I buy some company that is stagnant and even shrinking because they don’t know how to respond to AI? By the way, every debt collection company would be crazy not to look into doing AI on its own. So, it is this battle between startup and incumbent, but there is an interesting opportunity. We’ve done one in the MSP space—managed service providers for IT—because a lot of IT now is not, “Hey, come into my law firm office with 50 people and fix my printers.” It’s, “Onboard me into Microsoft Office.” All of that stuff can be done remotely. It’s a very, very digital experience. It’s a $100 billion market. That’s a little bit more interesting because I can actually ingest more clients that way, as opposed to having to buy hundreds of these things. Hopefully, that makes sense.

Awesome. All right, should we switch gears? I want to turn it over to Anish because all of these things that we’re talking about also apply to consumer. So, with that, why don’t we talk about why and how this applies to consumer?

Great. Actually, if we’re going to do that, why don’t we skip ahead a slide and then come back to this?

This is the application of all the categories that Alex outlined to consumer AI. It’s the exact same pattern. The first and very important one is traditional categories going AI-native. This is happening. If you look at Photoshop, it’s a fantastic business, but what do you do if you’re a young designer coming up in your career? You want to use the AI-native Photoshop. The AI-native Photoshop is Krea. That’s over 18 months old, so it’s a fabulous product. It has all the AI primitives built in, and it’s the one being chosen by people who are adopting their first design tool early in their careers. This transformation of existing categories is definitely happening.

The second is category creation. ElevenLabs is a fabulous example of this. This market for voice and audio models really didn’t exist 5 years ago. There was perhaps a niche market for voice actors and voice dictation, but it just wasn’t interesting. ElevenLabs has done something much more ambitious. They’re a model provider, and they have both consumer and enterprise SKUs. Because they vertically integrate, they’re able to go after this opportunity and create the category in a very short period of time.

Finally, proprietary data. Alex talks about proprietary data. It’s near and dear to my heart because I worked at a large-scale consumer company that was based on proprietary data—Credit Karma—for many years. I’ve seen this playbook, and it works extraordinarily well.

The area where we’ve seen it applied in one of our investments is a company called Slingshot. Slingshot is an AI therapist. How do they collect their proprietary data? They go to existing therapists and provide an AI scribe, a note-taker. The note-taker takes notes while those therapists counsel their patients. It then uses the generated notes to train a foundation model, and the foundation model trains a consumer product called Ash, which is then sold directly to consumers. Of course, OpenAI and ChatGPT are formidable, but they simply don’t have the data that Slingshot has. As a result, Slingshot is able to provide a differentiated, high-priced product, and it’s working well.

Each of the observations Alex made is absolutely playing out in consumer AI. We’re very consistent in our approach to the 3.

Do you want to go back one? I think this is an important slide as well, and an important concept, because a very fair question is: Why aren’t either AI labs or Big Tech companies that have real model efforts, like Google, going to win it all?

The reason is that, in many categories, being an aggregator of models is actually preferable to consuming just a single model. The metaphor that we’re all familiar with here, of course, is airlines. It’s much more useful to search for a flight from SF to New York on Kayak because I can look across the inventory of every airline, versus just going to Delta or United and looking at their inventory alone.

The same thing is true in categories like vibe coding or creative tools, where you really want access to all of the models. The reason is that the models each have their respective specializations, so they’re not exact substitutes. You want to work with them all. You want a single pane of glass, and the labs and Big Tech companies can, by definition, only use their own first-party models. This is why we see the aggregators winning, and it’s an important trend and investing principle for consumer AI.

The key thing—everybody’s heard this framework before—is that our job is to find, pick, and win deals. Once we win deals, we help these companies actually achieve their objectives and, most importantly, don’t screw them up by giving them bad advice and telling them what to do. The CEO knows what to do, and we’re there to advise and consent.

The way that we do this is by trying to be the leader and the expert on every market. We’re putting out more benchmarks. There’s actually a really cool benchmark we’re coming out with: an AI productivity benchmark. For all these different categories, it’s pretty cool.

Everybody on the team has what I would call a process-interrupt job. The interrupt is: there’s a very, very incredible deal. Incredible, incredible, incredible. Let’s go meet with them. Drop everything. This is, unfortunately, from my wife and children’s perspective, a weekly occurrence right now. It’s, “Ah, I’ve got to cancel this. I have to have dinner with this entrepreneur who has discovered the fountain—not of youth, but of perpetual motion.” Or so they think.

The process part is this: somebody’s going to outsell Salesforce—not for the hostages that Salesforce has, but by building the greenfield version of Salesforce. How is that possible? Everybody hates using Salesforce. There’s a new company that’s going to do this better and be AI-native. How do we make sure that we’re adept at finding, picking, winning, and supporting that investment?

We believe in adverse selection versus positive selection. A very inexpensive deal that has been hanging around the hoop for 6 months is probably bad. We don’t want to meet with them. We want to meet with the best company. If it’s the best company, every other venture firm also wants to meet with the best company, obviously. They’re going to send out their big guns to try to win that deal, and it’s very hard to win these great deals.

The best way of starting with this is to write an article. We made a video about this as well, which has had hundreds of thousands of views. It’s pretty incredible. The Death of a Salesforce: Why AI Will Transform Sales. Josh Schmidt and Mark Andreessen on our team wrote that. Everybody wants to talk to them. But ultimately, knowing what you’re talking about really, really matters.

Or, you know, death, taxes, and AI. We've covered the gamut on everything around taxes. What about companionship? We do something that we just came up with: What are the top 50 enterprise applications? What are the top 50 consumer applications?

You know, we often get a somewhat pejorative joke. I think it's a compliment: We're a media firm that monetizes with venture capital. But there is a method to this madness. The method is that it's helping us find deals, pick deals, and win deals.

This is the team that does that. Everybody, again, we've got a very prolific process calendar where we're publishing things, becoming experts in certain categories, and trying to find entrepreneurs who are positive selection and building the best things here. We always see them.

A good example of this is Rillet. If you talk to Nick Kopp, who's the CEO of Rillet, Seema and Marc Andreessen just knew more about this category. We were in a very competitive Series B process.

Right. The follow-on to that question is: Is there a process case study to check out? What's the process for investment decision-making? Is the right assumption that each partner is given a budget to invest rather than needing investment approval, or how does that change, if at all?

Yeah, so we try to be highly conviction-oriented. I feel like my job, David's job, and Anish's job is to make sure that the right process is followed. The automatic mistake in venture capital is: “I'm old. I don't use Snapchat. Why would anybody want to send disappearing messages? That's stupid. Let's pass on that deal.”

Meanwhile, you have the really smart—not to be ageist—24-year-old who actually uses this tool every day, knows the entrepreneur, and says, “This is the greatest thing that I've ever seen.” And then the old person—I'm the old person here—vetoes that deal.

The right process is that, yes, we do have some amount of a budget, and our investment committee is effectively making sure that we believe very strongly that the process was followed, that you've met every competitor, and that the work is top-notch. We will often defer to the individual who is in the arena. Our job is just to make sure the process is followed and turn that second key.

It's a 2-key process, and, again, it's much more conviction-oriented. I know that doesn't perfectly answer the question, but we don't have a committee where everybody votes, you have to have a certain number of votes, and then it's all political horse-trading.

It's all right, especially for seeds, where a lot of the younger people have been focused on doing seeds, where it's a little bit trickier. But for the smaller checks, which we are predominantly focused on, let's just defer to the person with high conviction, but make sure that our entire process is done end-to-end. This is the expert; it came from the content, you know what you're talking about, and so on and so forth.

May you just generally talk about team evolution and changes—how you're thinking about augmenting the check writers on the team, how you're evaluating the path to promotion for folks in light of some of the recent promotions, and whether you will hire any additional people as well?

Yeah, I think the main thing that we often debate, very candidly, is that what we want most is probably more leverage as opposed to capacity. We have the capacity to do lots and lots of deals, but if it's the best deal in the world, we need to assume that our counterparty is Roelof at Sequoia, a top partner at Accel, or Reid Hoffman at Greylock. All of these people are active.

If it's a great deal, the entrepreneur wants to talk to as many people as possible and will often be starstruck by the person who started a multibillion-dollar company, as they should. That makes a lot of sense.

I would say the one area that we might look to add to is somebody who has probably built a quasi-generational company and is still very hungry as an investor. This is not a retirement job. This is an anti-retirement job. It will drive somebody crazy to the point where they want to retire because you have to work 20 hours a day sometimes.

Working 20 hours a day is something that my kids make fun of. They're like, “You just have coffee with people. How is that working?” It's like, you have to have a lot of coffee. You have to have a very high tolerance for coffee. Then you have to switch to alcohol at around 5:00 p.m. It's a lot of work to do this stuff.

Joking aside, you really need to be able to meet with everybody. When it is a great deal like this, how do we know? These are the errors of commission versus omission. If we get one of those wrong—not only do we lose our money because we were wrong, but we lose infinite money because we didn't actually invest in the right one.

We have to make sure that we're on top of all of these people and that our team is made up of experts whom all of these entrepreneurs want to meet with. I don't know if that answers your question, Jen, but the only thing that I would potentially add is that when it's time to go win a superpower deal, we all show up together.

By the way, I jokingly call Mark the Air Force because if we need a big strike, what do we do? We call in the F-35s. Marc's got a few of those. We'll have dinner at Marc's house. Ben will show up. We all show up beyond just this team.

Having a few other people who can lead the charge on winning deals and have board gravitas is helpful. That's how we use Brian. That's how we use Andy. I'm doing that too, largely. We wanted to get as much ownership as possible, and we might need more people at a senior level—not to find the deals or pick the deals.

Of course, we don't want to just say, “Hey, you're just a monkey that helps us win deals.” But that is a very helpful thing to go do, and that's a capacity perspective. By the way, I know it probably frustrates folks on this call to no end because we can't cleanly attribute a certain deal to a certain GP on all fronts.

Hopefully, that also represents how much we think about this sport as a team sport, and one in which we bring the entire force of the firm to bear as part of that. Also, just in case people did not pick up on it, Mark does not actually have an F-35, but he's the F-35 that comes in to win deals.

Okay, we have 2 last questions. Maybe we can bundle them together. This was in reference to any observations on customer retention to date for AI-native companies, and then just the scale of spending required for enterprise sales for these types of companies. Maybe David or Anish, do you want to take this one?

Yeah, I can talk a little bit about the customer-retention point. So far, we haven't seen a bunch of price shopping and switching, and I think it's important that the startups selling into these companies build a rich software ecosystem around the primitive.

This is what David was talking about with voice. It's necessary but not sufficient to provide a voice capability. You've got to build a lot of things around that voice capability.

I think the companies that are building rich ecosystems have done a better job of retaining their customers. The other thing is that AI is moving so quickly. Many of these customers are looking to these startups as their AI solutions provider, and they're looking to them for a much more holistic set of things.

Because new primitives are being released every day, the startups are helping drive them into the future and helping them capture a lot of the top-line gains from the new technology.

I’d say that so far, certainly on the enterprise side, retention has not been an issue. I’m happy to speak to consumer as well, where we’ve also seen strong retention signs.

Honestly, I don’t think we’re seeing a tremendous difference from an enterprise sales perspective. If anything, we’re seeing more inbound than ever. Eve hasn’t had to have an outbound motion, which is kind of insane given the scale at which they’re operating.

There’s a lot of market pull for a bunch of these categories, but at the limit, I think they will all need significant enterprise sales. And I think, if anything, what we’re seeing—especially when companies are selling to larger corporates—is more of a forward-deployed motion on the engineering side.

I think many large companies are looking to startups to better understand where and how to apply AI within their organizations. If anything, we’re seeing people invest more on the forward-deployed engineering side than necessarily on the sales side.

It’s a very cultural thing. Before you hire somebody—this is happening in a lot of startups; it’s not happening at GE—you ask, “Can you use AI for this job?” In fact, Ben is the CEO of Andreessen Horowitz, and he’s asking that before we hire people here.

I think that mindset, if you do it correctly—if you’re Eve and you’re like, “Oh, I’m just going to hire people that play golf with lawyers, and that’s my entire sales process, and I’ll never use AI for anything, and I’m just going to use NetSuite, and I’m just going to use QuickBooks”—that’s not how these companies are actually orchestrated. They really understand the power both on a cost side and a revenue side, and they’re transforming themselves internally.

All right. With that note, thank you all for joining and talk to you all soon. Thank you.