Hello, and welcome back to the Cognitive Revolution!
Today I'm excited to be speaking with Marc Bhargava, Managing Director at General Catalyst, about their Creation Strategy, and a fascinating new category of AI companies they're building: AI-enabled roll-ups – companies that begin as vertical software startups, but – having achieved initial product-market fit – grow by acquiring traditional service businesses and transforming them through AI automation.
Marc brings a unique perspective to this conversation. Having co-founded Tagomi, the first prime broker in the crypto space, which was acquired by Coinbase, he has hands-on operational experience.
Now, he's helping to found and incubate companies that are targeting not just the $1 trillion software market, but the vastly larger $16 trillion services market, including accounting, legal, IT services, call centers, and much more.
These businesses have traditionally run at 5-10% profit margins – much too low for venture capital investment – but having done a systematic study of the tasks that make up the work across 70 different service categories, Marc and team have identified 10 leading candidate verticals where they believe that anywhere from 30-70% of tasks are already automatable with AI today, such that with successful implementation, they could achieve more software-like margins of 30-40%.
From there, they recruit and assemble founding teams that combine applied AI technical talent with industry veterans who understand service firm operations and have participated in M&A transactions before. These teams then work closely with initial customers to pilot purpose-built AI tools and automation workflows, and then, where the “pull factor” from customers is strongest, structure acquisitions that deepen data access and feedback loops and align interests by giving firm owners a mix of cash and equity in the growing roll up entity.
For now, the AI transformation at these firms is very pragmatic - you might even call it mundane. Teams shadow workers to map every hour of activity, identifying which tasks fall into automatable categories: customer support, data entry, content creation, and increasingly, complex reasoning. Rather than wholesale workflow redesign, they layer AI into existing processes – an HOA manager can still email for help creating a board deck, but now an AI agent takes the first pass at the task. The goal, for the short term at least, isn't to eliminate jobs but to enable each person to handle 2-3x more work, addressing the chronic labor shortages plaguing many service industries.
Overall, I have to say, I'm very bullish on this investment and modernization strategy. The end customer value proposition is super compelling: better service, faster turnaround, more proactive insights, all at similar or lower prices. And this feels like the right way to bring AI transformation to Main Street businesses that simply couldn’t build such capabilities on their own. So I expect General Catalyst and their portfolio companies – including the service businesses they acquire – to do phenomenally well.
At the same time, if there’s one overarching theme of this show, it’s that we might all still be thinking too small. And as you’ll hear, despite the fact that Marc is pursuing one of the few investment and company building strategies that meets the AI moment and could outcompete Big Tech offerings indefinitely, I still think there's a good chance he's underestimating just how far this transformation goes. While he believes that most local service firms will ultimately participate in this transition, my guess is that only a minority – maybe 20% – have the leadership vision, operational excellence, and cultural adaptability to be genuinely attractive rollup acquisition targets. Beyond that point, my guess is that it will make more sense for the AI-enabled compounders to grow by simply taking customers from lagging firms rather than trying to acquire and transform them.
And bigger picture, I do struggle with the employment math. Marc’s vision is about serving more customers with the same size teams, and that’s a great place to start, but … When we're talking about 90%+ task automation, potentially in just the next 1-2 years, do we really need 10x more accounting, for example, even at lower prices? I’m sure there’s some latent demand to be unlocked, but my intuition is that many workers in fields like accounting will ultimately be displaced, and if I had to guess, I’d bet on human-centric fields like nursing, childcare, elder care, and education as the sectors with sufficiently elastic demand as to potentially absorb those AI refugees. Bottom line, I believe that the "land of abundance" Marc envisions is likely to materialize, and look forward to it, while also expecting that the transition could well be far more disruptive than he projects.
Time will tell, but for now, I hope you enjoy this deep dive into the future of AI-enabled services and the art of creating modern compounders, with Marc Bhargava of General Catalyst.
Nathan Labenz: Mark Bhargava, Managing Director at General Catalyst. Welcome to The Cognitive Revolution.
Marc Bhargava: Thanks for having me.
Nathan Labenz: I am excited for this conversation. You are pursuing what I think is a really interesting AI-enabled investment thesis that I think says a lot about where things are going, and I'm excited to unpack it with you. Maybe for starters, give us the high level. I don't know if you want to introduce the firm. I mean, people I think will generally be familiar with General Catalyst, but kind of give us the, uh, you know, the high level theory, motivation, and tell us what is a compounder.
Marc Bhargava: Yeah, sure. Well, first of all, thanks so much for having me on. Really appreciate it and excited to tell you more about what we're up to. Uh, yeah, so as, as you mentioned, I'm partner at General Catalyst. So I think everyone in venture is out there really looking for outliers and how can we go and invest in them, so we have historically built companies like Kayak and Livongo and have a really strong DNA of incubating and building companies. And I joined GC, this is my third year, but previously was a co-founder at Tagomi, the first prime broker in the crypto space. Uh, we were acquired by Coinbase, became Coinbase Prime. Uh, my co-founder, Greg Zatusa, still runs Coinbase Cri- Prime and the institutional business there, so at my heart, certainly a, an operator and builder and founder. And what really attracted me to GC was what are companies out there that are multidisciplinary that don't exist today where we can put together the right team or we have some unfair advantage to really build it? And as you can guess, today a lot of the theme around that is AI, and especially applied AI, with the model companies getting better and better with the new releases and can do more and more with LLMs. What are the sort of applied AI use cases that have not existed yet that should exist that we can build? So that's really been the focus of creation. The compounder story and the AI-enabled rollup story came out of that. It came out of, well, if you look at where AI automation is happening, in things like automating customer support, automating data entry, automating the creation of presentations and papers and copy and emails, um, helping human in the loop with logic and reasoning, math and coding, all of these different use cases, it's not always easy to get that in the hands of the end customer. So the idea behind the compounder or AI-enabled rollup is essentially if you look at firms or places like accounting, legal, IT, etc., call centers being a good example with Crescendo, you might be able to get this AI-automated software built, but it's very hard to go to market and get in the hands of the end client. These industries are fragmented. They're all across the country. They don't really buy software. When they do, they call it IT spend, and it's a really small percent of their budget. And so our thesis was, look, at Creation we want to incubate the next great generation of applied AI companies. In many cases, these are in the services businesses where we actually will need to go and buy our distribution base or buy our client list in order to get these to market, in order to get the proprietary data to continue to train and improve the models, and as a technique, to improve these companies, get more free cash flow, and then buy more businesses. And that's really where the compounder story comes from. In the public market, you have folks like Danaher and TransDigm and Constellation Software and even Berkshire Hathaway, where as business models they go buy companies, improve them, create more free cash flow, and then with that free cash flow, buy another company, and, and hence compound. And so that's why those companies, which have performed incredibly well in public markets, are called compounders. They've been hedge fund favorites for the last decade or so. We think there's a new generation of compounders. We want to create those, we want to incubate those, and we think a big part of that story will be the AI automation around how they're improving those businesses.
Nathan Labenz: Yeah, you're giving new meaning to customer acquisition for, uh, SaaS businesses, for sure. I guess just one extra beat on sort of the history of this, 'cause you mentioned, and I don't know if those companies super well, even Berkshire Hathaway, of course I know, but I don't, I don't know that I really know. You know, the, I guess the general outside view is like, first of all, M&A is really hard, right? Broadly. My general sense from sort of corporate-level M&A is like most transactions fail in some general sense of like they don't pay off as promised, you know, at the time that they were, uh, executed. What has been the driver, in your opinion, of the ones that have been successful? Is it about just like, you know, consolidating operations and creating efficiencies of the sort of PE variety? It doesn't seem like it would be enough to create a Berkshire Hathaway. It seems like there must be something more cultural or more, I don't know, more, more core somehow-
Marc Bhargava: Mm-hmm.
Nathan Labenz: ... than just that.
Marc Bhargava: Yeah, that's the hardest part of what we do. So if we look at an opportunity and say something like, you know, customer support and call centers, we're seeing through our company, Crescendo, that 50 to 70% of call center workflows can now be automated with agents and LLMs. Or you look at something like MSP and IT. We're seeing through one of our investments there, 38% automation. Or bookkeeping with one of our companies, Kik, 80% automation. So we're seeing this massive automation every three to six months. It gets more and more in these different industries. And that's one big part of it, but a second part is our founding teams have to have...... MNA, private equity, kind of industry experience as well. And so that's why this is a perfect project for Kreation, which is all about incubating companies, putting people together who might not naturally know each other. And so in many of these cases, what we've done is, we've found people who have led applied AI teams at places like Ripley or Ramp or Scale AI, and we're kind of combining them with industry experts who have done MNA before, done roll-ups before, and we're putting together both of those legs. And so that's a really important part of where GC fits into this, is identifying which industries have high potential for automation, but then also putting together a founding team that can do the automation, has experience in applied AI, and has experience in the industry in MNA, in private equity, and doing the buyout part as well. So it is really important to have both. I will say that technology companies have done a pretty good job at MNA if you look at the track record. Like with Facebook, for example, the acquisition of Instagram, I think we'd all agree is absolutely instrumental to Facebook and what it is today. If you look at something like Google, the acquisition of YouTube an- and even Android is extremely instrumental. If you look in our own portfolio, like Anduril, it's done over a dozen acquisitions as it's become the best overall platform for defense tech. And so sometimes I think also tech gets a bad rap, but has actually been very successful in creating public value through their own compounder stories as well. And now some hedge fund analysts would put companies like Microsoft and Amazon in that compounder bucket of making great acquisitions, improving those companies, cross-selling them, and have become amazing free cash flow stories as well. So I think the AI world is no different. There is an opportunity now in many industries to have a high level of automation. If you can now go and buy the companies, improve them with that automation, you're freeing up cash flow to buy even more, and you can have an inorganic story that can also complement an organic story of naturally growing the product suite and getting more customers as well. So at GC and at Kreation, we're really thinking, what are the ways to get the most tools in the hands of founders? How can we put together these unique teams, and how can we help them get to market quickly, and also have the data they need to fine-tune and train their models and have a really fast feedback loop? So we think this hybrid services plus software, AI-enabled, with the capabilities to also go buy your distribution, buy your data, this ends up creating the best companies and the best products. And we're seeing it in HOA management space with Long Lake, or in, you know, u- the MSP space, Crescendo in the call center. So we've placed now eight of these AI-enabled roll-up bets, and five of them we helped incubate the company, it was the first check-in. And it's a strategy that's certainly now catching on with the market, um, and having other investors come in and a lot of excitement. But probably most importantly, having a lot of founders say, "Hey, I get it." Especially second-time founders. The hardest part is getting proprietary data. The hardest part is go-to-market. And I'm good at building and operating and running companies, and maybe even did MNA myself. I sold my last company. And so we're really attracting these second-time founders and experienced operators who could go work with any venture firm, uh, is choosing to work with us, with Kreation, because we have this specific thesis and strategy.
Nathan Labenz: How much of this is about... And I do want to go into each
Marc Bhargava: Yeah.
Nathan Labenz: ... one of the portfolio companies in more detail and understand the story and nuances. But at a high level, how much of this is about the fact that it's been said many times in the AI era thus far, software can now enter the services market? There's this blurring between software and labor. In a sense, that's the greatest opportunity ever for software. In a sense, though, it also creates a more challenging relationship. The people buying the software are also currently selling labor. So if you're selling billable hours at a law firm and you're like, "Geez, this thing might make me more efficient, but that also directly cuts into my billable hours. How do I think about this?" How much of this is about saying that what once was a sensible model of selling complements to labor is now that we are selling substitutes to labor? So we need to re-bundle or restructure the market to bundle those things together in one company with one set of shared interests, because it just doesn't work as well as it used to given that switch from, or obviously not a full switch, but at least a partial switch,
Marc Bhargava: Yeah.
Nathan Labenz: ... from complement to substitute?
Marc Bhargava: Yes, it really now is about selling work. I think a great example of this in the public market is Palantir, which has done incredibly well. It sells work. There are Forward Deploys Engineers, and that's services, but ultimately they're selling an AI platform, system, and technology. So it's a combination of software and services, and ultimately Palantir would say, "We're just selling you an end result or work." I think that is becoming true in more and more industries. You mentioned legal, for example. Most legal spend is on third-party law firms. It's not on the software component compared to the spend on lawyers. So an approach that one of our portfolio companies, Udia, is taking is they are going after Fortune 100 general counsels, in-house law firms, and they're selling to them a mix of services and software, all AI-enabled, so that these Fortune 100 brands can spend less on third-party lawyers, and instead do more in-house with the tools that Udia is giving them, which is a mix of software to do things like NDAs and contracts, but also services. They recently bought a company called Johnson Hanna that does a lot of services as well, more manual things. It has been reported, of course, that this hybrid services plus software is getting more and more interesting because the margin profile is changing in services. Globally, services is a 16 trillion dollar a year revenue opportunity. Software is actually only a one trillion global revenue opportunity. So services is much, much larger, but historically, services have been quite low margin, especially free cash flow margin. A lot of venture firms and even private equity and others have been less interested in the space. Now, if you're automating 30 to 70% of the workflows in some of these services industries, you can AI enable the workforce and take a lot more clients, and doing that really changes the margin profile. So now these businesses are getting more interesting to venture investors because ultimately this hybrid services software can have pretty high margin profiles. Instead of something like 5 to 10% EBITDA margins, 30 to 40% in certain cases, if you flow through the level of automation at 50% plus. So it's getting very interesting what you can do with an AI-enabled workforce, Udia being a great example of coming in and helping these large corporations have their in-house legal teams be AI enabled and take on more work with the same number of people, and outsource less to these third-party law firms.
Nathan Labenz: I want to double-click on the numbers there. We have one trillion for software, 16 trillion for services. Obviously, that's a vast opportunity. Margin traditionally in services is low, I think you said 5 to 10%. But if you can automate 50% plus, then you can bring those margins up a lot. One of my long-standing beliefs about AI has been, I don't know what to expect on most things, but I do expect high consumer surplus. So I wonder how do those before and after offers look today to customers? And how do you think that's going to evolve over time? Is a company, and we could go into specific ones here and it probably varies, but is a company going and saying, "Don't buy services the old way. Buy our new hybrid service. It's going to cost you 20% less." And then under the hood, the cost structure is 50% less, and that's how they get to a higher margin? Is that margin sustainable long term?
Marc Bhargava: Yes, absolutely. I think it is quite sustainable because the offering is generally, "We're going to charge you around the same we do today, but we can do a lot more with AI." If you're already using a product and you're enjoying using it, if suddenly you can have two or three times as much output and they're keeping the price the same, that's unlikely to be a client that churns. So we view this less as a cost-cutting game. When we AI enable a workforce in IT or in legal or in call centers, it's less about letting people go or automating people; we're automating tasks. So we're automating maybe 30% or 30-50% of the tasks that they're doing, which enables them to do harder tasks, which generally are higher margin and to cross-sell as well. That is our thesis, that this will really allow for a lot more growth. If you AI enable a workforce, you can now take on twice as much work with the same team because they have more tools and they're automating away the easier tasks. So that is the thesis. For the end client, what that means is you're getting even more of the high-margin tasks done for you without a change in price. So you're getting more at a similar price, and we do think that is quite sticky. A really good example of this is in Ramp where an angel investor and GC is also a large investor where they AI enabled a lot of the sales and marketing team, and folks were two to three times more effective. They got the question of, "Oh, so we can now cut our sales or marketing team." But it's actually the exact opposite. If people are now two or three times more effective, you actually want to hire a lot more of them. It's higher ROI. So I overall expect this AI opportunity to continue to grow the global economy. It's a tech innovation like the internet or like cloud that ultimately will create more jobs, more productivity. I don't see a pricing of margin compression because the offering is actually getting more and more output. So overall I think that it's going to create a lot of opportunity that way.
Nathan Labenz: The story of doing more with the same and embracing that, or even wanting to scale it up, because who wouldn't want more if it's better, does make sense to me in some areas. The things that I always point to as every organization would want to be doing more. Every founder would say, "I ought to be doing more cold outreach. I ought to be doing more lead gen. I ought to be doing a better job of nurturing my funnel." Similarly on recruiting, "I ought to be reaching out to top engineers ten times a day or basically as much as I can." "How many are you doing?" "Actually zero because I don't have time for it." Sure, if we could get AI to do that, it seems like for a company like Ramp, you're entering into an infinite market and grow as fast as you can makes total sense. It's a little harder for me to wrap my head around some of those things when it comes to, for example, let's say accountancy.
Marc Bhargava: Totally.
Nathan Labenz: At least the way I've always approached accountancy. I'm like, "I want to do as little as I can to be compliant and satisfy my board." It's not like I have a wish list of things that I would be doing if I only had more bandwidth or if I could only get more bang for my buck. That to me does seem like much more of a... Just like-
Marc Bhargava: Yeah.
Nathan Labenz: ...I should be able to do it for less, right? That seems like what most people would want.
Marc Bhargava: Yeah. So I'll give you-
Nathan Labenz: How do you segment that?
Marc Bhargava: Yeah, two thoughts on accounting. One is there's a huge shortage of accountants, so it's an industry fewer people want to get into. There's a generation of baby boomers retiring. Many of the top accounting firms, if they could, could easily take on 50% or 100% more clients because their constraint is their shortage of accountants. If you buy an accounting firm and you AI enable it, you might no longer have that shortage because a lot of the more basic tasks closer to, for example, bookkeeping, can now be automated. If you AI enable an accounting workforce, that accounting firm has two or three times as many clients they could go out and get because there's actually a shortage of accountants there. So that's how a lot of value can be created by going in. In terms of the end client and what they might want to use it for, suddenly your accounting could be giving you proactive recommendations. Today it might be around building out three statements and filing your taxes. You can easily see how once you have that data, you can actually start making recommendations on where you can save money, how you're spending. You can have tools that turn that into financial models that can help you think about your capital needs, should you be getting debt funding? So suddenly a lot of the tools which maybe today are very stagnant can actually be very diagnostic, very proactive. I do think that even if it might be considered a boring industry, the suite of products can drastically change with AI. Perhaps more near term and more importantly, the same accounting firm can do a lot more work with the same cost basis, the same workforce basis if it's AI enabled. We're really in the current phase going after that first opportunity, which is how can an accounting firm, AI enabled, do two or three times as much work. That obviously really increases the margin profile because it's a similar cost basis but a much higher revenue, and that's the focus. But certainly to your question, to the end client, maybe things get delivered more quickly, they're more accurate. All of those things are pluses, but even how you visualize the product will certainly change over time with AI, as we're seeing in things like legal being a great example of this where lawyers can also compare all the contracts and then recommend pricing changes because there are inconsistencies. So how does the law firm or the legal department go from just a call center to also making recommendations and maybe being even a business plus center? AI can make that change. So you have those two opportunities. One is just taking on more work with the same group of AI enabled folks that obviously increases margin and is a huge opportunity in services. A second is to change the output of these services businesses, and things like accounting or legal start to be more diagnostic, more prescriptive, more insightful, and become departments that aren't just thought of as a call center or a check the box, but can actually help you in running your business.
Nathan Labenz: Can you take me through the playbook of how you develop these companies? The general sketch that I have is you're, first of all, assembling a founding team that has a mix of technical and M&A roll-up experience. How do you decide what it is you want to focus on building, and how does that differ from a more traditional SaaS strategy? I guess I'm particularly thinking about how much of it is going after what you might think of as core tasks versus auxiliary tasks? A couple examples of core task versus auxiliary, it might be like, I just did an episode with Ambiance Healthcare.
Marc Bhargava: Oh.
Nathan Labenz: You know, they're not trying to do, at least not yet, we'll see, they're not trying to do a patient-facing doctor that would actually interact with you and make the diagnosis. That's the core task that for now, the human doctor-
Marc Bhargava: Sure.
Nathan Labenz: ...continues to own. But they do have the AI scribe in the room, which is transcribing and then translating the notes to medical coding. And that's all the stuff that the doctor currently has to do in their pajama time.
Marc Bhargava: Absolutely.
Nathan Labenz: After hours.
Marc Bhargava: Yeah.
Nathan Labenz: They are happy to offload that, and so far, everyone is pleased with that arrangement. I might contrast that, though, to a customer service call center scenario where I assume the AI is taking the calls, and that would seem to be the core task. So, whoever used to take those calls maybe can take other calls or may. But it seems like this complement versus substitute notion seems to come into play. So how do you think about that? As you begin, and again, maybe it varies by vertical-
Marc Bhargava: Oh-
Nathan Labenz: We've got all the time in the world for you to go vertical by vertical. How do you think about this? Do we enter here with auxiliary, do we try to take the burden off, do we try to empower, or do we go right at the heart of the core task?
Marc Bhargava: Yeah, it's a great question.
Nathan Labenz: How do those decisions get made?
Marc Bhargava: It's a great question. We are really underwriting complementary AI services. We do not think that lawyers or accountants or IT specialists or HOA managers or any of those jobs are going away, but we think 30 to 70% of the tasks they do can be automated, and this frees them up to do more elsewhere. So even in the call center case, there are times when people want to talk to a human, so you can have the AI agents doing the more basic calls, but you have a human in the loop, and it can get routed to them at different times. But one human with a team of agents can take a lot more calls than one human alone. So we really are focused more on the complementary side of it. We are not doing this strategy if we think something can be 100% automated. So one example is, maybe coding will end up being 100% automated to some degree, and so maybe that might be an example where it does not make sense. But for legal or for accounting, we absolutely think on large deals, large transactions, preparing your taxes, even on IT or with call centers, there are more emotional calls, more difficult ones. In all of these cases, it is about AI enablement of the human team and being able to do more. And that is why, at least my view is, this AI revolution is similar to the internet or to cloud, where there is a big boost in productivity, there is a lot of interesting tooling, and the winners will be folks who figure out how to use that tooling, and the ones who create it. So obviously, large language model companies and labs and hyperscalers are creating a lot of this tooling. Our thesis is, let us put together teams then who know how to use it. Let us add industry experts and M&A experts to those teams, and then let us go out and let us go to market much more quickly on the innovation that is coming out of these major labs. So our playbook is that; it is kind of two or three parts. One is identifying, what industries have this AI transformation potential? And we looked at 70, and we said 10 of them have 30% plus automation potential. The second piece is, how do we put together the right teams that come from applied AI backgrounds, but then also complement them with folks in industry and M&A? And the third piece is to actually build out the AI-enabled software and the product to get pilot clients, to get customers, and to prove the automation percentage. And then from there, both grow organically, but also by smaller mid-sized platforms and targets and be able to show the margin improvement, the free cash flow expansion through growth, and then from there continue to wash, repeat, and become these compounders that are growing organically and inorganically, that have this data moat from the companies they own and have a really deep understanding of what the customer wants, ultimately giving both services and software offerings to their customers. So that is our underlying thesis. We have placed eight or so bets now. Many of them are kind of off to the races, much larger, better well-known companies like Crescendo or Longlake or Udia or others that we have more publicly announced. And we are seeing a lot of traction in those companies, but then also with founders and others who want to go in on the strategy as well as they see every three to six months the models getting better, the level of automation getting better.
Nathan Labenz: The vision that you articulate for this sort of empowered human with AI as a complementary tool is a vision of the future I would like to believe in. I think I am going to come back in a few minutes and challenge how sustainable I think that is over the medium term. But just to dig in a little bit more on the playbook first-
Marc Bhargava: Hmm.
Nathan Labenz: What is the sort of happy story for the relationship between a software business and the service customer? Presumably you do not just show up one day and say, "Hey, I would like to buy your company and transform it with AI," right? So what are the steps, the milestones, the points of buy-in, the points where you determine that, yes, this local services business actually might be a good target for us to really, not just try to sell into, but actually, marry, for lack of a better term, uh-
Marc Bhargava: Totally.
Nathan Labenz: Bring into the compounding story?
Marc Bhargava: Crescendo, the call center rollup, is a good example where they got to 10 or 15 different clients. They had a smaller revenue base, but they were selling in and had a high degree of automation. And that is when we said, "Okay, we should actually buy a call center that does a more meaningful amount of revenue." So a lot of the times, existing clients or pilot clients are very open to testing software, and then once they are starting to see the results, it is hard for them to fully implement it because they might not have the engineering team, they might not have the CTO. So you actually see a lot of the pilot clients or customers say, "Hey, this is the writing on the wall. We are seeing this automation, but we might not necessarily be able to fully implement it. We would like to sell to you. We get to cash out, but we also get to roll over." So if they are rolling over 30% of their equity and they think there is a 10X from here because they have seen the power of the automation, the potential, the story, they are more likely to sell to us than, for example, to private equity because we will let them roll over the 30% and get another 10 or 15X on that amount. So the more classic version of it is, we go out, we get customers, we get pilots, we start working with folks. We also get a sense of which of these companies want the technology transformation, which of them get it, and we put a premium on that, and we try to buy those style companies because our goal is really revenue increasing. So as opposed to private equity that might buy a company, cut costs, improve margin that way, obviously find debt financing, our view is a little bit different. If we find some of these pilot customers, they can end up being the targets. And especially if they are aligned with us on the idea of bringing in technology and wanting to grow revenue. And you will see that alignment a lot of times in the amount that they are willing to roll over in these transactions. And that is always a really good signal. So we have been seeing this now across many of the different industries we are in where we are buying companies that are quite proprietary because we have built that relationship and we have proven to them the technology, and they have been really impressed, and they understand that for it to really be embedded in their company, we have to buy the company, and it is a major shift.
Nathan Labenz: I would assume that's required as part of these deals? You're not letting people roll over nothing, right? Or,
Marc Bhargava: Right.
Nathan Labenz: Leadership has to. I assume you want to retain leadership, and they have to have skin in the game.
Marc Bhargava: Exactly. But it's been really great to see two things. One, folks who said, "Lots of private equity firms have tried to buy this asset. We didn't really want to sell, but now we really like your team. We like the growth approach. We like that you're venture-backed. We like that you're not using debt. We think this tech transformation is real. We want to sell to you, and we want to stay on board. We want to AI enable our workforce. We want to do all these things we really couldn't do alone, and we buy into the upside, and we want to have," the second part, "We want to have this larger percent rollover than you traditionally see in private equity." So folks are really pushing for that. So it's been, I think, our secret sauce. One, is in these hybrid teams of technologists and industry experts, and then two is proprietary deals in the assets we're going after. We're building these relationships through trust, through tech transformation, and we're getting to do proprietary deals. So that combination of people arbitrage and deal arbitrage, I think, is a really special alpha for us.
Nathan Labenz: Can you tell me about the process of working with early design partner customers? I specifically have a couple things in mind here. One is, how do you think about the faster horse problem? There's always this debate, and it maybe can be answered with a sequence, but there's always, what does the customer want? "Oh, I'd love to have this go faster," or "not have to do this task," or whatever. And then there's the, maybe this whole thing could be transformed. Maybe it could be structurally different. I often use a customer service example. I have a talk on AI automation, and the example that I always go to there is just taking a customer service ticket. In many organizations, the first thing that might have to happen with a customer service ticket is prioritize the customer service ticket. So that's something that you could probably automate. You can think about what are the criteria, and you can talk to the people that are on the ground doing it. Through interviewing and looking at a bunch of examples, you can draw out of them how they are actually doing this. I find that that's often one of the stumbling points actually, especially if people have to, if these services businesses are trying to do it on their own. They don't necessarily know why they're doing what they're doing in many cases, so it needs to be drawn out of them. Literally eliciting the chain of thought from the human is a key step in the process. But then there's also the idea that, "Well, jeez, what if AI could respond to these customer service tickets?" We might not have to prioritize them at all because we could just respond to them all immediately. Then this whole prioritization step becomes a moot point. So how do you think about that as you're just getting off the ground and trying to figure out, is this a business that we want to keep the bones of and find these automation opportunities within, versus when does something need to be reimagined?
Marc Bhargava: Sure.
Nathan Labenz: And again, that might be a sequence thing, but tell me how you think about it.
Marc Bhargava: Generally, we think about it by going in there and shadowing the customer and understanding all the workflows in their business. So then we add up who are all the folks, what are their workflows, and we map out the hours at a company. Then what we do is we go through those workflows and hours worked, and we look at where AI could automate some of these workflows, and that's trying to get us to this 30% of hours worked automation, so 30% of tasks can be automated. Generally, historically over the last two years, those tasks have fallen into about four buckets. One bucket is customer service and customer support and frequently asked questions. So if you have any business that has a lot of customer support and is heavy on that, it could be a good potential industry to go after. A second one is the creation of presentations, company, marketing, emails. Businesses that have a lot of that work stream also have a high degree of automation there. A third is if the data tasks are very repetitive, so there's a lot of filling out forms or looking at sheets. Insurance is a good example of this, checking the box that this roof caved in, yes or no? Did they fill in the paperwork? So that's a third bucket. And now fourth, more recently, is around logic and reasoning. So again, with insurance, it's less automating forms now and more underwriting, like, "Can we make a suggestion of how to price this insurance coverage policy based on everything we know?" As models are getting better at logic and reasoning and math, you can have more AI tools provide a quasi-answer with a human in the loop to what you might want to do or say. Legal is also a good example of this, where now Anthropic and others have created models that are quite good at reasoning. So a lawyer can say, "What questions am I going to get on my case?" You can start to do more human prep. So those are four very broad buckets, but areas where AI, as they stand today, between LLMs and agents, can automate a good deal of things in those buckets. So then when we start going through these 70 industries, we break them down, and we actually get on the ground. Our companies that we've helped incubate are actually out in the field talking to clients. We are mapping out how they spend their time, and then we're overlaying where that mapping fits with the buckets we've identified have a high degree of automation. That then really helps us underwrite what is the opportunity in this space? What is the opportunity if we were to buy this client and instead of having it as a customer, have it as a company we own, that we can automate large portions of, that we can take on more revenue, that we can create more free cash flow from? So that's the strategy we go in and implement, and that's why it's important to have people who know the industry, know these clients, can get us in the door, but also folks that come from these really strong applied AI teams that know what best-in-class automation looked like at a Ramp or a Rippling or a Figma or a place like that.
Nathan Labenz: Do you do that initial mapping even before starting the company?
Marc Bhargava: Yeah.
Nathan Labenz: Was that part of the 70 to 10 cut-down?
Marc Bhargava: Yes. We do it at a certain level, from the 70 to 10 cut-down. Once we've given the first round of funding, to secure the second round from us, the team needs to gather granular data from their customers and pilot clients. They then present that data to us, and we provide the capital to acquire a target, often one they are already collaborating with. We conduct a first round of analysis, mapping and examining industries. As we place bets, the team has an even higher bar to build a useful AI-native software product, win an initial set of clients, and then demonstrate the automation with those clients to us.
Nathan Labenz: It's obvious why business owners would be interested in this. How did the staff at these companies react to this transformation? What have you seen regarding excitement, fear, rejection, or even quitting, protests, or sabotage? I'm sure it must cover the entire spectrum. Are there any general lessons you would share with business leaders, regardless of whether they are entering a structure like this, about AI transformation?
Marc Bhargava: Absolutely. I think there is, at first, a bit of apprehension around any new tool or technology. One distinctive aspect of what we're doing is that we state, "We're coming in here to AI-enable the workforce." We are not coming in to replace the workforce, but to AI-enable it, take on more customers, increase revenue, and foster a growth story. I think it comes across as authentic because we generally are not using debt, especially in these early transactions. This means we don't have interest payments to make, which would necessitate cost-cutting or seeking savings. We are adding costs in the form of software subscriptions and engineers, but we are placing a bet that an AI-enabled workforce can accomplish much more work and generate significantly more revenue. For the staff or the people working there, it is a challenge to get everyone up to speed on this AI technology. We're going to provide these tools, but then, of course, they also have to go out and acquire more clients and generate more revenue. Generally, however, we are saving them from the most mundane or repetitive parts of their day. So, we are freeing them up to do more interesting, more demanding, and more human-to-human interactive tasks. Once the process is underway, and as you're rolling up more and more companies, it becomes much easier because, many times, after a first acquisition, we have those teams—especially the leadership team, but even other employees—talk to the second team and help win them over for the second acquisition. That is a strategic part of it as well. Once you've done a few of these, you have the stories and case studies to tell. I think that really assuages a lot of the concerns you mentioned, like, "Are people going to get fired, or is all of this going to get automated?" It's like, "No, we've done two or three of these. Here's exactly what happens." We give you the software and tools to automate your mundane tasks. You get to do more complex workflows. It ends up being better for the company. Your shares are worth more. It is very compelling, especially compared to the private equity alternative.
Nathan Labenz: Are there any rules of thumb? For example, the Ambiance and Cleveland Clinic partnership required every doctor in their system to use the AI scribe once. I thought that was very interesting because AI tools do have an error rate. So, that projected confidence, even with a minimal ask, says a lot about how reliable your product needs to be if you're going to make those relatively minimalist demands on your team to try it out. From that one interaction, they report that now 75% of their 4,000 doctors use it all the time across 60 specialties. Presumably, that's product excellence.
Marc Bhargava: Great. Yeah.
Nathan Labenz: That translated to an amazing story. I assume one advantage of being more tightly integrated is that you don't necessarily have to reach that level before you can push adoption. But are there any mental models, rules of thumb, or frameworks for how to think about how far along the product has to be versus how much to push people to use it even if it's imperfect? What managerial takeaways have you seen?
Marc Bhargava: A big learning for us is to find assets and companies that have a pull factor. If you're purely underwriting in a private equity style, perhaps you're installing a new management team, cutting costs, and doing all these things—that's what you're looking for in the asset. For us, it's crucial whether the company—we've now incubated a company, we're going to buy one of our clients—truly wants the technology transformation. Have they tried different products? Have they tried hiring people? Have they truly pushed at the leadership level for technology and AI? We will pay a premium, or our companies will pay a premium in the acquisition, to have the right partner and the right partner mindset. So, one thing that's extremely important is that this is something they've been trying from the top, from the leadership position. Then, once the acquisition has been made, there are all kinds of techniques to ensure not just leadership's buy-in, but the entire company's. This honestly means not changing up the workflows too much, keeping them similar. For example, in HOA management, you can email saying, "Hey, create a version of the HOA board deck for me," and it will create it and send it to you via email. We're not trying to onboard you to some new software tool. In that case, you're using your existing text or email workflows. My earlier question to you was: do we totally reimagine it? Not really. We want to keep many of the same workflows, but we've already identified that a huge part of it, over 30%, can be automated behind the scenes. So, we're trying to keep folks' workflows very similar but use AI, sometimes without them even knowing what the large language model is doing behind the scenes, or how it's writing this copy and all that stuff. It feels like they're just interacting with an agent that is able to send them a board deck they need for tomorrow's meeting, and so on. We're generally mapping over those same workflows, so it's often not a big change for the workforce to learn and adopt these tools because our companies are building the tooling to be very similar to what they do day-to-day.
Nathan Labenz: Yeah. That makes a lot of sense. So there's not a, um... I mean this sort of speaks to, I think another major question right now which is like, when trying to get value from AI should we be creating what I traditionally refer to as smart workflows, where you have a sort of step-by-step process that is human designed and each step may or may not be an AI step, but the AI steps obviously you can optimize for performance and make sure they're working well? One way I characterize those is like you'll know you're successful when you don't have to check every single output anymore-
Marc Bhargava: Right.
Nathan Labenz: ... and I think for a lot of tasks you can get there today. And then there's like this new paradigm of AI that is the more agentic, like the AI chooses its own adventure, you know, maybe within some constraints but right now like, "Claude Code, go write this thing," or, you know, um-
Marc Bhargava: Totally.
Nathan Labenz: ... you know, some of these calling agents like, "Call this person and try to sell them a subscription," or whatever. These sort of like open-ended, you don't know how many rounds it's going to be, you don't know what the tool calls it's going to choose at any given time are going to be, you don't, obviously don't know what the environment is going to give back.
Marc Bhargava: Right.
Nathan Labenz: How much of that are you seeing deployed in your companies?
Marc Bhargava: So on the services side, I think it's going to be more of the first case that you outline where there are seven or eight steps and some of them will be automated and maybe a AI agent or... Does step one but a human does step two, and so it's going to be these hybrid AI enabled workforces where there are multiple steps and some are fully automated, some are not. But in services I think there's still going to be a huge portion of lawyers, accountants, to your example, doctors, and instead they'll just be different steps that are automated but they'll be key people and it'll be this hybrid AI enabled workforce. That is where I view the world ending with services for many reasons like one, the technology is not there, but maybe even more important, many of these spaces are highly regulated so you can't just have an AI doctor, even if it was great, and many of them are very human and emotional and it's... You know, a lot of it is talking through, learning medical news versus just what's the diagnostic and the prescription, et cetera. So I think in most services industries it will be the hybrid approach and that's the one we're employing. I do agree with you though that in certain industries it could be the full agent automation approach. Go and here's Airbnb, build me Airbnb but for cars instead of for locations and a swarm of agents will be able to code it up and you'll have sort of an Airbnb for cars that's built for you. So there will be, I think, some industries that are maybe fully automated. Certainly the holy grail right now is the automation of coding and you have obviously OpenAI and Anthropic working on that but also Microsoft and Google and then Cognition and Factory and even I think Cursor and others are moving to that trajectory. And so coding is a great example of obviously a massive industry where maybe building websites and products and platforms will be very close to fully automated using agentic swarms. But in the industries I'm focused on in applied AI, especially a lot of the services industries which we discussed is a really large part of the global economy, I think it will be a hybrid approach because of the human element, because of the regulatory element, because of the complexity and sort of gray areas of tasks. Writing code or doing a math problem generally doesn't have a kind of got to make a gut decision on something, while filing your taxes and how you classify something, what's the risk you're willing to take in an audit, et cetera, does. And so we're pretty clear on here are the industries where we think this hybrid approach makes sense and we want to incubate companies, maybe they have a relevant component, maybe they don't. And then also investing in traditional software and agentic companies that we think could maybe fully automate, um, an industry as well.
Nathan Labenz: Yeah, one other question I had that prompted is, how much do you build versus buy? 'Cause I could imagine, like, a lot of, you know... We have another episode coming up with the CEO from Intercom.
Marc Bhargava: Oh, okay.
Nathan Labenz: I could imagine, like, a lot of these companies might oughta just, like, buy Intercom and, you know-
Marc Bhargava: Yeah.
Nathan Labenz: ... have them do, you know, a, a big chunk of stuff. Like, it seems like there's a bit of a tension between we want to focus on automating this sort of auxiliary, you know, somewhat less core tasks from the verticalization or vertical specific approach. It feels like the more, you know, true to the vertical you are, the more you're kind of focused in on the core, and the more you're kind of trying to do the stuff that, you know, the, the pros don't wanna do, the more you're kind of entering into more of a horizontal, like, Intercom type-
Marc Bhargava: Mm-hmm.
Nathan Labenz: ... scenario.
Marc Bhargava: Mm-hmm.
Nathan Labenz: So, maybe there's just enough middle ground there that there's still, like, plenty left to build-
Marc Bhargava: Yeah.
Nathan Labenz: ... but h- how do you think about that spectrum?
Marc Bhargava: There's a lot of middle ground, because there are many point solutions out there in a lot of these industries, legal, accounting, IT, customer support call centers. There are a lot of great technology companies and a lot of solutions. Generally, they're point solutions, generally it's not a full platform play, generally they're not customized, and generally they're not trained on your proprietary data. So, there's still a lot left, and that's why the core of these transformations is staffing a great, backing great founders who are great technologists, because there's a lot to do in terms of stitching together different point solutions, there's a lot to do in terms of creating custom solutions, there's a lot to do on fine training models and making it unique to your dataset. And so, all of those tasks require a really deep understanding of AI and of engineering, and we generally have seen that kind of, uh, folks who are coming from applied AI positions make really good founders in these sort of companies because they absolutely can assess and buy things off the shelf, but then they're also figuring out which model we should use, where, how to stitch things together. Do we want to customize our own model, take an open source one, train it against our unique data? You have to have a lot of really expertise in AI and in building software and actually doing things. It's not just a research question. And we kind of tail off of all the amazing improvements made on the research side by an Anthropic or an OpenAI or a Google or others, where as they improve their models and can do more, we flow that through and we get the benefit. For example, with legal, at least Udia has found that, you know, OpenAI is really great at summarization of case text, for example, but the Anthropic model is better at reasoning and asking difficult questions that you might face in a courtroom. And so having kind of both OpenAI and Anthropic every three to six months improving their models, you know, we float both of those through at Udia to deliver value for the end client, which is the in-house general counsel at a Fortune 100 firm. And so you have to have this technology and AI expertise to understand the different models, where to use which one, and where there are also point solutions that are already built, you know, you have to have the technical expertise to evaluate them and integrate them as well. And so we still think you need a really amazing kind of technical CEO or CTO or founding team to be at the core of these projects.
Nathan Labenz: How much is about customization versus standardization? If you buy 10 businesses, is the idea that you'll have 10 different fine-tuned models that represent the unique flavor of those different local businesses and a bunch of different custom workflows that recreate the idiosyncratic ways that they did it? Or it seems like more of the economic promise would be in the standardization-
Marc Bhargava: More, of course, standardization.
Nathan Labenz: ... where you said-
Marc Bhargava: Yes, if you buy 10 IT firms, 10 call centers, 10 accounting firms, 10 legal service firms, 10 HOA management firms, the solution base that we're building for each is the same. So there is some slight customization, but it's extremely similar. The workflows in accounting, the workflows in MSP IT, the workflows for HOA management or property management are very similar. We're backing a team; they're the holding company. Then they're going and buying maybe 10 HOA management firms, for example, or 10 IT firms, or 10 property management firms, in the case of Dwellie, one of our investments out in London. The technology you're building is very similar across those 10. We don't see a huge difference there. But what we do see as unique is that as we buy these companies, we get their proprietary data. We can feed it in, train models, and adjust our software based on having proprietary data and much faster feedback loops on what the client wants. Ultimately, we think our product is much better on the market than someone who isn't doing a roll-up strategy, who doesn't have access to their client's data, or as fast a feedback loop. This is a very fast way to scale out applied AI, in our view.
Nathan Labenz: Yes, it makes sense. I will bet on you to do well, and I will also bet on the customers of your portfolio companies to be very happy with the level of service they're getting. I'm a little concerned about how much disruption there might be along the way, but I'm optimistic on some level that we can manage that disruption. How many? It varies by vertical, but it strikes me that you're not going to acquire that many companies. I think the whole market goes this way, right? Maybe with a very small residual that's just high-end. On TikTok the other day, I saw an artisanal pencil sharpener. I don't think there are too many of those out there, but there's apparently at least one. People are constantly debating in the comments whether it's just a bit for TikTok or whether people are actually sending this guy pencils to be-
Marc Bhargava: Right.
Nathan Labenz: - artisanally sharpened. Whatever, there's a long tail; there will always be an artisanal pencil sharpener, I suppose. But leaving that aside, it does seem to me like all these markets over a not super long timeline are going to go this way. As Tyler Cowen famously said not too long ago, it's the people that are the bottlenecks. You are the bottleneck. So how many of these local service businesses that exist today, your sort of Lebens, Bhargava, and associates, how many of those make it over the hurdle and get into one of these hybrid things, and how many are ultimately out-competed and go out of business?
Marc Bhargava: I think right now, you're already seeing a consolidation. Ignoring even AI, there's consolidation in accounting, for example, and the consolidation happening today in many of these industries is being driven by private equity. They are saying if we buy two firms that do the exact same thing, there's obviously overlap; could we remove some of the overlapping people? This saves cost. With the cost saving, we can pay interest payments on our debt, and so we can finance these with debt, and that's the model. So you're seeing a lot of industry consolidation in many of the areas we're going after already today, being driven by private equity. It was especially driven during the period of low interest rates, and it's continuing to be driven today as well, with just more of a push on cost savings. So we are in some ways offering an alternative to that. If you are running an IT shop, accounting shop, or HOA management, you're probably starting to feel the pressure in the industry, and you have private equity folks now saying, "Hey, we can consolidate you." You have the option to sell to them, and it's probably really going to change the profile of your company, and maybe even the management. So that's one option. We are trying to come in with an alternative option, which is same management, same team, let's AI-enable it, let's go for more revenue rather than less cost-cutting, and that's our path to higher margin. We're only right if the AI enablement works, because if we're asking you to go and take twice as many clients with the same group of people, that's just not going to match up unless we're really giving you the tools so that each person can do twice as much work, which generally means each person is automating the more basic stuff and focusing on the harder pieces. So we really need to get this AI automation right in order for our thesis to play out. But the good news is, over the last two and a half years of doing this, we are seeing we are getting the AI automation right, and every three to six months, we are seeing we're getting it even more right as more models are released that have more capabilities. Last fiscal year, 150 billion was spent on AI R&D and CapEx by the Mag 7 alone. This year, it will be closer to 250 billion. So there is a massive investment, there are obviously a lot of the smartest minds working on this technology, and it shows in the results. We are seeing an improvement in math, logic, reasoning, and automation. So we think our pitch is the right one, which is you can come in and change these companies to grow rather than come in and change on the cost cut and consolidate. That's obviously what we're fighting for and why I'm really passionate about what we're doing at Kreation and our vision and our mission. I think it's a different AI story than a lot of people are used to hearing. But anytime there's a new technology, you're going to have people say, "No, this is going to be really bad." And there are areas we have to be really careful about and we need to think through, but I think overall this will be a benefit for humanity and to the companies that are AI-enabling their workforces.
Nathan Labenz: Yes, I'm sold on all of that, with some tail risk that might imply it's not a benefit to humanity. But if it's not a benefit to humanity, I don't think it will be because the mundane task automation doesn't pan out.
Nathan Labenz: I'm confident we will get that.
Marc Bhargava: No, there's
Nathan Labenz: But, in terms of
Marc Bhargava: no need for the AI-enabled accounting software that takes down humanity, for example. So there are other questions perhaps that are outside
Nathan Labenz: Yes.
Marc Bhargava: of my expertise, but I think what we're doing for service industries, truly improving what people do day-to-day, like getting rid of many boring, mundane, repetitive tasks and freeing people up for other tasks. We've been hearing from the companies we've been acquiring that this is something they're excited about.
Nathan Labenz: We'll come back to the more extreme scenarios again in a second, but just in terms of the scale of realignment, if you will, or euphemistically, of the markets you're entering: If I had to put a number on it, it would be somewhere in the 10 to 20% of companies currently operating as traditional local services firms that might get over the hump to join one of these compounders. For everybody else, it's like, why bother? Your leadership's not that great, the pull factor is not that strong. You're sort of got your head in the sand. Maybe we'll just take your customers; maybe we'll hire some of your best people. But I guess my expectation is that the other 80 to 90% of those firms just won't make it. Is that too... And I don't know, you're striking a tricky balance here on the communication front, but to some degree, you might want to be clear about that if it is what you believe, because it would be quite favorable to your deal terms or your ability to get people to move fast if they feel like they want to be in that 10 to 20% and not
Marc Bhargava: Yes.
Nathan Labenz: not standing without a chair when the music stops.
Marc Bhargava: Maybe that will be good for deal terms, but I think the reality is a majority of firms could join these AI-enabled roll-ups because the tools we are building should be super intuitive. This is very different from the traditional software model where you have to go in and learn a system like SAP, Jira, or even Salesforce, and it's difficult to do. The way we're actually building these AI-native toolings, especially agents, is such that you can put it in your own workflows, so it should be easy enough for most companies to use. That's the first point. The second point is, there's a lot of churn in the workforce; many baby boomers are retiring, and a new generation is joining. That new generation is already fluent with ChatGPT, using many online tools, and certainly with email and Slack, and other things that we integrate very deeply with for our agentic workforce, etcetera. So, I think, first, the tools being built are very similar to existing workflows, and most firms can use them. Second, while there are, of course, some extremely old-school firms still using fax and other things like that—we absolutely have seen that as well—part of that is also just a generational shift, where that generation is retiring in the next 10 years or so. The folks coming out of high schools and colleges today, many of them use ChatGPT as an operating system, not just for search, where they're getting life advice, and in some ways, they're even more AI-native. Engineers today are using Cursor and Windsurf. So I overall think a majority of firms, even in the services space, will be able to participate in this if they want. There are other alternatives too. Folks, especially on the leadership side, might feel like they want a 90% cash out, they want to sell to private equity, and they want to take the money to a beach. That's how capitalism works if they want to sell their company that way. But we have a very different pitch, which is to take a second bite of the apple here: Transform what you and your family have built over the last 30 years with AI, make it much more effective, grow the business, roll over a significant amount, and be part of an alternative to that PE consolidation. We think that really resonates. Honestly, I think it can resonate for not just 10 or 20%, but for a majority of firms, and we're certainly seeing it resonate for more and more every year.
Nathan Labenz: Okay. So here I'm going to push you on just how crazy this all gets. But maybe we can start with, you went through 70 industries, winnowed it to 10 based on where you could see a path to 30% automation. What can't be automated, and what are the barriers? To some degree, obviously, operating in the physical world, we don't have robots yet. We do have quite a few humanoid companies that are, I think, on track to change that in the not too distant future.
Marc Bhargava: Sure.
Nathan Labenz: But clearly it's not here yet. Then we have this notion from AI companies that there will be a drop-in knowledge worker, and it seems there that one of the big barriers is context, like lack of context or inability to handle enough context. How do you categorize what can't be done? And do you have a map for what AIs specifically can't do that makes those things not yet accessible?
Marc Bhargava: There are two main categories here. One, as you already mentioned, is that much of today's automation focuses on continuous tasks like data entry, analysis, and presentation creation. It does not apply to hands-on work such as plumbing, roofing, HVAC, or firefighting, for example. So, there is a lot that still happens in the physical world where there might be some automation in routing, scheduling, or payments. However, this wouldn't meet our 30% criteria, meaning the automation wouldn't significantly transform roles like that of a firefighter, at least not right now. So, that's one category. A second category is that we avoid acquiring firms with high churn. When we implement AI automation, consider digital ad agencies or PR agencies where many clients use multiple providers. If we were to acquire one and begin transforming it with AI, any hiccups along the way could lead clients to switch providers because it is a high-churn environment with multiple options. Therefore, another major criterion is that we only operate in industries with strong business dynamics, where customers have sticky, potentially multi-year contracts, and very low churn. When we introduce AI automation, it is not always 100% smooth, so we need the benefit of at least six months to a year to implement these changes for our workers and end clients. These two factors have excluded many industries: either there's a significant real-world component or the business economics around churn are risky. These have been two areas that have recently been flagged, making it difficult to enter those industries. A third factor, of course, is the quality of the AI itself. Two years ago, you couldn't use AI to automate much of what we are examining today, like in insurance or even insurance underwriting. The AI simply wasn't capable. So, there are other industries where AI might perform tasks someday, but current models are inferior to humans, offering no real value.
Nathan Labenz: So, you don't expect to see this drop-in worker? It strikes me that there are several data points indicating that big tech takes all. A friend of mine coined the term 'the big tech singularity,' and I think that's basically my mental model of what will happen. Not necessarily that they'll own every customer relationship or anything. With Google and Facebook, specifically on their ad platforms, they had local agencies, sellers, and people translating the platform's value proposition to local businesses. But those facilitators, like 'Oh, I can do Facebook ads for you, Mr. Local Business Owner,' have been in a tough spot with significant margin compression. In medical, for example, multiple studies have shown that AI can diagnose better than doctors and even prescribe better than general practitioners, at least these days. Some studies suggest a hybrid approach, which is often seen as a refuge, where the doctor and AI work together. However, there's suggestive evidence that the AI performs best, then the hybrid, and then the doctor alone performs worst. I also recently saw a data point from Harvey where they released a stack-ranked leaderboard for their big law benchmark. Their own fine-tuned model has fallen behind eight other foundational models that seem to be winning based on 'the bitter lesson' and the continued benefits of scale. There is still this issue that the drop-in knowledge worker doesn't quite exist because it's difficult for them to absorb all the context. But that seems like something that will probably be solved. I wonder if you think it will never be solved, or if it is solved, how do we avoid a situation where we're just dropping these AI knowledge workers into the environment, and now you have a very literal complement versus substitute dynamic going on?
Marc Bhargava: For us, many individuals will manage teams of agents. You might ask if this replaces the people they used to manage. Not necessarily. I believe that overall economic growth and GDP output can increase significantly, leading to a land of abundance. People will still have jobs, but they will manage teams of agents. There will be far more jobs available because the global economy will be much larger. Every technological shift we've seen has raised the question of job loss. While it has meant fewer certain jobs, overall economic output has grown, creating newer jobs and different opportunities. It's very difficult to predict and precisely outline overall economic growth, what it means for new jobs, which jobs are lost, and which are gained. We absolutely must be mindful of that. But at least as we observe in cases of hybrid services, software, and AI-enabled solutions, our strategy is not about replacing jobs; it's about freeing people to do other tasks. The number of tasks that can be automated is certainly growing, but other opportunities exist, especially in client-to-client services and many businesses that require relationships. For the doctor or nursing example, in nursing, we incubated a company called Hippocratic, an AI-native nurse. We worked with many healthcare systems to acquire proprietary data to help build this company with the founder; it's part of our creation strategy. There could be ten times more nurses in this country than there are currently. Right now, we don't do preventative medicine calls or follow-ups after surgery six or nine months later. There could simply be many more nurses, and there is an absolute shortage. In many of these industries we are targeting, there is a shortage of people. We need to provide them with the tools to address that shortage. I believe much of this will involve AI-enabled agents working under them. That is the reality of the view we see on the ground today. Can we debate if robotics, especially AI-enabled robotics, will eventually lead to a different world? That's a very difficult question, and there are all kinds of technology and automation improvements. I would just say, personally, and at General Catalyst, we are very focused on the opportunities in front of us today. We have been very mindful of the companies we back. We are a large investor in Anthropic and believe they do an excellent job of considering the risks and safety of AI. So, responsible innovation and AI safety are key to our mantras in the companies we invest in. We believe that like every other technology, adopting this will present challenges. It is the mindset you bring to those challenges.
Nathan Labenz: I have a vision of that classic Simpsons episode where Homer is just hitting the enter button at every prompt as the nuclear power plant boots up until he finally decides to put the dippy bird in to hit the button, and then he just leaves. The plant, I think, does end up having a meltdown as a result. But this managing agent future, I'm a little bit... We were talking the morning after the Grok-4 presentation.
Marc Bhargava: Right. Yeah.
Nathan Labenz: I don't know, it feels like it's coming at us really fast. I want everything you're saying to be true, and I genuinely believe that the consumer surplus is going to be amazing. The quality of service, the access to expertise, the democratization of that, I think, is all super exciting upside. But then I see Elon saying things like, "New science next year." And I'm like, "I don't necessarily think I can count this guy out because for one thing, they keep climbing these... There's no benchmark that seems to be immune to their ability to climb the hill." And as they make this transition from other training signals to learning from reality itself, and they start to equip the models with the same high-powered simulation and other tools that their best engineers at Tesla, SpaceX, and Neuralink are all using. And it starts to actually compete with them to redesign a part, or find an efficiency, or can we eliminate this part. If those systems can start to do that in the next year, it seems like we are in for a really choppy ride, even if the consumer surplus story is totally true, which I do think it is. Is there a ratio of how much you spend on AI to how much you save on labor? I usually tell people, "Expect ten to one." But do you have a rule of thumb for if you can take some task and automate it, what's the savings? Of course, that could be you can do ten times as much of it.
Marc Bhargava: Well, for us it's very different because
Nathan Labenz: But what is that ratio?
Marc Bhargava: We're not really focused on the savings, to be honest. We're focused on growing the
Nathan Labenz: But how about, if you flip it and say, "How much more can you do with the same inputs?" What's that ratio?
Marc Bhargava: There, we have an objective. I think ideally, in a three-year period, we would like folks to be, with a 20% cost increase, able to double your revenue. So that would be a broad target. But in three years, if we come in, we spend 20% more than you spend today to make sure you have the right software, the right engineers, the right AI tools, increase the cost 20%, can we increase the revenue 100% in a three-year period? And if we can underwrite, yes, we think that is a plausible case, then we think buying that company could be a good investment. But we are not looking at AI as a cost saving, even though it of course can be that. That kind of limits your upside. We're looking at it much more, how do we grow and have venture-like returns and fast growth and scale by AI enabling our workforce to take on more revenue? So, our mindset is a little bit different on that metric specifically.
Nathan Labenz: Is there anything that big tech could do that you are afraid of? One that I always point to is, and they swear they'll never do it,
Nathan Labenz: up until maybe they do it, is not allowing, basically using their best models for their own products first, right?
Marc Bhargava: Mm-hmm.
Nathan Labenz: Like so far, we've had this generally even playing field. ChatGPT has a model, and the API gets it roughly the same time, maybe even a little before or whatever, but it's pretty much parity. But you can imagine a world where Claude 5 and Grok 5 and GPT 5 are all proprietary products only for some significant period of time, if not indefinitely, and that could potentially really shift the power. Now, from a consumer surplus side, they probably still can only charge me so much a month because I've got four or five of those
Marc Bhargava: Yeah.
Nathan Labenz: options.
Marc Bhargava: Of course. Yeah.
Nathan Labenz: But how does that play into the app layer? It seems like it could be a real problem for the app layer.
Marc Bhargava: I think that's true if there were two or three players. Even two years ago, you could argue OpenAI was in the lead, followed by maybe Google that fumbled the bag there since they were the original with the transformer paper, but still credible, and Anthropic, which had many of the early OpenAI best folks. There were two or three players. But then xAI came in and entered stage left, DeepSeek and Alibaba and all these other folks. Now, I don't think there's a world where, because you have seven to eight pretty top quality model providers, obviously Facebook with LLaMA is doing a whole refresh on talent, which is in the news. But there are just so many players now that the idea that all of them are creating closed models and shutting the door is extremely unlikely. And customers can always go and use DeepSeek. The Chinese labs are also quite good, and there's very good technology coming out of there. So there are just too many players at this point who are open source, and it was really a shortage of researchers two or three years ago. You could kind of count who had worked on these projects because they were really Google, OpenAI, and starting to be Anthropic. Today, a lot of PhDs are being done in this space. There are a lot of other folks hiring. Facebook, xAI, I would say both have caught up as well to a certain degree. The companies out in China, Mistral in Europe. So I think we're at a point now where there's a lot of AI talent, AI researchers, AI companies that are creating these models. For all of them to turn off and go closed model, maybe we can trade that on Polymarket or something, but I would put the odds at sub 3% there.
Nathan Labenz: Yes. The dynamics do change a lot from-
Marc Bhargava: Totally.
Nathan Labenz: ...one to two to-
Marc Bhargava: Yeah.
Nathan Labenz: ...eight.
Marc Bhargava: Exactly, yes.
Nathan Labenz: Yes, there are some other interesting factors around safety and reliability. There could be some other factors that could make it shift. But I do think, on a purely competitive market basis, there is a strong incentive to try to win the API market. So there's always an incentive, even if you have something better in reserve, to leapfrog your competitor and try to win share. Why are you telling this story publicly? The first rule of having a great investment strategy is to keep it to yourself. Like,
Marc Bhargava: Yeah.
Nathan Labenz: You're coming here and talking to me, you're just talking to a bunch of AI obsessives, so maybe there's a talent angle. But I've also seen you on CNBC. How do you think about the public communications you're doing?
Marc Bhargava: Yes. We've decided to go public on it for two major reasons. The first is to continue to attract amazing founders. We want people who are leaving applied AI programs or research labs and thinking about building a company to consider working with General Catalyst and to think that we are the firm that can help you go to market in the most diverse ways. We are hyper-focused on how we work with the best people, and we think that's really the thesis that will win out. We need to get the message to the best people, and a lot of them are watching this show or watching CNBC or-
Nathan Labenz: Oh, that's right.
Marc Bhargava: ...other places. So the talent angle is definitely important, and I wouldn't minimize it; it's the most important part. We can't incubate and build these companies and do this strategy without the right people. So that's one part of it. A second part of it is to push back on the doomsday AI scenario or the private equity consolidation play and tell people who may be wanting to sell their company that there is a third way here as well. You could sell to an AI-enabled rollup that believes in compounding, that doesn't want a three to five-year flip, but really wants to give AI to your workforce, gain more customers, have you own shares in what we're doing, go public, and be compounders like Transdigm and Danaher. So I think a lot of the press is also honestly directed at the companies and the people who are thinking about maybe selling their business to a private equity firm or to a strategic. We would just say, 'Hey look, there's a third way here, and we'd love to tell you more about it and talk to you more.' At the end of the day, our secret sauce is essentially incubating companies with the best people and buying the best assets. So I think it makes a lot of sense to tell this story.
Nathan Labenz: That could be a great place to end. Is there any other aspect of a positive vision for the future? I always say the scarcest resource is a positive vision for the AI future. Any other aspects of your positive vision that we haven't touched on that you'd want to share?
Marc Bhargava: No, I think that's generally it. There's a good and fair question around whether so much investment being done in this space will really pay off. Maybe we haven't seen that yet. Even last summer, about a year ago, it felt like a lot of the models were flatlining. I'm sure you were covering that as well, but we weren't really getting many interesting updates. Then a lot of that changed honestly at the end of last year with the reasoning models; there was more context. We've now seen the deep research product from Google, OpenAI, and Anthropic releasing their own versions. So I would say, on one positive note, for a while it did feel like there was a lot of spending and maybe fewer results, even in the productivity of the models. Today, we're seeing a re-acceleration of improvements, which you also referred to. So I think that's really exciting. It's a very cool trend to be a part of, and I'm really appreciative of folks like you for covering it. Especially for engineers and students in university or computer science, or those thinking about what to do next, I think having more and more people look into AI, especially applied AI, which I'm so focused on, is extremely exciting.
Nathan Labenz: This has been excellent. I really appreciate your time and all the deep-dive answers. Marc Bhargava from General Catalyst. Thank you for being part of the Cognitive Revolution.
Marc Bhargava: Thanks for having me on.