Welcome to the Prof Markets founder series. I'm Ed Elson. Since the start of 2026, investors have been asking one big question: how much of the economy will AI disrupt? We've already seen the fallout in software, where waves of SaaS apocalypse sell-offs wiped out hundreds of billions in market value. But that may just be the beginning. Now attention is shifting to the legal industry, a trillion-dollar market built on manual, time-intensive work like document review, due diligence, and compliance — exactly the kind of workflows that AI was built to transform. Well, my next guest saw that as an opportunity. In 2022, he built an AI company designed to streamline legal work at scale. Now, the question is, can it actually disrupt one of the world's most archaic industries? Well, with over a billion dollars raised, an $11 billion valuation, and adoption in over 60 countries, it's well on its way to doing just that. This is my conversation with Gabe Pereyra, co-founder and president of Harvey. Gabe, welcome to the show. Thank you so much for joining me. So much to get into here. I guess maybe lay out for listeners what actually is Harvey, and how did this company get started?
Thanks so much for having me. I would think of the problem that we're solving with Harvey as: how do we help large law firms and their clients, which are large enterprises and increasingly all law firms and all companies, go through exactly the transition that you talked about. So when we started the company four years ago, we built, I think, what most companies built of some form of co-pilot for a professional. Cursor and Cognition built this for programming. We built this for legal. And I think what you're starting to see the shift as these models get better and better is you need to start thinking not just about the productivity of individuals, but the productivity of entire organizations and what is the infrastructure that they need to be able for the entire firm to operate effectively. And so that's a lot of what we're building at Harvey.
When did you come up with this and how did you realize that this was going to be a huge opportunity?
Yeah, we started the company summer of 2022. I was at Meta on their large language model team. GPT-3 had just come out. I had been doing AI research for about 10 years before that. And so even in 2014, I think I was at Google Brain, then DeepMind. A lot of people in that community had the belief that we would figure out a way to build these systems. I don't think the path was super clear. I think we're kind of still being surprised by the way it's going. But there was just like I had this strong belief of we will be able to build superintelligence, AGI, things like this. And I think when you do a lot of that research, you're constantly thinking about what problems are the models good at, what problems are the models not good at. And in 2022, my roommate was Winston, who's now the CEO of Harvey. He was a lawyer at O'Melveny. And I'd been brainstorming startup ideas as I saw these language models get better. And one day he kind of showed me the work he was doing and the workflows. And that was kind of the light bulb moment where at the time GPT-3 couldn't do that work, but it was very clear they would keep getting better. And that felt like one of the big industries that would just be a very clear application of this technology.
It does seem as though legal work is basically ground zero for AI, or at least that seems to be the way people view it at this point. I mean, what we've seen is that AI appears to be able to do all of the grunt work of the white-collar jobs. That is kind of like the starting point of what AI can do. And it does seem as though law is exactly that. Just for those who maybe don't understand how law firms work and what kinds of work they're actually doing, could you tell us a little bit about the legal industry? What are the workflows of a lawyer? What kinds of things could AI potentially automate in that business?
I think when most people think of law, they think of consumer law. And so I need to review a lease or, you know, I need to kind of look at one document. And the models are obviously great at that. And I think to a large extent the base models can do that. And then there's kind of corporate legal work and particularly big law, which is the massive law firms. And you can think of the work they're doing is the highly specialized legal work where you need these incredibly talented, very specialized partners. And I think kind of the two best examples of this are: you're doing a massive merger, right? Or an acquisition. You want to go buy a company for $10 billion, $50 billion. Like you need the highest-tier partner to advise you how to structure that transaction. Or you're doing a bet-the-company litigation, right? Like a big antitrust or something like this. And the way that I would think of the workflows of all these firms is these projects take thousands, they can take tens of thousands of hours, teams of associates. And the big challenge is, for example, when you're going to buy a company, you need to go understand all of the contracts in that company, all of the things that are going to happen when those contracts change because of the merger or the acquisition, all of the legislation around it. And then there is also all of the negotiation dynamics, right? It's a semi-adversarial or could be an adversarial process. Same on litigation. And so to your earlier point of I think when language models came out, legal was kind of this great application where typically your workflows as an associate is you're getting emails from senior associates or partners and they're just giving you tons of tasks. Go research this. How do I write risk factors for this document? And what these associates are incredible at is they can just absorb these tasks. They know how to go use all these tools and solve all these problems. And that's kind of what you're seeing these agents starting to get better and better at. The part that I think is so difficult about these legal workflows and similar to programming is the boundaries between the tasks are super blurry. And so it's not easy to go to a law firm or go to a programmer and say, 'Hey, the coding models or the legal models can do this and humans can do this.' Like the boundaries are very blurred and the work is so complex that a lot of the challenge we're working with law firms is how do you rethink your workflows and what humans and agents should be doing when you're working on these large projects.
When you talk about those boundaries, that's an interesting point that the boundaries are blurry. Are you saying the boundary between what AI is best at versus what humans are best at? That it's hard to draw a distinction between those two things?
I think it's both that, but also the distinction of even what do you delegate to humans. When you're doing a merger, you kind of have a pyramid, right? You have like a senior partner, some junior partners, senior associates, junior associates. There isn't like a concrete rule of when I'm doing a merger, this task goes to this associate. And there isn't even a concrete rule of what defines a task, right? Because it's all text-based. And so it's a partner just saying, I need you to figure out XYZ. And that could be something super simple like go look up this one case and tell me who is the other party in it, or it could be something super complex that is like go write the first draft of the merger agreement, right? But even that mapping, and this is kind of what you see with ChatGPT, where when we started the company, people would be like, what does your product do? And it's kind of the same as asking people what do you do with ChatGPT. It's like everything, but it's really hard to define why you do something one way. And this is exactly what makes the transformation so difficult because to your point, given this kind of such an open natural language shape, how do you start defining the boundaries of this is what models are good at? Because it depends how you prompt, it depends which model you're using, it depends on the agent harness. And so there's just this massive challenge of how do you organize all this work in this new way given the models can do some stuff but they make mistakes in ways that aren't intuitive. And so it is this huge change management problem and like upleveling problem for not just legal, like all these industries. Like you're seeing the same thing in programming right now.
Yeah, it does seem as though the great thing what ChatGPT enabled us to do is ask questions that are actually blurrier and that cannot be answered in binary. That cannot be answered. I mean, it used to be that you had to spend a long time trying to phrase your question for Google search very, very specifically. And then what was kind of remarkable and liberating about large language models is that you could be a little bit more blurry and rough and it would be willing to go to those more ambiguous places and try to come up with more creative answers to more complicated questions. So on the one hand, I kind of think, well, that's exactly the strength of AI. So maybe this is exactly the place where AI should thrive. But then at the same time, you're also pointing out like there are places where it still gets confused. I mean, large management work is still actually very complicated and it's a lot more complicated in a large organization versus when you're just operating as a single individual trying to figure out questions on your own time. I just want to point out for people who might be listening because I mean there are AI startups for everything now and I'm sure there are probably hundreds of legal AI companies that are trying to eat your lunch at the moment, trying to compete. I would just note for people, I mean, from my understanding of the AI industry, Harvey is the number one AI company in law right now. You guys hit $190 million in ARR in January. That's the most recent number we have. If you want to update it, go ahead. That was nearly double what it was five months earlier. So, you guys are growing incredibly quickly. You are partnering with basically all of the biggest corporate law firms. You guys are kind of spearheading this transition. I guess the question then becomes: when you went to these law firms and you said we can do what you guys do with computers, what did they say? Were they excited about that? Were they scared by that? I mean, how did these big corporate law firms react when you went up to them and made the pitch?
So, I mean, the pitch is definitely not we can do what you can do with computers, but I think what helped early on was we found kind of certain partners or innovation leaders that AI isn't new to law firms. They had been using things like TAR and other kind of AI technologies to do parts of legal work. I think this was just such a large step change. But early on, for example, our first client was A&O and David Wakeling there. When we showed him kind of we got early access to GPT-4 and we built a product around that and showed that to him. He just had this light, the same light bulb moment we had where he's like, 'Oh, this is going to change how we do work.' And I think a lot of our pitch to law firms has been there will be parts of the work you do that these models will do. The same way now when you do discovery, you use TAR and you use contract attorneys and you don't use associates. So that's going to happen. But there is also going to be a lot of work that these law firms do that these models aren't going to do, right? Like I don't see a world in the next 10 years where you're doing a large merger or large litigation and it's fully automated, right? Both for technology reasons but also for regulatory, insurance, all these reasons. And so a lot of the problem we want to work with law firms to help solve is what is the future of their business model going to look like, right? Because there are parts of this technology where you are selling expertise on an hourly basis. Like there are parts of this that it will be complicated to figure out. There are new ways to collaborate with your clients. And so there's just going to be all these questions that this technology is going to raise for law firms, for all professional service providers, for most companies. And so I think a lot of the pitch is just we want to be your partner and help you think through this entire transformation, not just the technology.
What you're essentially saying is, you know, it's what we can do is powerful, but not that powerful to the point where we're going to automate everything and basically eliminate all of these jobs. And you mentioned that there are technological constraints, regulatory constraints, and also insurance constraints such that you couldn't just automate a big complicated legal contract with your product and with AI. At the same time though, someone like Dario Amodei, who is leading the frontier of large language models over at Anthropic, is also saying that we could see roughly half of white-collar jobs wiped out over the next several years. And so there is this tension where it's like, you know, on the one hand, a lot of people in AI would say might make the case that actually no, you can do that. And I think a lot of people maybe listen to what you're saying now and they might be thinking you're trying to tell us that it's not going to be that bad so that maybe people would be less afraid of your technology, less afraid of your product, maybe root for your product a little bit more. So I guess to press on that, what exactly are those constraints? Like why couldn't you do this all with AI?
I think the biggest is just the change management. So to be very clear, I think the technology is good enough now that it's like maybe the timelines Dario is talking about, like I roughly agree with those, but I think it's going to be close. It's going to be somewhere between what he's saying and like self-driving cars where it's like self-driving cars are better than most humans at driving, yet they're 0% of the cars on the road and they've been better than humans at driving for 5 years, right? It's like it's clearly harder to roll out something like self-driving cars than this digital technology. But from the law firms we work with and the large enterprises we work with, I'm still skeptical that like in a year or two if you're working with like a large regulated bank, right? It's just like I don't see a world where you just deploy these code models across the entire bank and say like the regulatory agencies just won't let you do this. And so I think there's going to be challenges like that. I think it's not clear to me how it plays out in terms of like a lot of these arguments you could have made with computers and with the internet and it's like there was a huge class of jobs like librarians and all these things and it's like oh now you don't need this. And I think with like a lot of the legal work that gets done, there is stuff that is just not digital, right? And so I like from a capability perspective, I think I'm actually like I did AI research. I'm pretty aligned with like Dario and like the way these things are going. And I think to me the biggest thing that I see in the legal industry is people don't think this is going to be as serious or it's going to happen as fast as it is happening because we're seeing this happening in engineering right now, right? Like these models are getting so good at programming that they are better than most human engineers, right? And to me, that's like there's no clear research blocker that that's not going to continue and it's going to keep speeding up because I think that research path is pretty clear. And then a lot of what we're doing is how do we take those models that are really good at programming and like translate them into legal. And so I do think we will build systems that have the capability of like automating large parts, most parts of transactions. My point is just there's parts of that that are not technical problems, right? Like if you think of a negotiation, there is a part of a negotiation that has nothing. It doesn't matter how smart you are or how technical you are, right? It's like it is a human-to-human thing and it's like when you meet the GCs of a lot of these large companies, it's like when is the GC of a large private equity firm that is raising a $20 billion fund going to be comfortable with AI automating that entire fund, right? It's like the downside's so big, it's just not worth it, right? It's like, okay, maybe you can do that 30% cheaper, but the cost of a mistake is like that fund is structured incorrectly and now I just lost $2 billion. It's like, I'll just have one of these law firms do it and I'll have the partner review it, but hopefully they use AI for parts of it. And so, I think there's just enough structural things where I think a lot of people think about legal as can you review this contract? If you can do that, then we've reached automation. But like what these large law firms are doing and what these complicated regulatory industries are dealing with is so much more complicated than just purely like a capabilities intelligence problem. And so I think there will be things like that that this is definitely going to happen on some timeline. I just don't think it's like in the next 2 years.
We'll be right back.
We're back with Gabe Pereyra. I think the autonomous vehicle comparison, the analogy is a good one because it's true. It's like self-driving cars are technically better than humans at driving right now or at least certain technologies. We'll go with Waymo for example. But at the same time, whenever there is an accident, it seems to be far more abrasive and scary and more concerning than when we see an accident which happens at higher rates among the human population. And it sounds like maybe what you're describing is the same system where I mean you mentioned like a GC, one of the main roles of that job, one of the main responsibilities is negotiation. Would you really let an AI do that for you? I would argue maybe you would, maybe it might be possible that actually the AI is better at negotiating. But it sounds like what you're saying is that if the AI were to make a mistake and if it were to eviscerate however many millions of dollars for the firm, that would be more difficult and more concerning and more of a problem than if a human being were to make a mistake and therefore we might be slower to roll this out than a lot of people think. Am I getting that right?
Yeah. Yeah, I think it's like the accountability is one piece. I think the other thing is like so to me the argument against why this goes faster than self-driving cars is like the obvious, you know, beyond regulatory for self-driving cars. It's just you need to build all these cars and it's like it's hard to get all of them on the road. And so this will go faster because it's digital and you don't have that restriction.
And I think to me the biggest challenge is actually just how correlated all of the risk of adopting this technology all at once is. Like if you think of what's so nice about self-driving cars is even for people not working in AI, it's quite intuitive what a self-driving car is, right? Like you're just like, 'This drives, it didn't crash. I can kind of evaluate it. Accidents per mile makes a ton of sense.' But now if you think of a bank for example adopting this technology, these agents not just in legal across the entire bank at the scale you would need to do the disruption you're talking about. You're taking on all of this correlated risk on this technology that makes mistakes in ways that isn't intuitive and it's like company destroying if you take this risk incorrectly. Right? And the same reason that self-driving cars freak people out, it's because they cause accidents in ways that you don't anticipate because you're like, as a human driver, I wouldn't make that mistake, even if statistically it's safer. And I think there's just going to be so many variations of that where all of the second-order effects and things like this. And if you think of legal, it's this is like the underpinning of our entire society. And it's just like, does anyone understand the entire legal system and all the implications well enough that if you just automated this whole thing in the next two years, we'd be like, that's going to go well. It's like it's just too complicated. And so to your point, will NDAs be fully automated in a year or two? Definitely. Will merger agreements for multi-billion dollar mergers? Like, I would be skeptical of that. And I think that's also the hard part of legal is there's just such a massive spectrum, right? And like that's a lot of what we're helping law firms and enterprises figure out where it's like you want to roll this out slowly. You want to start using it on the low-risk use cases so you can build this mental model. You want to find pieces of the higher risk that you can take out. But it's not this super intuitive just like here's a bunch of legal agents, all your problems are solved. Like it's the same way if you hired a thousand lawyers and you'd never run a law firm, you're like, I don't know how to manage them. And I think that's the big challenge everyone's facing.
It does seem a huge problem for AI that AI does make mistakes and it still seems to make mistakes like relatively frequently. And on top of that, your point about an NDA versus a multi-billion dollar merger agreement. Humans make mistakes, too. But at least you can tell a human, hey, you cannot mess this up.
Yeah. Like you better get this right because there's billions of dollars on the line and if you don't get it right you're fired or maybe you're going to get into some legal trouble. I mean you can really explain to a human the stakes of the problem and have some more assurance that they're not going to make the mistake. Seems like you can't really do that with AI at the moment. That's one of the biggest values that I think people underappreciate of law firm partners, right? It's like at the end of the day when you're doing one of these litigations or one of these mergers, there is like a single human being that has spent the past 20 years of their life doing all these similar transactions and being held accountable to that. And they're willing to bet their entire career that they're going to do that correctly. And it's like that is a level of trust that I think we will get there with these systems, right? It's like the stock market runs on like an automated financial system that like we now all trust. And so there's ways to get there. I just think there are a lot of problems we need to figure out to get there.
