How Harvey is reinventing legal work with AI

0:00Starting a business comes with its share of ups and downs, which is why staying true to your vision is essential. a non-negotiable for Romeo and Milka Bregali, Capital One business customers and co-owners of Ra's plant-based restaurant in New York. Romeo and Milka took a leap of faith when starting their own restaurant, gutting an empty space and building it from the ground up, every pipe, every wall, every detail. But building from scratch came with a heavy financial burden, which is when they turned to their Capital One business card. With the flexibility of the card's no preset spending limit, they were able to spend more and earn more rewards while bringing their vision to life.

0:36Today, Raz's success is proof that with passion and the right support, it's possible to make your dreams a reality. Learn more at CapitalOne.com slash business cards. If you've spent any time building AI products or leading technical teams, you know this. Transformation doesn't fail because of ideas. It fails because teams can't move together. Enter Atlassian's Teamwork Collection. It has planning in JIRA, documentation in Confluence, video updates in Loom, and now AI agents in Rovo, which connects the dots across your work so nothing gets lost. It's one AI-powered teamwork platform designed for how modern teams actually build.

1:21Learn more at Atlassian.com slash TeamChanger. That's A-T-L-A-S-S-I-A-N dot com slash TeamChanger.

1:38When you picture a good lawyer, the kind of lawyer who you'd want on your team, this guy might come to mind. Guys like you always think other people get lucky. I don't get lucky. I make my own luck. That's Harvey Specter, the charismatic, sharp-witted lawyer with a heart of gold from the TV show Suits. So when Gabe Herrera and his partner were coming up with a name for their new AI-powered legal services company, Harvey seemed like a good fit. So you guys named the company after a character from the TV show Suits, which I don't watch, but is that true? Yeah, I think loosely based on that. And then I think we also joke that it sounds a lot like Harvard and I think kind of that name association.

2:26But for Gabe, the company name is about a lot more than its namesakes. It's about the impact. I think what was most important for us is, I think even then we thought that this technology would be personified. Like with our product now, a lot of people say, hey, did you ask Harvey? And I think that is like, as these models get more and more powerful and more human-like, I think you're seeing more of this with companies kind of naming themselves that way. In just three years since their launch, Harvey is now valued at$5 billion and boasts a client roster of some of the biggest law firms out there.

3:01And for good reason. They're totally disrupting the legal landscape with an AI platform that supercharges the work lawyers do. On this episode, we're talking to Gabe about how Harvey supports lawyers using AI, building a scalable AI company, and the shifting legal playbook. I'm Rana El-Khalyubi, and this is Pioneers of AI. A podcast taking you behind the scenes of the AI revolution.

3:37Hi, Gabe. Welcome to Pioneers of AI. I'm so excited for this conversation. Thanks so much for having me. So before Harvey, you had a career as a researcher at DeepMind and Meta, and then you kind of had the opportunity to co-found Harvey. What was that aha moment? Yeah, so when I was working at Meta on their large language model team, this was around GPT-2, GPT-3, you kind of saw these models getting better. And in the past 10 years before that, I had always been thinking about what is the right application of AI to the world and thinking about startup ideas. And at the time, my roommate and now co-founder Winston was working at a large law firm, and he showed me his legal tech, his legal workflows.

4:22I showed him these models. And as we started playing with it, the more we dug into it, it just seemed like this was one of the perfect applications of large language models. You know what's really interesting about this is sometimes founders start a company having a personal connection to the problem the company's solving. In your case, you have a background in technology, and then you saw this massive business opportunity in a field that you weren't really an expert in. I mean, Winston is kind of the domain expert. What was that like? Yeah, I think I had always been very industry agnostic when I was doing AI research.

4:57I think when I was doing research, you always had to be curious about how do humans solve problems in a bunch of different domains. I had done startups in a bunch of different domains. So for me, when I was thinking about what type of company to start, it was much more important of, is this the right industry? Will this have a large societal impact? Then am I an expert in this specific domain? But I think to your point, having someone like Winston, who is an expert on the legal side and really understood kind of how law firms work in the legal industry, I don't think you can do something like Harvey without like a legal founder and an AI founder.

5:32And I think that combination has been super important. I agree when I invest in early stage AI startups, but I typically look for that combo, right? You want the AI expertise, but you also want a fair amount of domain expertise if it's a vertical AI application. So I do think that's a winning strategy. We're going to dig into what Harvey does in a second. But before that, I do think there's still a lot of skepticism about using AI in the legal field. And in particular, as I was prepping for this interview, I kept thinking back to the story from 2023 where there was this lawyer and he cited a bunch of court cases.

6:07But of course, they were all basically AI hallucinating. What's your view on that? Yeah, I think this is still a big challenge. I think a big part of why we partnered with the largest law firms is they are actually some of the best companies in the world at avoiding, not hallucinations, but mistakes. If you think of what these large law firms do a great job of is they have these super talented associates that don't have partner-level expertise, but they build these hierarchies of review where associates will do research, they'll draft parts of contract, they'll get reviewed by a more senior associate, another more senior associate, a junior partner, a senior partner.

6:47And so when we sold to these law firms, Harvey wasn't being used as, hey, go write this memo and file it directly. It was being used as help this associate do their work better. And then that still gets reviewed. And so I think that was actually a really counterintuitive thing where we saw this very fast adoption because it fit nicely in their workflows while avoiding, I think, a lot of the like, maybe from the outside problems that you would say, oh, this doesn't cite cases perfectly. And then now as we've gotten bigger, we've started partnering with Lexus, Walters Kluwer, kind of data providers, working with these law firms to kind of improve these systems to avoid those challenges.

7:24But yeah, I think the industry itself is very good at that. And so working with them has been super important. So let's talk about Harvey. What exactly do you do? And maybe personify Harvey AI. Like what kind of lawyer is Harvey? So I would say Harvey started out as more of a transactional lawyer. So law firms are typically split into transactional and litigation departments. And when we started, a lot of the work you do in transactional is I have a lot of contracts. I need to analyze them. I need to go look at previous transactions, SEC filings. And so what we wanted to build was an AI associate, right?

8:01Kind of like a co-pilot that if you're a busy associate, you could delegate these tasks. And what we've been building in the past couple of years is how do you build a single interface, kind of like an IDE that doesn't really exist for lawyers where they can do all their work and collaborate with this associate. That was kind of where it started. Now what we're starting to increasingly do is how do we help teams of lawyers complete entire matters? So if you're working on an entire transaction, an entire litigation, how do we help that team be more effective? And then how do you help the law firm be more effective?

8:34So law firms are managing tens of thousands of these matters at a time. How do you organize them? How do you query them effectively. And then I think the most interesting is how do you work with your clients? And so as we sell Harvey to in-house teams, the immediate thing they request is, hey, we know our law firm is also using Harvey. Can we share data? Can we share workflows? Can we collaborate on the legal work we normally do together? And increasingly, can we do this with other professional service providers as well? That is so fascinating. So, I mean, the legal field is huge, right? There's corporate law, there's immigration law, there's family law.

9:09Are there specific domains of law that you're focused on or is it general? I would say right now our primary focus is kind of the largest law firms in the world and the largest corporations. And so what we don't do is kind of consumer work because I think there's regulatory issues and like the type of legal work that I think most people are familiar with, which review my lease or I need this document drafted for my personal use. We don't focus on that right now. Much more of the work we're focused on is I'm having large antitrust litigation, or I want to acquire this company. Anything that you're doing, if you're kind of a massive company, and that's where we kind of see the complexity where I think right now the off-the-shelf models are quite good at upload a lease and, you know, tell me, hey, is this like a standard lease?

9:58But what they're not good at is, you know, I'm Adobe and I I want to try to acquire Figma. Can you draft me like the merger agreement? And can you do all the due diligence? And so that's where we see a lot of the complexity where you need these like vertical models and products to be built. So maybe it would be cool to bring this to life. Like, let's imagine, you know, I'm a lawyer at one of these big firms. And yes, we're about to acquire a super awesome AI startup. What services can Harvey offer, right? Like, are you doing doc review? Are you doing drafting? Are you doing research? Is it all of it?

10:31Yeah. And so I think it's all of it. And I think it's useful to frame how do these lawyers work before things like Harvey? And it's basically just email and then a set of like legal tech tools. So if you're doing research, you're using something like Lexus or Westlaw. If you're doing an acquisition, you have a data room, which is kind of just a drive with a bunch of documents. And then maybe you have some tools that help you analyze those. But for the most part, as an associate, you're getting an email from a partner that says, hey, I think we did a similar transaction. Can you go in our document management system, find that?

11:03Can you check these terms, see if they're similar? And so I think what's so hard about legal is the work is so text-based that it doesn't fall into these very nice buckets, right? Like research bleeds into drafting, bleeds into all these things. And so a lot of how we've thought about building the product is you want to build it to match kind of the existing product surfaces that they're familiar with? And then how do you merge that with kind of the email experience, which is why I think chat and that type of experience is so successful for lawyers because they're incredibly good at working with language and prompting and all these things.

11:37And so we've seen that part work really well. I imagine a lot of these law firms, like they're the equivalent of their IP or intellectual property is actually in the history of all the cases that they've done before, right? Like if They've done 100 acquisitions before. It's all the data rooms and all the documents from that. How do you incorporate that into a custom AI? I think the most valuable data for the most part is still in the partner's heads and actually in the law firm's emails. And so it's like a lot of the final work product that these law firms file is public, right? So if you do a public merger, if you do a litigation, those final documents get filed publicly.

12:15And so those, to some extent, are in the foundation models. The thing that isn't that I think now is the very valuable data is like the reasoning traces. Like, why did you draft that motion that way? Why did you draft that transaction that way? And then I think the very big challenge is obviously the naive solution would be go take all their emails and all of their documents and just fine tune some model with them with all this data. And obviously the reason you can't do that is like a lot of this data is client data, right? if I'm working on an internal investigation for Walmart, a lot of that data is derived from the emails of Walmart, which I can't train on.

12:50And so a lot of the research and technical problem we're solving is how do you go to a law firm and take all of their client matters, all of their internal know-how, and help them build some model that respects confidentiality, privilege, the ethical walls, because a lot of these law firms are working for competitors and that data can't touch. And so I think there's a bunch of legal specific, and then also just general privacy specific of like, how do you train these models with very sensitive data? In a minute, we go behind the scenes and dig into the AI models powering Harvey. The key, customization.

13:27Stay with us.

13:42Are you curious about the hidden side of everything? Then I have a podcast for you. I'm Stephen Dubner, host of Freakonomics Radio. Each week, we hear from some of the most fascinating scholars and thinkers as we tackle big topics like how whales became the face of environmental activism, how to succeed at failing, and whether public transportation should be free. Go ahead. Listen to Freakonomics Radio wherever you get your podcasts.

14:14Harvey isn't building foundation models from scratch, right? Like you're using existing foundation models, but it sounds like you're building an additional layer on top that is customized based on every law firm's private data. Is that correct? Yeah, I would think of like we don't pre-train models. And so we will use, you know, pre-train models from the foundation or cloud providers. We do post-train and build kind of all the brag and agent infrastructure to make our general purpose product better at legal. But where we see the really exciting part is how do we help every law firm build their own custom model and custom workflows with all their data?

14:54Because I think the biggest question we get from law firms now is, you know, everyone is buying Harvey, how do we differentiate? And I think the same way, if you look back, you know, everyone bought Lexus, everyone bought Westlaw, you don't differentiate by having a case law subscription. You differentiate by having the best litigation partners and the best litigation associates. And I think by analogy, the way these law firms will differentiate is how well can they build custom solutions, custom models, innovate on kind of how they deliver services and even price and their business model and things like that.

15:26Are there any applications that you have a hard line against either because the technology isn't there yet or there are just too many risks? I think probably where we draw the line is we don't give legal advice, for example. So we think of this as a tool for lawyers. And actually, when we were first starting Harvey, something we did was we went on r slash legal advice, which is like a Reddit subcommittee where people ask legal questions. And we just downloaded a bunch of those and had 3.5 and 4 answer these questions. And then we went and just gave those answers to lawyers and said, would you give this to one of your clients without editing?

16:06And it was something like 80, 85 % said yes for a lot of these answers. And so I think the opportunity there was obviously like quite big of like democratizing access to legal services. And so I think that's something we're interested in. But in terms of like core business right now, it's very much like how do you augment lawyers? But I think there are some things we're starting to think about and we're working with some court systems largely outside of the U.S. of how can we provide this to pro se litigants or people who can't afford lawyers to give them advice on here's how you navigate our court system or in these small claims cases like can we help with this?

16:42But I think that that area is challenging and you need to be very careful. It's fascinating. That's a kind of direct to consumer model, right? where if I actually I'll give an example my daughter is 22 and she just signed an agreement with her first job ever and she uploaded the contract to chat GPT and she was like are there any red flags here and I was like well let's also get proper legal advice here and so we retain kind of you know just legal counsel and it was interesting she actually iterated on the contract with chat GPT and I mean I think on the one hand this is really cool because it does democratize access to legal counsel, especially for people who can't afford it, right?

17:20Legal counsel is, for the most part, pretty expensive. But it's also a little scary, right? Because you're delegating these important decisions to an AI. No, and I think this is why, rightfully so, the legal industry is regulated and you need a bar to give legal advice. And I think we're starting to think about, and I think the industry needs to think about it, it's like there will be some use cases where you want to do this. There'll be some cases where you don't. But to your point, the challenge is like everyone has Chachipiti and so the boundaries get very blurry. But for corporate, I think this ends up being kind of fully separate, which is nice.

17:53You know, my thesis around AI is what I call human-centric AI, which is basically AI that augments and amplifies human abilities and not replace it. It sounds to me like Harvey's not designed to replace a lawyer, but augment what a lawyer does. Was that kind of intentional? Yeah, I think like when you look at the work that these large law firms do, they are doing some of the most complex knowledge work in the world, where if you think of something like Microsoft acquiring Activision, just thinking about how you structure that transaction. I don't think the models are anywhere close to doing it.

18:29What I do think happens is there is parts, pieces of the work that the models do automate, right? And there's even full end-to-end tasks that these models will be able to do, right? Like already internally we have systems that automatically negotiate our NDAs. But I think to me, a lot of it is how do we move humans to doing the things they can uniquely do? And when we think of like, what are the most valuable lawyers? It's these partners that are these like strategic counselors to these companies where you have this super complex litigation, you have this super complex transaction, and they can tell you beyond what is in the documents.

19:05This is what you need to do if this is what you want. And so how do we enable that? Yeah, I think of my days when we were selling my company Affectiva and we ended up selling it to a Swedish publicly traded company and our counsel was Jay Hatchigan at Gundersen. And honestly, you're absolutely right. Like, yes, there were all a ton of mechanics that had to get done, but he was so strategic. And it's a result of like 20 plus years of experience doing M &A and right. And it's also the personal relationships, right? Like most of these legal problems don't have objective answers, right? Like the outcome you want in that transaction is really different than the outcome maybe someone else would want.

19:46And I think that is still what these like very good lawyers are uniquely good at. And we want to free up their time to do that instead of like, hey, let me read all these red lines, for example. You also talked about how like some of your customers are the law firms and they're using Harvey, but sometimes it's a big organization like Bridgewater and they're using Harvey too. And then the two Harveys can talk to each other. How does that work? What's an example of that? Using Bridgewater and I think just generally the like private equity and financial services firms, one very big use case is fund formation.

20:20So I'm an investment firm. And I mean, you probably did this as well. It's like, I want to raise a bunch of capital. I need to make a fund and structure it properly. And I think what you see is there is a lot of work in fund formation that is not profitable for law firms. It just gets written off. Oh, interesting, because now I have a little bit more empathy for our law firm because I'm like, this is like so expensive. Yeah. Oh, I mean, I think this is where the like dynamics get really interesting, where it's like I think a lot of people think about the billable hours. Because obviously in the legal world, the business model is the billable hour.

20:54But how does integrating AI change that? So I think we're still in the early stages of these law firms thinking about how they change their business models. I think with PwC, we've talked with them a lot and they're kind of, I think, ahead of the game in terms of thinking about, okay, we need to provide services in this new way. I think some interesting things we've seen, probably the most interesting thing is there's some litigation firms that are just contingency based or fixed fees. So it's purely I'm defending you and if I win, I take some percentage of that win. And there the incentives are very clearly aligned.

21:27What's the business model for you in that case look like? I think we're starting to think about can we take a percentage of the client matter, for example. And so if you use Harvey on this litigation, Harvey takes some percentage of that and we automate some percentage of that work. I think where we see this going long term is the reason you need billable hours is this work is so complex that it's, I think, until now impossible to price. It's impossible to estimate how long this is going to take, all the things that are going to come up. But I think now these models are getting so good where you can actually do much more accurate pricing of legal work.

22:05And so you can move to more like value based, fixed fee based pricing. I think there's kind of like a win, win, win for like Harvey clients and law firms. You know, I'm an investor in a number of companies where they basically it's almost like an AI staffing agency. So instead of hiring, I don't know, a human health care administrator, you're hiring the AI equivalent of it as almost like a fraction of a full time human head. Would that be a business model where you're charging for Harvey as, you know, the equivalent of an associate or, I don't know, a paralegal or something? Yeah, I mean, I would say this is kind of what like our enterprise SaaS model right now is like we charge seed-based pricing.

22:45And then I think a lot of the question is like, how much of that value are you able to charge for or capture? And I think it's like that analogy is very useful. Like a lot of times in building the product and these things, it's like very useful to analogize it to human labor. And then I think there'll be very unintuitive things about this technology that don't map to that. and you need to think about like where the intersections are. Can you talk about what the Harvey team looks like? I mean, is it mostly like machine learning engineers and software developers or is it also like a lot of lawyers?

23:19Like often in these like vertical AI applications, a lot can get lost in translation, right? Because if I'm a machine learning engineer and I know nothing about law, like how do you actually understand the pain points of your customers? So how do you deal with that? Yeah, we think of building the team a lot like kind of Winston-I's relationship where you need kind of the technical expert and then you also need the domain expert. And so both on our sales team, we have kind of a traditional like GTM org, but we also have a bunch of lawyers from top law firms that were associates, senior associates that help sell the product, map kind of workflows of these law firms into the product.