美国律师时薪352美元，Harvey把4000亿美元的法律生意讲透了

id="podcast-transcript" data-framer-component-type="RichTextContainer" style="transform:none">Winston Weinberg: Something that’s really important in professional services is prestige and trust, right? The reason prestige is so important is because trust is the most important thing in professional services, right? And so the reason we went after the larger firms is if you earn the trust of a few of those firms, the rest of them will trust you, and the rest of the firms downstream will definitely trust you, right? And their clients will trust you, right? So I think that something—you know, something that we thought about doing in the beginning was, well, just go straight to enterprise, right? And there are a bunch of problems with that, but one of the main reasons is there’s just no reason for them to trust you, right, so you can actually build these systems. Pat Grady: Greetings. Today on Training Data we have Winston Weinberg, co-founder and CEO of Harvey. Harvey has occupied a special place in the AI ecosystem over the last two or three years, becoming the canonical example of what it means to build an application layer company on top of the foundation models. A couple years ago when these companies were derided as wrappers on top of GPT models, Winston and his co-founder Gabe realized that this is where a lot of the value would actually be created, dealing with the messiness of real world problems, not replacing human beings, but giving them superpowers. And so over the last couple of years as much of the industry has taken a “tech out” approach to build their business, Winston and the team at Harvey have taken a “customer back” approach and now stand positioned to revolutionize the legal industry, which is $400 billion in the US alone, the same size as the global cloud markets. We hope you enjoy. Winston, welcome to Training Data. Thanks for joining us on the show. Winston Weinberg: Thanks for having me. Beyond the GPT wrapperPat Grady: All right, so you have built, you and your co-founder Gabe and your team have built a company that has become to many people sort of the canonical application-layer AI business or sort of the defining application layer AI business. But when you started back in July of 2022? Winston Weinberg: Yeah, end of July, 2022. Pat Grady: So when you started in July of 2022, pre-ChatGPT nonetheless, this wasn’t really a thing, and as soon as it became a thing, people looked at this whole category and they said, “Eh, these are just wrappers on top of foundation models. There’s no value to be built here.” So they’re immediately sort of derided. So I guess the question is as a starting point, what gave you the conviction that this was the right business to go build, and that there would be value between foundation models and customers? Winston Weinberg: Yeah, I think the simplest answer to this is quite literally that these industries are incredibly messy, right? And so I think that the largest misconception with kind of the GPT wrapper companies was a lot of what you were starting to do didn’t have, like, a massive delta between what the foundation models provided and what your small company could do, right? So one of the largest advantages that we had in the beginning was how much emphasis our product had on citations. Like, literally just it was very important in the legal industry to be able to cite, like, line by line and the accuracy of those citations, right? There are a bunch of other things as well but, like, from day one, that was something we put resources into. If that’s all our company was, if we just said, “Okay, we’re going to be a citation company,” right, then yeah, I think you’re just a GPT wrapper and it’s going to go away. But if the ambition of the company is to actually partner with an industry to completely transform it, this is a trillion-dollar industry. It’s incredibly messy. There are tons of different data sets, there’s tons of ways that people work, there’s tons of different specialized workflows, et cetera. And so I think over time it became clearer and clearer that these models weren’t going to just automate entire industries away, right? They were going to change and, like, fundamentally serve as kind of like the groundwork for changing an entire industry. But I mean, one way to even look at this is when a new feature comes out, we usually think of that feature as this is a great piece that we can put into part of a workflow. So an example of this is, like, OpenAI will release Deep Research, right? And my first reaction to that is, “Oh my God, there’s all of these capital markets use cases,” right? For, like, studying markets, looking at all of this data that’s on the internet, et cetera. It’s not everything a capital markets attorney does, but you could put that into a piece of a hundred-step process in your product and it makes that process better. And so the best way to think of this is literally whenever I see either the foundation models can do something better than they could before or completely new feature, it just, like, unlocks part of the TAM or part of the actual process that you can build on top. Pat Grady: Let’s pull this thread a little bit. So at time zero, you got the model down here, you got the customers up here, and you got Harvey doing citations in the middle. Winston Weinberg: [laughs] For a month or so, yeah. Expanding and collapsing productPat Grady: What came next? Kind of walk us through some of the major leaps in the product, or some of the major value creation opportunities you’ve seized upon along the way. Winston Weinberg: Yeah. So the best way to kind of describe how we’ve been building the product, and this is, I think, one of the things that we’ve gotten the most right, and it has also been the hardest to manage, which is—and from an org perspective and from a kind of like, which pieces do you focus on, et cetera. And the piece here is at all times you have to basically expand the product and then collapse it back, right? And so what I mean by this is actually if all of these models work perfectly and humans were perfect at communicating with each other, the best interface is literally email. Like, that’s it. Like, there is no interface. It’s literally just email. And it’s perfect. It has all of your contacts, it can read your mind, et cetera. Pat Grady: I mean, it could just be a neural link. Winston Weinberg: Yeah, sure. That’s perfect. Yeah, that’s even better, right? But that’s not how it works, right? And I don’t think that we’re just going to randomly one shot be able to do that, right? And even if we could do that on the model side, we can’t do it on the human workflow side. In other words, like, the example I give a lot is even if the models could do—you know, somehow chain all the steps together to do a large merger like the Activision and Microsoft merger, the user can’t send that to the model and say, “Merge, please,” and then it just like does all those steps, right? There’s all of these communication sides on the UI side as well. But okay, so holding that still, going back to the, you know, you have to constantly expand and collapse the product, when you’re expanding, what I mean by this is the chat UI doesn’t work for every use case, right? And it doesn’t work for every use case right now. I don’t think it will for the future, right? And so an example of this is if you were trying to build something that does really good case law research, right? There’s multiple steps to that. You want to build a system that is very good at doing retrieval over all the cases. You want to build a system that is really good at comparing and contrasting all the cases. You want to build a system that is really good at synthesizing the facts in your case to all of the case law, et cetera, right? And if you’re doing that, the best way to do it is you expand your product by building specific vertical—you could call it agentic systems, whatever you want—that do that thing from start to finish, right? And then you collapse them all together. So what does this end up looking like? We will take a bunch of use cases that are very high value, we’ll build out a specific workflow to do that use case, and then we will chain them together so that you can complete a task from start to finish, right? And so the difficult piece of that is if you’re trying to sell something that is seats, you have to sell something that is applicable to as many users as possible. And so you have to balance, are you building something that is really good for a securities attorney, or are you building something that’s really good for all of the attorneys, right? And so that’s kind of the collapse part, where you want to build these specific workflows, agentic workflows, whatever you want to call it, whatever terminology works, and then you want to combine them into the same service level on the product, right? And so what this looks like eventually is you upload a share purchase agreement onto Harvey, right? And we have—the user might not see this, but there’s tons of different workflows that are like extract the reps and warranties from that or summarize it or whatever it is, right?  And we’ve built them all separately, and you can use those UIs and actually do that workflow separately, or when you just upload an SPA, Harvey says, “Would you like to run any of those workflows?” And that’s the collapse version, right? So you’re building these specific solutions and then you collapse them back in. Sonya Huang: How much of the magic do you think is in the models themselves versus what I think you just described as, you know, an agentic system or a cognitive architecture? Winston Weinberg: Yeah. Sonya Huang: And is it the workflow that’s hard? Is it chaining things together that’s hard? Is it the combination of all of it that’s hard? Winston Weinberg: So I think I would categorize this in three areas. So all you need to do for kind of like every single workflow is basically what does the user want? What is the intent, right? After that, it’s—so what do they want? What’s the intent? How do you extract that intent out of them? The second piece is what context do we need? And the third piece is this right? Right? And so my point is there are different systems that work really well for different versions of those patterns, right? So routing, like, using the models to actually do predictions and routing, is really helpful for the first one, right? And so that’s the example of either follow-up questions is really important there. Routing someone’s query to the particular task you think they want to do is very good. And so there’s kind of like an orchestration element there, right? Context is, okay, do you have predefined systems that will search for your internal documents that are relevant to the question and then your external documents are relevant to the question? A lot of that work is retrieval, right?  Like, most of what you’re building there is retrieval, and then routing to make sure that you are accessing the external documents when you need to and the internal documents when you need to as well. And then the third one is, is this right? Right? So going back to my citations example, which is silly, but it’s actually really important, I think there also a lot of the work can be done by the models, but you have to make sure that the models are very good at checking for certain things. So let me give you an example of this. There’s a thing in legal that is what is “market,” right? And the models don’t know what market is. And there’s various versions of what market is. There is market for a particular private equity firm, like what terms they are used to, you know, doing on an LBO or a deal or in a side letter, et cetera. There are terms that are across all of private equity, and then there are just kind of like general M and A terms, right?  And the models don’t have access to this data, right? And so on that third piece, a lot of it is can you build a system that is very good at retrieving all of those different data sets when needed and comparing them against each other? Pat Grady: Hmm. And I’m sure it helps if you have all of those different folks as customers. Winston Weinberg: Yes. Yeah. So I mean, that I think goes to another piece, which is the biggest problems that we have, and I think—I mean, this is a very interesting problem, and I don’t think the models are just going to solve this is the process data for a lot of these tasks doesn’t exist on the internet, right? So the process data for how do you do disclosure schedules or what is market, right? Those are not things that are just kind of like on Reddit somewhere, right? And so what we do is we actually hire domain experts who sit down and say, “These are the steps that I would take,” right? And then you just chain the models on top of that, right? Or if there’s a gap between what the model can do, then you do fine tuning, right? But the best way to do fine tuning in post training is task specific. It’s not here there was a large kind of ethos maybe in the legal industry and a bunch of other industries that somehow if you got all of the legal documents and then you just trained a model on that it would do law, right? And that’s kind of like saying read all your caseload and all of the textbooks in law school and then just put you into, like, a profession and you somehow know how to do it. And that’s not how it works. So much of it is like you actually have to learn how to do the different steps, right? And then the other side of that is not only is the process difficult, but the evaluation is really difficult, right? So you have to hire kind of like mid-level folks to do evaluation for a lot of these things because if the junior folks could do evaluation, they would be mid level or they’d be senior, right? Like, that’s the reality. And a lot of what you do in a law firm or professional services is you’re actually evaluating the work of junior folks, right? It’s incredibly expensive. I’d argue that maybe 20 to 30 percent of the revenue of these places is doing that. Sonya Huang: I want to go back to this concept of expanding and collapsing the surface, which I love. What is your view of the ideal end state for how a lawyer should be interacting with Harvey? Is it kind of that email chat interface, and it’s just merge company A and B and I mean we’re good, or …? Winston Weinberg: I don’t think that we will get there anytime soon. And that is not for me to say that I’m not bullish on the foundation models getting better. I mean, I—you know this, I am incredibly bullish on the foundation models getting better, and we have designed the company kind of with that as a constant push, as a driving force. I think the biggest problem with that is that doesn’t allow the user to have enough kind of—exercise enough judgment on the workflow itself, right? So I think one thing that when people are talking about agents, they’re talking about these tasks that are, like, quite simple and don’t have massively high economic value. When we’re talking about building workflows and putting kind of agents in those, we’re talking tasks that cost hundreds of thousands of dollars, right? I mean, one reason the legal industry is so good for LLMs is if you can kind of think of this as is the industry text based, and then how valuable is a token, right? And a token is incredibly valuable in legal and professional services. If you look at, like, a merger agreement, like a 50-page merger agreement, the token in there, like, each piece of a word is worth so much money if you think about how much it costs to produce. So this is all to say I think the end state is you keep building these agents and workflows, and then you chain them together as much as you can, right? And that makes it so that your UI actually looks the same, but your suggestion and your routing model gets better and better, your orchestration level model gets better. So you can kind of think of this maybe like—let’s go to the law firm. I think that we’re building out the specialized associates that can do different tasks, but it’s also incredibly important that you have the partner or the managing partner operating model as well. And I think that as the models allow us to build more and more of these specialized kind of specific associates, you also have to put all of this effort into actually having the orchestration layer that pulls all that together, right? And so I think our UI actually looks similar in the sense that it does look kind of like a text window, right? But the ability for a user to upload a bunch of documents and then for it just to suggest what you do or the ability to say, “This is what you did last time, would you like to do it again?” things like that improve. Kind of like a colleague that’s just, like, really good at knowing what you want to do. Selling software or selling work?Pat Grady: On that note, are you selling software or are you selling work? Is this legal software or is this AI lawyers? Winston Weinberg: Yeah. So so far, I mean, software. I think that the best way—like, most of our clients in the beginning were law firms and professional service providers like PwC. We now have a lot of enterprise customers as well. And the way that the product is evolving is you can kind of think of it as actually two products. One is a productivity suite, right? And that is lawyers in the loop at all times, et cetera. And, you know, the ROI on that is incredibly helpful. It saves you tons of hours a week, et cetera, right? The other side of that is you’re building these workflows that do part of the work from start to finish, right? And that is closer to selling the work. The way that we’re actually approaching this is we are building those with law firms and helping them get more business, right? So we will get these revenue split agreements with law firms or for professional services, and we will combine their domain expertise with our tech and then they go out and sell it to their clients, right? And so we are transitioning from just a seat-based company to actually selling the work as well. And I think we’ll see a lot of that this year. Pat Grady: What are the ripple effects of that? Kind of like inside the building and outside the building, how do you manage that as a company, and what change does that require of your customers? Winston Weinberg: Yeah, so managing it internally is all about taking bets, because it’s—you know, if you’re building software that is—and every single feature that you’re adding is good for your entire user base, you’re kind of like reducing your chance of failure, right? Because every time you add something to the software, it is hopefully increasing the value that you get from every single one of the users. If you focus on building one of these specific workflows and it doesn’t work, it’s zero and one. Like, it either works or it doesn’t, right? And it’s for a specific use case. The value of it is really high, but if you can’t sell it to people, then you’re in trouble, right?  And so the best way that we’ve handled this internally is we’ve actually let the law firms do a lot of the discovery for us, right? So we will do joint projects with a large company and their law firm, right? And so we will use them as design partners to make sure that this is something that is repeatable. You can actually do it, right? Like, the technology is there to do it. And number three is appetite. So there is a large problem in legal where a lot of kind of what you need is like compliance checks or you’re doing the work for insurance, et cetera. And so the third piece is also really important where you have to make sure that the in-house team is actually okay with AI handling that use case, right? So that’s kind of how we handle it internally, where we spend tons of time on discovery, right? Tons of time talking to customers and we have a lot of design partners for the early stages of that. One more piece about how we handle it internally as well: In order to generalize what we’re doing is we’re building, like, AI patterns. So here are the 15 types of actions in the legal industry that we need, right? So research, like, case law research is a huge one. Regulatory research is another one. Clause extraction is another one, right? And so we’ll put tons of effort into those kind of generic or widespread horizontal things, and then we’ll chain them together kind of like a Ford factory line. So those are the two ways. So on the GTM side, it’s discovery, on the product side, it really is kind of building that Ford factory line of AI patterns. Externally, I think this will have a very large impact. So we’re working with a firm that is almost a hundred years old, and they became famous from literally advising, I think it was King Edward the something on the abdication of the throne. So in other words, these firms are really old. Some of these firms are hundreds of years old, and they have had the same business model for a hundred years, right? And they’re actually willing to work with us and change some of their business model, right? And, like, take a bet on this and explore it, right? And so I think most people know, a lot of professional services, and especially legal, has been the billable hour forever, right? And so the problem with that is if you’re selling them efficiency software, there’s only so many hours in the day, right? Unless you come up with a software that invents the 25th hour of the day—which, by the way, would be the best thing you could do to sell to professional services. But if you can’t do that, efficiency is a hard sell unless you can convince them that there are ways for them to actually transform their business, right? So what we’re starting with is let’s help you take a bunch of these workflows that you normally do at not a very good cost or a loss, right? And let’s turn those into software, and help you kind of spread out and get more market share in areas that you don’t normally have. So I’ll give you a pretty good example of this. In private equity, one of the ways that law firms go and kind of get new private equity business is they’ll do things like side letter compliance or kind of like lower end work. And they’ll do that at a loss so that eventually, you know, the private equity firm will pay them for the LBO or whatever the big M and A deal is or something like that. And so we’ll build them software to help them get that work and not operate at a loss to get it, charge for a flat fee, et cetera, and then take some of the cost. Sonya Huang: Is the tech ready? Like, how much, I guess legal work can already be automated with today’s models? And if you were to freeze model development, how much can be automated? Winston Weinberg: Yeah. I mean, I think that’s a good question in terms of—and especially the second one of I guess, like, how there’s the reasoning ability of the models and then the capability of the models, right? And if you froze the model’s reasoning ability, I think we’d be fine, actually. So if we somehow kind of like froze the ability for the models to process data, that doesn’t mean that they know which process to actually analyze that data, but just their ability to ration over it, to make rational decisions over it, we’d be in a really good spot. So in other words, I think we’re at, like, small percentage points right now of, like, legal and professional services, but if you paused, we would increase pretty high even if the models didn’t get better, right? Because I think the reasoning ability of these models is there, actually. I think the bigger problem is, like, evaluation, improving process and collecting more data, too. The data isn’t there. Sonya Huang: Is model development going to freeze here? Winston Weinberg: I don’t think so, no. I mean, I think we’re seeing a pretty large evidence of if you throw more compute at the model and you let it make more and more reasoning steps, it just gets better and better. Pat Grady: At inference time. Winston Weinberg: Yeah, exactly. Right. And it’s interesting, if we go all the way back to 2022, in the early 2022, the thing that Gabe and I had done was we were doing a bunch—we got access to GPT-3—this is public by the time, this is, like, early 2022. And the thing that we found was we were just doing chain of thought prompts before anyone was thinking about chain of thought prompts. And the way that we actually started was we were doing that over a bunch of kind of like legal questions, and we cold emailed the general counsel of OpenAI and sent him that. And he basically—his name’s Jason Kwon, and he basically responded, “Oh my God, I had no idea these models were this good at legal.” And I think the main reason people weren’t looking at it is because they were just doing one model call over a set with one static prompt and just calling it a day, and they weren’t doing what are the actual process steps. They weren’t telling the model these are the steps that you need to take in order to answer this question, right? And so my point is that’s what we saw in, like, early 2022, and that’s, like, the direction that these models are going, right? And I think that it’s really good for legal and for kind of these industries where you have to have these complex decision-making processes for two reasons. One, the models get better at that; that just increases the ability of them. But two, is that process data doesn’t exist like I said before. Like, it’s not online. Like, how to book a flight is online, right? There are easy ways to train that. It’s harder to do how do you do an LBL? Building trust with lawyersPat Grady: Well, and you guys, you now deservedly so have an unfair advantage around the process data because you have some of the very best law firms in the world as customers of Harvey, which I think was a contrarian strategic decision that you and Gabe made a couple of years ago. And you made it with high conviction. And I remember a time when there were lots of law firms who wanted to come work with Harvey and you basically said no so you could focus on some of these big prestigious firms. Can you say a couple words on what gave you the conviction that that was the right strategy? And then maybe more importantly, once you determined that that was the strategy, how in the world did you get them to trust you? Winston Weinberg: Yeah. Pat Grady: This is a very, very scary new world, and you managed to earn their trust. So how did you do that? Winston Weinberg: Yeah, so I think—okay, so the reason for doing it, there was a GTM reason for it and a product reason. From the product side, the bet was that the models were going to get better, and that you want to build systems that the next generation of model cannot do, right? And so you want to go after the incredibly complex international merger type work, right? You want to do the very complex work and build systems for that because it is the most defensible by far, right? So that was kind of the product side of it. From the GTM side of it, I think that something that’s really important in professional services is prestige and trust, right? The reason prestige is so important is because trust is the most important thing in professional services, right? And so the reason we went after the larger firms is if you earn the trust of a few of those firms, the rest of them will trust you and the rest of the firms downstream will definitely trust you, right? And their clients will trust you, right? So I think something that we thought about doing in the beginning was well, just go straight to enterprise, right? And there are a bunch of problems with that, but one of the main reasons is there’s just no reason for them to trust you, right? That you can actually build these systems, right? How did we do it? We did a bunch of things that do not scale at all. Pat Grady: [laughs] Winston Weinberg: Like, in all honesty. I think that one of the things that we did really well is I don’t think there is any excuse for someone who is building an AI product and trying to sell to not do hyper-personalized demos. Like, there is no excuse. I think it used to be really important to do that. Now it’s paramount and it is so easy to do this, right? And so one of the things that we did in the beginning was whenever I would demo to a partner, I would try to use something that they recently worked on. And then the other thing, too, is lawyers are argumentative, like, very argumentative. And I mean that in the best way. Pat Grady: So just let them fight with the model? [laughs] Winston Weinberg: Yeah, I’m serious. And I mean that in the best way. So I would sometimes say, you know, “Was this a good argument, and how would you improve it?” And if they were really bored on the demo and then you say that, they are reading every single word that comes out of Harvey, like every single word. And, you know, it wasn’t always the perfect response, but I think it engaged them in a way that they’ve just never been engaged with software before at all, right? And we found—I mean, one thing that’s interesting is a lot of the older partners at firms, sometimes we might be their first AI product that they’ve used, right? And so it’s really important to actually show them kind of the basics as well, not just exactly what is special about your product, too. Changing behaviorPat Grady: How have you elicited the behavior change out of your customers and kind of taught them to use AI? Winston Weinberg: Yeah, this has been really hard. I think that it starts with product, by far. So when I was talking earlier about the expand and collapse, the most important piece of the collapse is to make it so you don’t have the blank page problem, right? Like, that is by far the most important thing. So that when you go onto the landing page of Harvey, there are all of these buttons that you can click that will help you get started, right? And that is super important. And now we actually have it to a point where you put in how many years you’ve been practicing, what kind of lawyer you are, things like that, and it will change what that screen looks like when you start, right? So that’s been super helpful because if you make it so that those very specific systems are exactly what they do day to day, it is really easy to get them to start. And if they start using it, they’ll be creative. They won’t be creative right off the start. It’s hard to get people to do that, right? So on the product side, I think that’s the most important piece is just making it very personalized and making it so the time to value is really short, like as fast as possible, and then they’ll go out and explore. On the CS side, we’ve hired a lot of lawyers, and I think that’s been super helpful is you need to hire domain experts who can say—and we’re doing the same, too, in tax and these other areas, you need domain experts who can come in and say, “This is how I would use it.” And this is—you know, I’ve been doing this for six years. I was in the same shoes you were in. I know exactly what work you do. Hallucinations, bug or feature?Sonya Huang: Talk about the trust thing. Hallucinations. I think if you look at where AI has worked really well, it’s where hallucinations are a feature. Like in the creative industries, hallucinations are a feature. In your industry, I’m assuming hallucinations are an absolute bug. And so what do you do to make those argumentative lawyers really trust what the model was saying? Winston Weinberg: So I’d actually push back on the second piece a little bit in a lot of what lawyers do is creative, actually. Like, a lot of it. Pat Grady: Hmm. Winston Weinberg: So, like, I’ll give you an example for this. Litigation would not exist if there were just 10 really simple rules and everyone knew which fact pattern fell into the boundaries of those rules. Like, there would be no litigation, right? And so I was a litigator. I mean, transactional is also incredibly creative, but litigation I think is just a better example because what you’re trying to do is, I have these facts, here are somewhat of the rules of the game, let’s figure it out, right? And that’s actually incredibly creative. Like, very creative. But still, like, accuracy is more important than creativity, to be clear here.  And so on the accuracy piece, so even given that creativity is actually super useful in some instances, and we try to make it so that our product doesn’t, like, get rid of that aspect, on the accuracy side, there are a lot of things you can do to improve it, for one. But for two, the main thing is law firms are hierarchical. So a junior associate, if they get a task, they do the first draft of it, and then a second year reviews that, and then a fifth year might review that and then the partner reviews that and it goes out, right? And so actually, the minimal viable quality of your output can be lower for a law firm than it can be for an in-house team, actually. So selling to the law firms was also helpful in the beginning because so much of the work gets reviewed, right? And so you aren’t selling them something where they kind of click it and forget it, right? They’re actually in the loop at all times. When you’re selling to an in-house team, it is better to sell them the specialized versions of tasks that they can kind of see an insanely high accuracy level on. And the more specialized you’re building, the higher accuracy you can get because it’s easier to fine tune, it’s easier to do evaluation because there are less steps, and the surface area is just smaller. The law firm of the futurePat Grady: Let’s talk about the lawyer of the future and the law firm of the future. Foundation models are going to keep getting better. You guys are going to keep doing valuable stuff on top of those models. What does the job of a lawyer end up being? What does the law firm end up looking like? Let’s go, like, five, ten years out. Winston Weinberg: Yeah. So start with the job of a lawyer, because I just care a lot about this. I think that it goes back to actually what it used to be. So 50 years ago, the role of the lawyer was an advisor. And a lot of what they did is look around corners, give a different angle of advice, et cetera. And we’ve actually seen this evolve to a degree where the CLO, Chief Legal Officer title is, like, pretty new. And really it’s like, most of the CLOs that I have met, they are part of the business. Like, they are business drivers. It is not just the “no” person, right? And I think that that is going to happen. So I think what’s going to happen over time is kind of this, like, lower end work is going to get somewhat commoditized and automated, but then the high level strategic work is actually going to be more valuable, right? And so this is really good for a young lawyer. It’s fantastic because, you know, most people, they go to law school and their goal is not to sit in a data room or do discovery or doc labeling for 10 years and then maybe go to trial once. That’s not what they want to do, right? They want to give advice to clients. That is why you want to be a lawyer. They want to help people, right? They want to help people or they want to win, one or the other, right? But sometimes both. That’s what you want to do. And it very much is, like, close to like professional athletes where they want to be the best at their craft, and you get better by having more hands-on strategic experience than you do just sitting in a data room forever, right? And so I think that experience will be really good. The law firm, I think there will be a lot of transitions in how law firms operate. So I mean, I now am a client. Like, I use a lot of law firms, right? And one thing that I’ve always got annoyed with is the really high billable hours for very low end work. Like, looking at, you know, whether this change of control clause was triggered or not, things like that, right? And as a customer, that can be annoying, right? But actually for the high end strategy of, like, how should I go about buying this business? How should I think about restructuring this part of my org? Things like that, I would pay more than you pay the best lawyers on Earth right now. Like, the delta between a junior associate and the best partner on Earth is like 3 or 4X, right? Which actually doesn’t make sense to me. And so I think what will end up happening is there will be a lot of fixed fees for kind of the part of a transaction, the part of a litigation that is somewhat more commoditized. But the insight, the value of strategic advice, looking around corners, things like that, I would argue you can charge more for that. Expanding access to justicePat Grady: I know part of your mission is related to providing better access to justice. What does that mean to you? And kind of draw the line for how Harvey helps us get there. Winston Weinberg: So a little bit of background info on this. The average price of a lawyer in the United States is $352 an hour. So almost no one can afford a lawyer, right? And I think there’s a bunch of arguments about, you know, how much latent demand is there for lawyers, et cetera. The reality is there is a massive population of folks that do not have access to our justice system, right? Like, either way you slice it, that is the case. And even if you had every single lawyer work 20 hours a week on access to justice, it still wouldn’t close that gap. Still. Literally. Without anything else being fixed, right? And so I do think that we are going to have a large transition to lawyers using AI to actually help and increase access to justice. There are a bunch of things on kind of like the regulatory side that prevent a lot of this, but I think a lot of those will change. So the best example of this is there are kind of two conflicting rules here. It is unauthorized practice of law, right? So businesses cannot give legal advice. Someone who has not passed the bar actually cannot give legal advice, et cetera. And the second one is you cannot make an equity investment into a law firm unless you are a lawyer, right? And so that has basically cut out any sort of traditional financing in the legal sphere, right? Utah, Arizona, and now some other states are also thinking about getting rid of these rules, or creating sandboxes to kind of experiment with that. And it’s been going really well. I do think that there is going to be a lot of change in this in the next couple of years, and I think it will be amazing. Like, there really is—I think that if you can figure out the way to make sure that people are getting the same quality of legal advice with AI and maybe a lawyer in the loop, depending on kind of how that pans out, this will be incredible for folks. And my last piece on this is one of the things that I think people don’t think about a lot is most people don’t know when their rights have been violated. They do not know, right? And so I think we take for granted that we have all been in this kind of bubble, have been super well educated. When something happens, we know whether we have legal recourse or not to some degree. A lot of folks don’t, right? And so they don’t know if a person in authority is doing something that is illegal or not, right? There’s a huge problem in housing, huge problem in kind of like collecting unpaid fees for things that shouldn’t even be a fee, et cetera, right? So I think that’s an area where you can do a lot of not just providing the services, but a lot of education at scale that you couldn’t do before. The AI marketSonya Huang: Can we zoom out to the AI market more broadly? Winston Weinberg: Yeah. Sonya Huang: OpenAI was lucky enough to partner with you in the early days. What do you make of the recent developments that are happening in the AI ecosystem, and what do you think are the developments that are most interesting for you? Winston Weinberg: Yeah. I mean, costs going down are always good. I think one thing to think about here is there are so many use cases that we have gotten to work 70 percent of the time, right? And the o series models have been incredibly helpful for us because not even just the o series models themselves, but it unlocked kind of like our product strategy for the next six months to a year, right? Where there were so many systems that I don’t think we thought that we would be able to build in the next year if it was just the GPT series models, right? And I think that the o series has massively kind of changed that for us. And so our product roadmap has drastically changed because of that. Sonya Huang: What’s an example of something that you couldn’t do before that now you can do? Winston Weinberg: Yeah, so examples of this are, like, things that you need to do multi-step reasoning for and pulling from many sources at once, right? So the thing that the o series models is really good at is orchestrating a plan and also executing on that plan, right? So if you build a bunch of systems that are really good at basically extracting information from EDGAR, then extracting information from case law, then extracting information from all your internal documents, the missing piece for us was what do you do once you extract all that information, right? Like, once you have all of those different pieces, how do you combine them into the correct work product? That’s what the o series models allow us to do. Switching places with SamPat Grady: If Sonya had a magic wand, and she were to wave her magic wand, and you and Sam Altman all of a sudden switch places and you’re now CEO of OpenAI and he’s now CEO of Harvey, what would you do different at OpenAI and what might he do different at Harvey? Winston Weinberg: Yeah. Well, at Harvey, he’d raise more money. Pat Grady: [laughs] He is pretty good at that. Winston Weinberg: He’s really good at that. Maybe I’ll start with what I think OpenAI has done an incredible job of, and I would maybe even double down on it more. I think that they have done an incredible job of capturing the consumer zeitgeist. So I am not from Silicon Valley and I don’t have—I mean, now I have tons of colleagues here and tons of friends, but I didn’t beforehand, right? And almost none of my friends know what any AI tool is other than ChatGPT. That’s it, right? And the thing that I think I would do differently, and to be honest, I think they are going this route anyway, is just, like, put way more effort into productionizing that for consumers as much as possible, right? And that’s not just model performance at all, right? So I think that’s still one of the biggest problems with ChatGPT and just kind of like all of these tools in general is they’re looking so much on the performance from the model side and not at how do you make the experience easier for the user, right? How do you make it so that it extracts more information from the user? How do you help the user figure out what it can do and what it can’t do? How do you help the user figure out how it combines with different pieces of information, et cetera. So I think on this side, it would just be putting more and more effort into understanding the consumer behaviors, and how they use AI right now and making that easier for them. Pat Grady: You guys have been partnered with OpenAI from the very beginning. What has it been like and how has it evolved over time as it’s gone from kind of undiscovered to being, you know, the center of the universe in many ways? Winston Weinberg: Yeah. I mean, so I think the thing that we have kept up that has been really awesome for the engineers at Harvey but also at OpenAI, I think, is we have always been working on things that I think a lot of companies aren’t. And so what I mean by this is OpenAI would give us models all the time and say it performs way better on all of their benchmarks. And we would respond and say, “Sorry, but it doesn’t on ours. Like, it’s actually not better for us.” And I do think that the reason our relationship has been so strong is we keep saying, “Okay, here’s a model that can do XYZ. I would like it to do XYZ and the rest of the alphabet too, right? And that’s what we’re trying to do. And so I think we’ve actually helped them a lot, too, with at least, like, applied use cases, and how they think about post training and how they think about what are some of the things that companies, like, are really pushing to try to do, right? And so I think that has been really, really good for the relationship overall. Pat Grady: Do you have any hot takes on Microsoft OpenAI? Winston Weinberg: Yeah. Pat Grady: [laughs] Winston Weinberg: One of the hardest things about being an application layer company is you have to bet on model providers. And it is not—like, we build our system so that you could kind of just pull out one model and replace it for another, but you could do pieces of that, right? So, like, we don’t just use one model for the entire system. You’ll build a piece that’s really good at this and then a piece that’s really good at this and then you’ll chain them all together, right? But the reality is, you know, who’s in the lead changes so often, and different models are good at different things, right? And you have another problem too, where all of your customers—or most of them—I said that partners probably aren’t using a bunch of AI tools, but the associates for sure are. And the associates are using all these different AI tools, right? And so what I’m trying to say is we, I think have done a good job where we work with all the model providers, and we are constantly testing out models with all of them. And I think that Microsoft, which has, you know, such a massive customer base, they have to figure that out too, right? They have to figure out, “Oh wow, we have customers that think that Claude is better at certain things. What do we do about that?” Right? I mean, they can’t really do anything about that, but they’re—like, you know, Mistral or whatever it is, right? And so I think that that relationship is evolving as it naturally would. The biggest threats?Pat Grady: If the opportunity for Harvey is to revolutionize a trillion-dollar industry and provide better access to justice, what is the threat? What are the biggest threats to Harvey? Winston Weinberg: I think not moving fast enough. I mean, I tell this to my team a lot, and I think it’s becoming very obvious in the past couple of months, too, that we’re really living in a time when all of your timelines are compressed. I would argue this is, in all of human history, the most compressed timeline in terms of what you can change in the world, right? And I think that you have to move so incredibly fast in order to keep up with that. And the speed is compounding in a couple different ways. If you are constantly moving fast, you are required to constantly test everything around you, pay attention to every single part of your industry, how every single customer is using your product, how your sub processors are changing their models, everything, right? And if you do not do that fast enough, you’ll make a bunch of mistakes, I think. And you constantly will make mistakes, right, by moving very quickly, but the scariest mistake is you move too slowly and you miss a massive thing, right? So there is part of a feature that someone releases something and you have steps 12 through 13 complete on that feature. But that 13th step, you just cannot get the models to do it. And somebody releases something that unlocks that, you need to put that in the product immediately, right? Because you need to start testing it, you need to see actually did it solve it, et cetera. And if you aren’t moving quickly and you just say, “Ah, that’s something new, we’ll try it out eventually,” I think it’s a huge problem. Lightning roundPat Grady: Speaking of moving fast, good segue into our lightning round. Winston Weinberg: Okay. Pat Grady: Since starting this company two and a half, maybe three years ago, how many days have you taken off? Winston Weinberg: It’s a loaded question. [laughs] I have not taken a day off. I probably should though. In all honesty, I think, like, I’ve definitely a little bit worn the—you know, in, like, full transparency, I’ve worn the badge of honor of, like, you know, don’t take any time off, obsession, et cetera. And I think on one side, I actually very strongly believe in that. So going back to my point about the timelines being incredibly compressed, you need to be obsessed, right? Like, you massively need to be obsessed. But I do also think the other side of that is, like, you need to transition how you are a leader. Like, that needs to change. And, you know, our company, last year we started the year with around 40 people. We have 260 right now. And you just need to change how you are spending your time, right? And I think that I’ve definitely learned how to spend my time differently. And there’s also stuff I’ve held onto and I actually, like, deeply believe in. So one of the things I deeply believe in is I actually think you should do part—for a little bit—of every single job at your startup. For a little bit. Like, I actually—I do it too much and I did it for too long. But it is all—most of my hiring mistakes have been I didn’t understand what the role did, like, at all. And a bunch of people told me what it was, but it doesn’t help you hire if you don’t understand what the actual role is. And so that one I feel really strongly about. But at the same time, I also took way too long to hire, and I was probably doing too many low level things for too long, right? And so it’s a combination of both of those. Pat Grady: For those 260 people who work at Harvey today, what is the best thing about working at Harvey, and what is the worst thing about working at Harvey? Winston Weinberg: Yeah, I think the best thing is every day something new is happening, right? From the product side, from the GTM side, you’ll see a bunch of changes with the model providers that you want to integrate very quickly. It is stimulating above all else, for sure. And I think we’re seeing a lot of impact, too. And it changes, too. And I think if you ask folks that have been here for a year or a year and a half—and there aren’t tons that have been for a year and a half—the thing that I think is most fun for them is how much our market has changed, right? Where I mean, we used to—it was brutal in the beginning for kind of how you do your sales process, the requirements people had and things like that. And now most of our customers, like, they’re partnering with us, right? And they’re letting us do their onboarding. They’re doing all of these things where it really has seemed like we’re kind of this wave of all of us together. Pat Grady: Doing it together. Winston Weinberg: Yeah, doing it together. And that changed a lot. There was a lot of pushback in the beginning, like, a massive amount of push. And there’s still some, but I think that has changed pretty drastically. The worst thing is the expectations are really high. Pat Grady: [laughs] Winston Weinberg: And, you know, we had a really good year last year, and I think—we had our off site recently, and I went up to the off site and I basically said, “Hey, we did a really—you know, we had a great year last year. This year needs to be, like, significantly better and we need to raise the bar.” And, you know, I think that’s not always what people want to hear. Sometimes people are like, “Wow, we did such a great job last year.” Pat Grady: Let’s take it easy. Winston Weinberg: Yeah, let’s take it easy. Or do the same as we did last year and then it’ll be fine, right? And the reality is again, going back to those timelines being so compressed, you can’t do that. And I think that the main thing I ask people is basically look, like, I don’t know what your goal is. I don’t know if your goal is to make a large industry change, to learn as much as you can, make money. Whatever it is, your options to do that in the next decade is the best it will ever be. Like, it just is, right? You will make more impact than you ever will have the chance to in your life. Sonya Huang: How have you changed as a CEO, and what prior have you updated the most? Winston Weinberg: Yeah. Hopefully I’ve changed somewhat. I think the prior that I’ve updated the most is teach, not do. Like, I’m bad at that. Because I kind of want things done so quickly, I have a massive problem of whenever I start seeing friction, I just go, “Okay, I’m gonna go do it.” Right? And that’s actually really, really bad. I think it makes it so other folks can’t learn. It makes it so that it is kind of like too top down, right? And it’s something I’m working on a lot. And I think I’ve gotten better at it, hopefully. You can ask my direct reports. I’m not sure if I have, but I think I have. And I think that that’s the area that I want to keep getting much better at, is taking kind of—you know, slowing down in some instances and actually, like, setting ourself up to scale instead of everything being like, go fix that, fix that, fix that, fix that. Yeah. Pat Grady: Who’s a better athlete, you or your co-founder, Gabe? Winston Weinberg: Oh, Gabe is a much better athlete than I am. Pat Grady: [laughs] Winston Weinberg: I mean, it’s really unfair. So he played professional soccer, and he is just a much, much better athlete. But he had a knee injury recently. And I will say I do a little bit—like, we live together, and every morning I’ll get up to go to the gym, and I definitely slam the door a little bit loudly just so he knows that I’m going to the gym and he can’t quite yet. We met each other before this, and we became best friends and had no plans of doing a startup. And he was just always a better athlete than me, and so this is my revenge a little bit. It’s not gonna last long. He’s healing right now, and so he’s probably right now working out and getting better. Pat Grady: All right, last question. I’m gonna steal the last question from Guy Raz. I don’t know if you ever listened to his show, How I Built This, but he has the same last question every time, which is: How much of your success has been luck, and how much of your success has been skill? Winston Weinberg: It depends how you define “luck.” We have been in a place where we have the options to apply skill, and we have the leverage to apply skill, and if you apply it correctly enough times to actually have a large impact from that. And so the luck is the timing. The luck is, do you actually have the option to make a difference, to make an impact? And actually, I think—you were kind of asking earlier about what have I learned as a CEO? One of the things that I think I’ve actually tripled down on, quadrupled down on is young talent. Like, by far. And that goes back to giving them the luck or the opportunity to actually try something they’ve never done before. It works out really well, and they don’t get it right every time, but I don’t either, right? But you adjust really fast. And their ability to adjust, I think, is better than a lot of folks that have been doing this for a long time. And so maybe the best way to phrase that is their skill is their ability to adjust to luck and seize luck opportunities more than anything else. Pat Grady: I like it. Winston, thank you so much. Winston Weinberg: Yeah, thank you. <div class="framer-mcvrsd" id="