How Winston Weinberg built his $11 billion AI Company | Term Sheet

I think it's really hard to figure this out without failing. Like you just have to fail like a million times. Like, what is your relationship to failure at this point in your life? Marriage. Hello, hello, welcome to Termsheet. I'm Ali Garfinkle, senior writer here at Fortune. And this is the podcast where we talk about the weird and wonderful world of private capital tech and startups. This week we're talking to Winston Weinberg, CEO and co-founder of Harvey. Harvey is widely viewed as among the leaders in Legal AI and the company is valued at $11 billion, despite being just a few years old. But first, let's talk about the news. We're now several weeks into the war in Iran, and energy markets have very publicly been hit hard. But it's not just energy markets. We've seen the public markets react, and the private markets are also reacting. A number of sources have been telling me they believe the IPO pipeline has been slowed substantially for the time being, in part because of the war. The pipeline was already slowing down because of the SaaS apocalypse, and this is expected to compound those effects around uncertainty. At the same time, a lot of folks I'm talking to in the M&A landscape do believe deals are done. This is at best a proceed with caution environment. There are some folks you'll talk to in the private markets who will say, "We invest on such long-time horizons that it doesn't actually matter." To which I say, "How could it not?" At a certain point, it will create challenges for companies that otherwise wouldn't be there. And if you're looking to make a deal, if you're looking to go public, going public in this kind of environment is extremely challenging. In short, stability breeds liquidity, and this is proving to not be the year that everyone was expecting in the private markets. People really came into 2026, I think hoping for a sense that everything would start to stabilize, things would get a little bit more predictable. And I think it's safe to say at this point, that's not the case. In much less serious news, Travis Kalanick, founder of Uber, is back. He has a new company, it is a robotics company called Adams, ATOMS. And he's going to be rolling his current Ghost Kitchen company called Cloud Kitchens into the specialized robotics company, Adams. He's saying he's going to be focusing on robotics in the food, mining, and transportation industries. I don't know exactly how Ghost Kitchens and these things are connected. I can sort of see some through lines. I can try to spin it out. But Travis, call me. I would love to talk about it on the Term Shoot podcast and a little bit of a blast from the Uber past. And now, onto my interview with Winston Weinberg. Winston's someone who I've been wanting to talk to for a long time, because he in a lot of ways is one of the quintessential founders of the AI era. And having interviewed more than 1,000 entrepreneurs at this point, there were a couple of things Winston talked about that really stood out to me. The first was failure. Some people say they like talking about failure. This guy loves talking about failure. He also took me inside his process around what it actually means to prioritize appropriately. He has a formula and it got really granular, honestly, to the point where I was like, "Gosh, should I start doing this?" And he told me the real story of how Harvey was founded. And it was genuinely surprising. Here's Winston. Winston Weinberg, hello. Winston, I was so excited to talk to you for a couple of reasons. The first is that you looked at the legal space. One of the most entrenched industries out there, one that literally works on precedent, and said, "I can disrupt that. You also still live with your co-founder, and you're now sitting on an $11 billion company. So I want to understand who you are." And I think, you know, we'll get into long-term scaling, where AI is going. But I'd actually like to start by just setting the table a little bit for viewers. What does Harvey do and why? Yeah. So Harvey is a legally-eyed platform. And I think you can kind of think of what we're doing as we're kind of scaling out and developing different systems over time. So in the beginning, we started really as an assistant or a co-pilot, and we're moving closer and closer to an entire platform. And I think probably soon we'll end up becoming closer to an infrastructure. Right? And what I mean by that is, how do you coordinate all of the different agents that are very good at doing a specific legal task into creating an entire full process? Right? So you can think of this as moving away from just like a co-pilot or an assistant to how do you have the solution that is the infrastructure that gets a deal done? Right? And it coordinates all the different parties. It has agents that do part of the work, and then they sign that work off to a human. The human reviews it, passes it to a different agent, and you may have like multi-parties working in the system. And so really what we're starting to see is, how do you do this evolution from an application layer company from, you know, kind of these co-pilots that I think are going to get decently commoditized to really deep workflow integration, and then from there to kind of like full infrastructure for getting work done? I was going to say, how do you actually even know when you are taking the steps towards all of those things? To our listeners, that might sound a little bit vague, but there is a substantial difference, actually. And in terms of the company itself, but also in terms of what it means to survive the AI era, let's be real. Yeah. I think like we've always had the same plan for kind of the stages that we go through. I think that they've happened faster than I thought they would. So I mean, for instance, we kind of started with selling to large law firms, right? And now we serve large law firms and then a lot of the Fortune 500. And that Fortune 500 part of our business is going really, really fast. I thought it would take years for law firms to adopt this technology. And I thought it would take much longer for the profession to start to think about how we are actually going to change, how we deliver these services, how we evaluate, how these services are delivered, and honestly, what is the net new things that a company can do? And the reality is the models have gotten so much better, like the underlying systems. And I think the economy, especially in the past couple of months, is starting to say, wait, I actually want to transform my entire business based off of this. And so you can kind of just feel the market pull, right? Like a section. Yeah. And it's not just what are the capabilities that you can create, but what are the preferences, right? And I think that this is really important in highly-- Because you're sort of dealing with a form of consumer behavior. 100%. Right. And so like if you're a large enterprise, a lot of what you're starting to think about is, okay, what type of work do we have that's fully agentic, right? So maybe you don't even need humans in the loop for, right? Like routine daily work. What kind of work is going to be agentic, but with a human in the loop, right? Internally at a company. And then what kind of work is going to be a combination of agentic plus a human internally at the company in the loop plus outside counsel and outside law firm involved too? And I think that those type of discussions are starting to happen in the past like six months, which changes your trajectory as a company. Because it makes it so that you're greater, your grander ambitions, you can start pursuing that way earlier on. Now I think this is an opportunity to ask you a question I've wanted to ask you for a while, but has we have not had the opportunity because we've always had a task. How did this all start? You went to law school, right? Like take us back to the beginning. Yeah, so I was a lawyer only for eight months. So it wasn't an incredibly long time. So you loved it. Yeah, no I did. I mean, I actually did. What I wanted to do was I had an internship at the US Attorney's Office. And what you do in your there is you kind of like work with the FBI on these really cool federal cases. And that made me realize I wanted to be a lawyer. But what I probably wanted to do or at least back then what I thought about doing was you go and you work at a law firm for a while and then you go to the US Attorney's Office and you get a lot of trial experience. I want to be a trial attorney. That's like what I wanted to do. And you can't do that if these big law firms. It's like you can do that like 10 years in. You work at the US Attorney's Office and then I wanted to start my own law firm. It's a very common thing in Southern California. I was going to say you had a plan. I didn't really had a full plan. Yeah. What motivated you at that time? I think like for me, for me, I was never an incredible student. At least before I had this internship and I had this internship and I think I just didn't I didn't hadn't been exposed basically to a type of profession that got me really excited. And when I had the internship, I just kind of looked at all of the people that were working at the US Attorney's Office. It was in New Orleans actually. It was really cool. And all of them just love their job. They worked like crazy. The whole notion that like government employees don't work by the way. Not totally sure if that's true. And if you've ever met a United States attorney. They work. They work really hard. They work just as hard as the big law attorneys. And it really was just I wanted to do it and I felt like passionate about something. And that was the motivator then. And then what happened is I knew my co-founder Gabe who had the research background and he was at Google and then met after that. How did you meet Gabe? We actually met in San Diego through like friends. So it was not like what were you guys at a bar? Like it was happening. We got invited by a mutual friend to a brunch basically. And we just met each other at that and we became really close. I remember that some of the first things that Gabe was talking to me about was basically using like prediction models to predict like different chess moves. And how could you actually translate that into a learning platform? And then how could you take that into a general learning platform? And I just remember that was like one of our literal first conversations. And we really got along and then about a year and a half, two years later, what happened is he showed me GB3 at the time. And so in early 2022 or late 2021, there was an API available to open AI, GB3. Right. And this is available to the public. Like everyone had access to this, right? >> It's a different time. >> Yeah, everyone had access to this. And he showed me it and we had some friends over that we're actually thinking about, they were thinking about joining OpenAI and they were working on large language models at Google. And we looked at these models and then we looked at legal work. And I started using the models for like legal work. And it was incredible. But I didn't know, you know, when you're practicing securities or antitrust law, which is what I was doing, you're actually not an expert in all these other areas. And so what we did is we went on to our slash legal advice and we grabbed 100 landlord tenant questions. And we came up with the chain of thought prompt and we ran those over the landlord tenant questions. And then we just gave them to landlord tenant attorneys and we said, would you give this answer to the person that asked this question? Like assume there are clients and nothing else. And 86 out of the 100 questions, three out of three folks said yes. Yeah, it was that good. And we were you surprised? >> I was insanely surprised. >> Yeah. >> And because they worked decently well for the type of work that I was doing, but not well enough to build a product around it, right? Like they weren't good enough to handle big law. And so in the beginning, we were actually thinking about doing consumer law. And but we bundled all those results out from the landlord tenant attorneys and we actually called emailed Sam Altman and then general counsel at OpenAI at the time. His name was Jason Quan. He's now I think the chief strategy officer. And we emailed the results and we said, did you know that the models were this good at legal? And then we, you know, a couple weeks later pitched them on the idea of the company. And what really changed for us is once we got access to GPD4, we said, wait, you can do much more complex legal work. You can do corporate law. You can do, you know, the big law type of legal work as well as consumer. >> This was a very different time to be emailing Sam Altman, but what happened? What did he say for at first when he responded? Like how did the conversation evolve? >> Yeah. I mean, so they, we set up a first meeting with the general counsel and then we pitched them the idea there and then we set up a meeting. It was actually the morning of July 4th with the rest of the open AI. >> Very patriotic. >> Yeah, exactly. With the rest of the open AI, I see sweet and we pitched them on the vision there. So and the vision hasn't changed that much since the original pitch. I think the timelines have compressed. Things are moving much faster than I think we originally thought they would. >> So what was the moment where you said, okay, we're actually doing this? >> When we had the results. >> When you had the results, that was the moment. >> Yeah, I think for me, the biggest thing was, can you get these models to generalize, right? So at first it was basically, okay, with a lot of work, I can get the models to do certain things. But the bigger question was in order to create a real like product here, can you get the models to actually generalize and be good at all of these other types of legal tasks and until we went outside of a domain that I understood as well and we got good results there, I thought, oh, maybe this is just something you can tinker with, but you can't build a product around. >> What were your fears at that time? Did you have any reservations? >> I know. I think like the, I mean, part of this is like, I don't have a tech background and before Gabe, actually, I had almost no friends in the tech industry and so a lot of this is just ignorance, right? Like, you don't, you have no idea how hard any of these things are, you don't know how hard it is to build a tech product, you have no idea how hard it is to build a tech company and so you jump into these things mostly with ignorance. And I think like, it's, it, what ends up happening from there is you find out things are harder and you learn from it, which I think has been, it's been somewhat of an advantage, I think, to not have that background because you try to do things differently, right? And think from first principles, etc. >> Well, there's something very interesting about the idea that this is your first company. Legal AI now is hot, but at the time this all started, it absolutely wasn't. >> It was not. >> Yeah, it's hard to overstate and I think memories and tech are very short, but if you said legal AI in what 2023, people were like, I don't know. >> Yeah, there were two massive problems. One was people thought that because the models hallucinate, the legal is a bad area because everything needs to be insanely accurate, right? And I think- >> Which was a reasonable thought. >> By the way, a reasonable thought. I think that was a, b, I don't think people really were bullish on the models getting better really fast. They weren't. And if you look at, and to both of these two things, for the first one, actually the legal process and a lot of professional services are perfectly situated for these models making mistakes. The reason why is they already have a pyramid system. And so what ends up happening is you have the client that asks a question and then the partner sits here and they get that question and they break that question down into 10 subquestions and then they give that down the pyramid, right? And the bottom wrong of the pyramid does some of the work, right? They do that work. It gets reviewed by the next level of seniority, then the next level of seniority, then the answer goes to the client. And actually that's a perfectly ingrained review flow, right? Like you already have that. And I think that a lot of people didn't realize that. >> The backstops are already built in just how law specifically works. >> Exactly. How law works and a lot of professional services are this way, right? And so I think that it was perfectly situated for that. And then the second thing is I think that people have really for a long time until honestly probably the past two months have constantly gone back and forth between us, AI, a bubble, like does this technology actually work? And my co-founder Gabe and I have been just incredibly bullish on the models getting better from day one. And we've constantly designed the company with the idea in mind that eventually long-term your competing against the models because they're really good. >> Well, one of the things that you've said over and over here is that this is all happened a lot faster than you thought. Let's do like an expectation versus reality. What was your expectation for how fast it could go and what has actually happened? >> Yeah, so some things have happened faster and some things have happened way slower. So actually let me start with the slower one. The one that I do not understand is we still have folks that are kind of like I'm not totally sure what the value of AI is. I do not understand that. Like when Chanty B.T. got released, I genuinely thought that it would be integrated into every single business within a month. >> You actually thought that. >> I felt like every single use case. 100%. >> You're not accounting for human nature. >> My brain was basically just while these systems are so good, they can do a certain percentage of every single task. People are going to figure out how to use them. They should go everywhere. So that was pretty wrong about. The thing I think, and that, so that was definitely we were wrong in terms of how long it would take for that. What I was surprised by was how fast adoption at law firms happened once a couple law firms started using it. I think that that has moved way faster than I thought. It was really hard in the beginning, incredibly difficult to get any law firms to even try our product or jump on a call with us or anything like that. >> For a while, no one was taking your call. >> No one would take our call at all. >> One was the moment that you saw that shift. The biggest shift was we had a large announcement with a big law firm that's now called ANO Sherman. They actually announced that they had trialled us and were deploying it across the law firm. I think that we needed that one moment of a firm that has an incredible brand saying, hey, this is actually real. This product works. Once that happened, I think most things snowballed after that. Then I think in terms of the speed of adoption, it took longer than I thought for it to get to enterprises and to get to all those things. Now it's going faster. The industry is starting to adapt to this faster than I thought. We have law firms that are literally building software on top of us and trying to resell that software to their in-house customers. We have law firms that are pitching completely new ways of doing work for customers, doing completely different types of tasks. I think that that has happened faster in the past six months than I thought it would. It's been strange where the initial part was slower than I thought and then the past six months, the adoption rates have been much higher than I thought. >> Tell me a little bit about how you keep up. One of the things I found very interesting about this moment when we think back on it is everything is just moving so much faster than we ever expected. Harvey was founded a few years ago. >> Three and a half, yeah. >> Three and a half, I was going to say we're at less than five years easy. I was being generous. >> Perfect. >> Now worth $11 billion. What does it actually mean to build a company in an environment where not only can you grow that fast but you have to pivot that fast? >> Yeah. The thing that I care the most about with our culture is the size ofness. I think you have to basically build a company that has a culture of making decisions very quickly and being okay to make mistakes and learning that. The reality is the folks that I have found that haven't scaled and when I myself think that I'm not scaling, it's because I haven't learned enough in the past couple of months. You basically have to, one thing I say a lot is you have to re-earn your position every six months. You need to re-earn your role at Harvey every six months. >> That includes you. >> Yeah, 100%. It includes me 100%. >> It includes you the most. >> Yeah, potentially the most. I think that if you don't reinvent yourself as a company and as a leader and whatever your role is at a company right now at an AI company fast enough, you will lose. That involves making a lot of mistakes and that's okay. It is okay to make mistakes as long as you pivot after that mistake and you don't repeat the same thing. same mistake again and again. And so when I look for folks that I'm saying that I say, oh, wow, this person is going to scale. This person is going to go from never managing a team to managing a team of 20 to a team of 100, et cetera. The main thing I look at is, can they make decisions own that decision and then pivot when they make a mistake? Like instead of penalizing the mistake, penalize not making the decision or not learning from that mistake going forward. So you are perfectly okay with a series of mistakes. What you are not okay with is lacking size of it. Exactly. And on the mistakes side, if you continue to make the same one, right? Definition of insanity. Yeah, exactly, right? Yeah, yeah, exactly. And I think like even more importantly, what I'm starting to see is there folks that are able to say, what do I need to get done in six months? Like, can you have a very good plan for what are the top three things that you need to do that are not a problem right now at all? But in six months, they're going to be an astronomical problem. Can you predict that and start taking movements toward that right now? So I'm going to pull this thread a little bit. How do you figure that out? Especially for you personally, where everything changes in AI every six months. And even six months seems like it might be a little long. So there's maybe two ways to do this. One is everything is changing, but a lot of company building I think is quite the same. In terms of, can you hire the right people? Can you put those people in the right roles? And can you create a culture that actually works for your company and for your customers, right? And that part of the job is exactly the same as I think it's always been. At least when I talk to mentors and things like that, it seems like it's exactly the same, right? You have to be able to do that at a faster rate, though. And so you have to be able to say, hmm, is this person going to scale into this or not? Or are we going to have to completely have a different type of org here because the communication between EPD and GTM is going to work for this product line or whatever it is, right? Those things are quite similar. Like they have nothing to do with AI, right? On the product side, that's where you have to just assume that the models are going to get even better than you think they're going to get. And how do you do a product roadmap that makes it so you don't get destroyed by the tsunami of the models, right? That's a harder skill set to learn. But I do think the first limiting gate factor for it is, do you actually believe that? Like, do you believe that the models are going to get this good that fast? If you don't, there's no way for you to ever plan for that. So that's like number one. Why is that belief so important? Otherwise, you get trapped by ego. So you get trapped by, oh, there's something that we can build that they can't. And in six months, that's going to, you know, there's no way they're going to be able to do that. It's much healthier to actually say, huh, these models are going to double every six months and get, you know, that much better at reasoning capability, tool use, etc. What are the things in our domain that are really hard to replicate? That is a much better mindset. And I think that's hard because a lot of times when you try to think about like what you're building, you think, oh, we'd be able to do something generalized better than somebody else. But it has to be vertical specific. When you think about situations at Harvey where someone has scaled successfully for a while and stopped, what does that look like? It's the six month problem. It's a six month problem. It's always the six month problem. It is defined the six month problem. Yeah, the six month problem is you look at your org right now and your customer base and your product and you go, there aren't problems right now. Everything's going great. And you can't do, this is about to break. This is going to break in one month, three months, six months, right? People are usually good at this is going to break in one month. And then about 80% of people fall off when you ask them to do three months. And then 99% of people fall off when you ask to do six. And so that is like the main skill set. And I think the problem with it is it's really hard to learn how to do that unless you've messed it up before. And so like the the number one thing that I try to focus on is every week I want to do something that is hard and like makes me really anxious. Like every week there should be one night where I can't sleep because the next morning I'm nervous to do something, right? And a lot of the times the thing that people are the most nervous to do is the thing that's the most important and the hardest. And I think people like to rotate on busy work. They like to say, oh, I've done 12 meetings today or 15 meetings today. And this week I worked 100 hours, but they don't do well, what are the top three things that I need to do so that we win in six months? And how many of these meetings have anything to do with those top three things? People are it's really hard, right? And I think that it's a it's a different type of discipline where it's not just about performative work. It's can you do the right things? Can you spend the time on the right things? And that's hard for people. Well, it's a form of forecasting, but it's a form of psychological discipline. It is. And not getting lost in well, I did 12 meetings, I can't think about this anymore. I guess the thing I would wonder for you is as you think about keeping all those balls in there, all of the juggling, I'm sure you do. On the days where you can't sleep at night, how do you go back to center and say these are the only three things that actually matter? It's I find it really hard. Yeah, it's really hard. So I do have a tracking system. So I have a document that's just called the list. And on it, I have what I need to do every day. I have like all of our customer metrics are what platform is this on? It's just Google docs. It's just a Google doc. And I updated every single day. And I've been doing that for a little since I started the company. So I have like, and then it breaks at around 200 pages. Google, if you're listening, if you could please fix that, that'd be great. So I basically have to like, I have to like legacy out and then come up with the new list from date to date or something like that's really annoying. Once you look in Google docs, you're listening, please help me with this. Winston has a request. But I so I track everything on there. But the most important thing is I have a top three goals. And I literally bold them and nothing else on the document is bolded. And what I tried to do is at the end of every single day, I go, what it was my list of like 15, 20 things that I did. And how many of them actually like back up into that list? And I actually like try to penalize myself if I've done too many things that I have nothing to do with that top three lists, right? And then you rotate that list and you have to be able to rotate it, right? And I think it's it honestly is a discipline problem. But when I hear people talk about discipline, I feel like they're mostly talking about like how many hours I worked or how many meetings I took. But the harder discipline is actually these are the three things that matter and how do I stay focused. And that's what I've definitely found is the hardest part with people, whether they scale or don't scale is those three things like stay the three things for too long. They either stay the three things for too long or they turn into four things and then five and then 10, right? I mean, I remember we as reviewing, we were doing like quarterly planning and someone put p zero and then they wrote above it p zero zero. And I'm like, okay, so nothing is the priority, right? Like there is no priority. Everything is the priority. Yeah, and then it's going to be like p zero zero zero, right? And I think that like that's also really hard to do as a founder. Like it's really hard. And you're getting pulled a thousand different directions. The company gets bigger. There are more requests of all kinds. You have to start saying no to things. And again, it's I think it's really hard to figure this out without failing. Like what is your relationship to failure at this point in your life? Marriage. No, I think like you're married to failure. I think like it's a very good way to learn. And I think you learn a lot by the way from winning too. And we've had a lot of success. And I'm very grateful for that. And I've learned a lot from that too. But I think you learn from both, right? And I think that one of the problems with the with either of these is like what was the thing that contributed to you failing? What was the thing that contributed to learning? Right? And so it's not just you have to, you know, have a bunch of wins and then have a lot of failures, but you have to get good at taking some time to actually like analyze what did you do? Right? What did you do wrong? And most of that is like destroying your ego 24/7. Like that's mostly what it's just constant ego death effectively. Yeah. And I think like one way to do this is just have very high ambitions and very high standards. Right? And if you have very high standards, then a lot of the successes you don't get trapped in them. Right? You don't kind of get obsessed with what is the valuation of your company or how much revenue do you have or anything like that? You just stop caring about it because it's so far from like what you're trying to build long term. And if you have a really long term mindset for what you're trying to do and you're really ambitious about it, all of the wins don't really look like wins. They're just like, okay, next thing, right? And I think it's actually better to and by the way, I think that's the same with failures where you want to just move on really fast, right? And I've definitely have been given feedback for people that I work with that sometimes it's like an adjustment working with me because I will call out like 15 failures a day. And I think that a lot of people aren't used to getting that feedback. Like they aren't used to it. And the reason they aren't used to it is because they're, you know, they want to be perfect. And I do not care about perfection. I care about rate of improvement. That's it. Like I only care about rate of improvement, right? Only rate of improvement. That's it. Yeah, that's all that matters, right? Because otherwise what you're going to end up doing is you're going to hire a bunch of people that are really good for six months. And then if they aren't improving, then it's irrelevant because your business has changed massively, right? They'll stall out. Yeah, they'll stall out. So rate of improvement is is only in the matters. And I would also argue right now, like with how fast the technology is changing, all that matters is rate of improvement in team. That's it. - I was gonna say, sort of on the subject of success, Harvey's become a darling among venture capitalists, right? $11 billion dollar valuation. It's not nothing, right? Was there ever a time where fundraising was on that list of three things, or has it always found you? - Yeah, I think that if fundraising is on your list, you're not gonna do a good job fundraising. - Why? - There's basically, I think there's kind of like two ways to raise. You can either, I know people say you can raise off of vision basically, or numbers is what people have said. I don't know if that's necessarily true. I think you can raise off of like performance, right? Numbers, whatever you wanna call it, or resume. Those are like basically the two ways. I don't have a very good resume. And so there's only kind of-- - So you had one option like that. - I had one option, right? And so I think that I've just kept that, right? Where I have always thought of fundraising as something that will happen if you tell people you are going to do something and then you do it. And so for every single fundraising round, I've-- - It is a downstream effect. - Yeah, and I've always actually known who I want to do the next round. And I build a relationship with them. And I tell them this is what I think, these are the products that we're gonna ship. This is the DAU of our MAU that I think we're gonna hit, et cetera. And I start that relationship six months before I raise. And I don't go do a process. I don't do all of those things. And then what I do is every month, I just update them on our progress. And I don't think there's any pitch or anything that you could ever do with someone that builds more trust than saying you're gonna do something and then do it. And I actually think this is the same with customers. The number one thing you can do with customers is say that you're going to do something and then deliver on it. That's it. And at the end of the day, almost everything is trust-based, right? So I think that that is a much better way to build a customer base that cares about you, investors that care about you, rather than kind of pitch a bunch of vision and things like that that you might change a lot. And that might not actually be what you're trying to do. - Well, the other thing you're saying here is what you're setting up as a system of accountability and a system where you make mistakes and say, "Actually, we didn't hit that goal." - Oh, 100%. - But I'm telling you now. - I'm gonna tell you a month ahead of time. - Yeah. - Right? And I think that's another thing too is this is something that I have found that people that have been very successful. And we've hired a lot of people with incredible backgrounds who have been very successful. And I think something that they struggle with is this where they have never failed before. And so they hide it and they wait until the end, right? And if there is one thing that I've struggled with with hiring folks that come from incredible backgrounds is this is something is dying or something is rotting at the company and they hide it and they won't tell me until three months. And then it's like very hard to solve. Versus, oh no, I think this thing is becoming a problem. Tell me and my bad, I should have hired XYZ or built this part of the features at first. - Or I think this whole thing just doesn't work. - Yeah, exactly. And that's totally fine too. And I think that being transparent about it early and then moving it along really helps, right? It's the same thing with customers. Like the best thing you can possibly do is tell them that you're going to do something and then do it. And the second best thing you can do is say we aren't going to do that. That's actually like very important. Like and I think that that's actually a huge mistake that a lot of vendors are making right now is because you can build so many things and there's so much noise, you can just promise everything, right? - A lot of vendors in legal AI? - In general. - Or in general. - I think in general. And it's because it's just there's so much chaos out there and so much noise that a lot of folks are going around kind of promising, yeah, we'll build that of course, we'll build that of course we'll build that. And the problem with that is if you don't deliver on that, if you're trying to build a company that lasts decades, people will remember and they're smart. And I have always tried to be very honest about a, we aren't actually going down that path or that's six months from now. And if that's not helpful to you, I'm sorry, but let's talk in six months. And I think that in the long run will help. - How has this applied to agents? Because I feel like we've been talking about AI agents for a while and I am on the record saying, I am still skeptical about how this applies to the enterprise in the near term. But I know that's something you're thinking about. How do you know what it's working? - Yeah, so this is again iterative. Like the best way to do this is what we're doing is we're picking use case by use case. And then we're thinking-- - Like what's an example of a use case where you're like this worked? - Yeah, so one of the ones we're focused on right now is, so we're focused on M and A and fun formation or two of the ones that we're really focused on. And in fun formation, you break that down into like 50 different pieces, right? And one of the pieces that we've gotten pretty comfortable with automating parts of is the comment letter piece. And basically what that is is, if you're raising a fund and you have ADLPs, right? They all care about like different terms and you have to negotiate with all the LPs simultaneously. - All AD of them. - Really easy, yeah, super easy. But it takes a really long time and because you have to basically, like it's hard in negotiating with one party, imagine negotiating with AD and like making sure that they're all-- - It sounds like my personal nightmare. - Yeah, it's really hard. But what we've done is we've said, okay, what you can do is take like the historical comment letters for the LPs that you have again in this round, compare it to what they're asking for now and then create like a massive chart that compares every single one of the ADLPs and like what they want now, what how that's different than what the other ones want and how that's different than what they've done in the past, right? And then what you need to do is create a bunch of synthetic data to get evaluation. And I think one of the huge unlocks that people are not talking about enough is one of the reasons why coding models are so good is the data is out there, right? The data is out there and it's easier to evaluate whether it's correct or not. - Well, code is binary. - Correct. - It works or it doesn't. - Right, and the models though are getting to the point where they're so good at generating synthetic data that you can use that synthetic data to create evil sets. And that's like one of the huge unlocks that we have recently is a lot of this data isn't available for like how do you do common letters or XYZ? And it's not public, right? And so it's really hard to get access to it. And one of the things that we're starting to figure out is what if you just create synthetic data sets for all of this, right? And use the models to create that synthetic data set and then get a bunch of lawyers and have the lawyers use that synthetic data sets to actually try to create the fund formation or the M&A. And that's starting to look like it'll work. And there you can get just incredible performance from agents when in the past, really all you were doing was just like a random model call and that's kind of it. - Why do you think that that particular task is especially good for agents? - Well, I think a lot of legal tasks are especially good for agents because most of them are text-based, right? And if you think about like one of the biggest problems with agents is there's all these different bottlenecks, right? And in legal, at least what you have is a lot of it is text-based, you either are analyzing text or you're writing text or you're comparing text. And the data might be all over the place and that's a separate problem. And that's a lot of what the problem, like what you have to do is like put all of the data in the right spot, but it at least exists, right? Whereas there are a lot of other areas where if you take like a task from start to finish, part of it is like, you know, some conversation you have to have with someone or something like that. And so it's really hard to have the agents get access to that data ever, really. - So one thing I sometimes say is that I think in industry can sometimes develop in the image of its customers. And I've heard a lot of stories about shenanigans and legally, the competitors can get very vocal about one another sometimes perhaps. What is your relationship to your competitors? - Yeah, good question. So I actually really agree with how you started, which is, but I would even take it a step further. I would say that in order to be a successful vertical AI company, it is really important that you take on the brand and the honestly, like morals, principles of the industry, right? And so something that I have massively tried to push with our team is even if you get into competitive agreements and there's a lot of noise and things like that, you just take the high road every single time. And I think that maybe that causes you some damage in the short term, but in the long run, I think that's what's going to win. And legal is an incredibly trust-based industry. It is a very relationships-based industry and it is a very moral industry, right? Or at least, I mean, that is what we are trying to do as a lawyer, right? - Well, if you're a lawyer, you're interested in justice. - Right, you care about this. And I think that in the long run, the companies that have the same type of moral or principles as the customers that they serve are gonna be the most successful. So I think it can be noisy kind of there, but to be 100% honest, I mostly think about our competitors as the big labs almost entirely. And they are very above board, to be clear, both of them are. And-- - All two of them. - Yeah, they are. They're not going around making market noise and things like that. And really, with them, it is going to be best product wins. That's it. Like it is, if we can build a better product than anthropic or open AI can for lawyers, we will win against them. That's it. That is the end of the story. - Well, and the origin story of Harvey involves open AI, right? What is your relationship to this partner in compete model? - Yeah. I mean, I think like the best way to say it is, as the models get better, if we're doing our job, our product gets better, right? If the models get better and we're scared, we're not doing a good job, right? And this again goes back to my piece of, can you look six months into the future, right? And not just for your team, but also for like what your product is supposed to look like, and what are the underlying capabilities of the models so that you don't get eaten by a tsunami, right? And so right now, we have a very good relationship with them, and to be honest, like the model's getting better, is so far good for us. Like so far we've been right. And that's like, yeah. Always be right, but so far we've been right. At what point does it stop getting good for you? I think that if you live in a world where the models have access to every single data source on Earth, can reach every single security system, and have perfect like recall, accuracy, and memory, we're probably in trouble, but I would argue that every single business on Earth is in a lot of trouble. I also think like the reality is, there are just so many things in this particular domain and every regulated domain, that you have to be able to build around, right? Like security, ethical walls, permissioning, a huge problem like go back to the fund formation with ADLPs. Sometimes like you'll set up that fund formation in like Luxembourg, and the data can't leave like Luxembourg for tax reasons. And then you have to coordinate the handoff from the agent to one of the ADLPs, and then like there's a huge coordination problem within the product as well. And that's why I like to think of our company more as building infrastructure. How do you get to the point where you coordinate the entire M&A with all of the law firms on one side, plus the private equity firms on the other? Do an entire deal with Harvey. Do the entire deal, right? And by the way, humans are going to be in the loop during that deal, and a lot of what we have to build is, how do you review all of these things, right? Like we talked about how these agents can get to the point where they can create synthetic data sets, right? OK. Well, if they can create synthetic data sets, and everyone is also using them to negotiate contracts, like what is in data room going to look like for an M&A in 10 years? I'm not totally sure humans are going to be able to process that information at all. Like I think you will probably need to have agents working with humans in order to process the data. It's just going to be too complex, right? So I think that there is a very large place for a player like us in this vertical, just because of the complexities of the domain. One thing I would be curious about really pinning down. You're suggesting at it, but I think it's worth saying it directly. In a world where open AI's chat GPT and Thropics Cloud can get really good, why does there need to be a Harvey? Yeah. I think the orchestration layer is really, really important. And what I mean by this is what part of-- say you're doing that large M&A, right? How do you review all of the outputs of the agents? How do you coordinate all of the review processes? How do you make sure you're connecting to all the correct data sources instead of using online data, things like that? And how do you make sure that all of that is done securely? And so I think a lot of what you're going to see is in these really, really complex fields and regulated fields where you're going to need a human in the loop, you're going to need humans revealing, you're going to have a lot of multiplayer, where you're going to have 17 different parties working on the same project. There's a lot to build. There's an incredible amount to build. And even if the models get access to all of the reasoning of data in these different domains, they don't know the data for that particular deal, right? And they don't know the processes and preferences of, OK, we're going to hand it off to this law firm to review this. And then we're going to have this in-house team produce that output, et cetera. It's not binary, exactly. No. And I think these companies are going to start looking closer to the underlying infrastructure for how these verticals are ran, right? And that maybe is closer to an operating system for a law firm, but for an enterprise that is, if you are going to spin up an M&A, you use Harvey to spin that up and to coordinate that process. And I think in a way, you also are kind of absorbing the liability here without actually being a law firm because you need to make sure that's secure. You need to make sure the correct parties have reviewed everything. You need to make sure the data is in the right spot. And this gets really hard for these complex pieces of work. That's actually a great segue to one of our reader questions. I sent out a call to Termshoot Reader saying, hey, guys, I'm interviewing Winston. Do you have questions? I got to tell you they came in hot. OK. One of them from Demier said, what guarantees do clients have that their cases, data, and sensitive information do not leak? Yeah. So I think this is really important. One thing that's different for us is we basically-- so A, we're multimodal. So you can use basically any model you want. But in general, we have our own instance. We have dedicated capacity. So what's happening is instead of using the normal systems that are used by everyone, we have a dedicated instance, and then we have BYOK and all of the security concerns for each one of our particular customers. We also don't train on any of the data. So the only time we train on data is when it is like a general purpose thing, like improving citations within Harvey, which has nothing to do with any client data. Or we have a client that asks us to do that. And we do have some clients that are asking us, hey, can you create custom solutions? Now, I wanted to turn to you as a person. What do you read these days? How do you spend your time when you're not harving? I don't have tons of time when I'm-- I was going to say, I suspect you don't. But there's got to be something. I mean, I think the one of the things that I still do is I still pay an incredible attention to what cases are being decided and things like that. So I'm very much like I still listen to a lot of the oral arguments with the Supreme Court and things like that, too. So I pay a lot of attention to the industry. I still pay a lot of attention to antitrust law. That was like the area that I was really interested in. And then I still-- less than I'd like to, but I've been a huge fantasy fan forever. And so when I do have time to read-- Tell the audience what are some of your favorites because they will. I don't know. I mean, definitely like all the OGs things, like Lord of the Rings, Malazan series is incredible. King Killer Chronicles are really good. But I haven't read a lot of stuff recently. I go back and I kind of like reread things that I've read in the past. I'd like to have more reading time. And if there is any reading time that I'd like to spend more time on its fantasy, also if there's anyone that is building like an AI world of Warcraft, please let me know. I'd love to invest in that, too. I literally just watch Twitch. I just want to watch requests. I can't play-- I don't know time to play video games. So I just like watch other people play video games sometimes before I go to bed. And like that's like the closest I have. So I would love to invest if you're building an AI world of Warcraft. And the last thing that sort of-- I'd be really curious about from your perspective is-- the legal industry is clearly changing. And I think that does actually matter to the average person. So I was curious about your perspective about how you think AI is going to change your daily lives. Yeah. I think it's a really good question. I think one of the biggest things I care about is-- and maybe this is just because of my age-- but it's like education, A. And when I say education, I don't just mean school. I also mean like how do you get educated in the professional world? And the thing that I am actually weirdly confident in is I think this will be a massive forcing function in how we train young professionals. It will just force us to do this. And I'll give you an example of this. So at a law firm right now, they spend a decent amount of time training folks, but a lot of the best attorneys that I have spoken to, they've kind of just learned how to do it themselves, or they've found a particular mentor, and they learn through them, right? It's an apprenticeship, isn't it? Yeah, exactly. But I think it's going to become more of one. And one thing I like to talk about is firm should start thinking about time to partner and reducing that. In other words, what makes someone a really good lawyer? And how do you train them how to do that earlier on? And the reality is like you do learn a lot from doing document review, discovery, all these different things. But by the 20th time you've done it, I'm not sure how much you're learning. And I think it's going to force the legal industry and force a lot of other industries, by the way, to actually think about what makes people good at their jobs. Like actually what is it? Like what are the human elements that make people good at their jobs? And how do we promote based off of that? And how do we teach people that earlier on in their career? It's a weird thing where all of the visuals around AI are robots, right? They're like, okay, we're going to become like mesh with the machine and whatever. We're going to become a robot. Terminator-esque. Yeah. I weirdly have like an opposite opinion on this, which is, I think the way that most of the corporate world right works right now is a robot. Like most of the corporate jobs today, you are a robot. Like you are in a Ford factory line and you're a robot. And I think that this is going to be so disruptive to that type of work and that type of organization in a large enterprise or a law firm or anything like that that is going to be a reckoning. And it is going to force people to actually think about like what are the most, what are the top three things? I was going to say this all goes back to your top three things. What are the top, it's not- That eventually work will be about that. Yeah, it's going to be about that. It's going to be, you know, what are the things that actually move the needle, what make our business who we are? And I think if you think about every company that, like you start small and everyone talks about that, everyone talks about the mission. And the bigger you get, the more you're kind of caring about like internal politics, who like looks the best, managing up all of those things. These models are just going to get rid of all of that. Like you're going to be able to use the models for so many different things that you can't really like hide the busy work anymore. And I think that's going to be actually really good, especially for the younger generation, because a lot of the corporate world I think is going to start rewarding creativity, focus, discipline, and not just, well, I work this many hours and I produce this many documents. And it's like, well, AI can produce that many documents who cares. And actually what matters is, I decided to try something. And it part of it worked, part of it didn't. I learned from that now I'm going to try something new. And do you have people and do you have a system at your company that iterates and supports that? And if you don't, if what you mostly reward is whoever writes the most documents, well, I can tell you one thing. Claude can write way more documents than you can and way better than you can. And so that doesn't really matter. And I don't know. To me, that was something that I was frustrated when I joined the workforce. And I think that that's actually a really positive thing that's going to happen. I think it's going to make us less robotic. Because it's going to make us focus on what actually we can do as humans uniquely. Winston, thank you so much. Thank you. So week after, I am still thinking a lot about Winston's process around finding the three things that matter most to win in the next six months. And I think part of the reason I'm thinking about it is because it is so symptomatic of this era. What Winston's effectively suggesting is a method of finding focus in an environment that's inevitably going to change on you and a competitive environment that could crush you. But I think his central idea is that if you find focus, that's how you really have a fighting chance. If you should decide to take the Winston Weinberg method to heart, let me know if you start your own Google Dac or come with your own top three priorities to win the next six months. I would love to hear about it. I think I'm going to try it. In the meantime, thanks for watching. And I'll see you soon.