Why Harvey's Hardest Problem Isn't AI—It's Multi-Entity Collaboration

Hi, I'm Connie Loizas, and this is Alex Gove, and this is Strictly V.C. Download. Hi, and welcome back. This week on Strictly V.C. Download, Alex and I sit down with Winston Weinberg, co-founder and CEO of Harvey, the AI legal assistant that's quietly become one of the most talked about companies in Silicon Valley. Winston shares the origin story of how a first-year associate at O'Melvin and Myers used Chatchy PT-3 for dungeons and dragons before realizing it could transform legal work, leading to a cold email to Sam Altman on July 4th, three years ago, that changed everything. We discuss Harvey's rapid growth to more than 100 million in annual recurring revenue. While 33% of revenue now comes from corporates rather than just law firms, and the technical challenges of building a truly multi-player platform that navigates complex ethical walls and permissioning across 63 countries, Winston also addresses the Chatchy PT rapper criticism. Explains why he believes professional services will be less disrupted than people think, and why he would really prefer that law firms not band together to anoint a winner in this particular vertical. This was really a fun, interesting conversation, so let's get into it and we will see you back here next week. I'm so thankful to have time with you. It's kind of crazy to me. I've been hearing about Harvey for fully two years. I feel like it comes up in nearly every conversation I have with a top VC, and so I'm really glad to be connecting. I think our audience knows of Harvey, but I just think it's time for a little deeper dive. So thank you for making time for this today. Yeah, of course. Thank you for making that time. You started as a lowly, if you'll forgive the word first year associate, totally fine. A first year associate at old Melvanie, working what I assume are 80 hour weeks. What was the moment you realized that either a law was not for you and/or AI could transform legal work rather than just speed up drafting? What ended up happening was so my co-founder was working at Metta at the time. He was also happy to be in my roommate. He was showing me GPT-3, which GPT-3, the API for it, I think came out in late 2021. And then there was a pretty big update to it in early 2022. And so this is end of Q1 of 2022, I'd say something like that. He showed me it. And in the beginning, I swear to God, the main use case I had for it was at the time I was running a Dungeons and Dragons game with a lot of like friends in LA, which is I guess maybe the number one overlap between San Francisco and LA would be like there's an equal amount of people that like Dungeons and Dragons, I would say that's like the number one overlap. I was using it for that at first. And then what happened is I was working at O'Melvanie and I was assigned to this landlord tenant case. Landlord tenant, I think something that a lot of people don't necessarily realize that aren't lawyers is lawyers do not know everything about every area of law. And so you trained to be like a big law attorney at one of these large firms and the reason these firms exist is because you need so many specialists to handle a massive international merger. You need so many specialists to handle the IP, the antitrust, the contractor viewing all of these things, right? And so I didn't know anything about landlord tenant law, basically nothing. Just what I had kind of remembered from law school, some classes on property law. I started basically using GPT-3 to work on it. And the way that I started first was just asking kind of like general questions like here's the fact pattern, here's the statute and go from there. And then what my co-founder Gabe and I figured out is we started doing chain of thought prompting before I think like chain of thought prompting was not like a thing that people thought of. And now obviously the reasoning models are kind of in a way doing this automatically. But we started, it's really easy in lay in their tenant because we just said there are only so many statutes in California. And let's just make a prompt that is basically if fact pattern A, then go look at this statute. And then if you're looking at that statute, there's 20 subsections. And if you know fact pattern B, then you look at subsection one of this statute and answer the question based off of that. And so we have the super long chain of thought prompt over California landlord tenant statutes. That's what we did. And a lot of them work super well for like safety security deposits. Like what do you do if someone puts your safety security deposit in Bitcoin? This is like a famous one for us. And then they, you know, the price of Bitcoin goes down and they give you back to safety security deposit and it's worth like $2. Who do you so? And what we did is we grabbed that chain of thought prompt. We ran it over a bunch of questions from our slash legal advice, which is a subreddit for just consumers asking legal questions. I would not recommend by the way you go there for answering your legal questions. And we basically went on there and we ran, we took 100 questions and we ran that prompt over the questions. And then we gave question answer pairs to three landlord tenant attorneys. And all we said, we didn't say anything about AI, we said a potential customer, law firm customer, asked this question, here's the answer. Would you make any edits to this answer or would you be okay just sending this in an email as is on 86 of the 100 samples that we found two out of three attorneys or more more commonly three attorneys said they would send it with zero edits. And that was the moment where we were just like, wow, this entire industry can be transformed by this technology. And my co-founder, we just called the emailed Sam Altman and Jason Kwan, Jason Kwan used to be the general counsel at OpenAI. And we figured that we had to email a lawyer because otherwise the person is going to look at the outputs and they have no idea if it's right or not. And he said he didn't realize that the models were so good at legal and that you could basically do these like chain of thought, reasoning, prompts and things like that, and structure out these really good outputs. We then talked to them for a while and then on the morning of July 4th at 10am, I remember this specifically because it was July 4th, we got none of a call with them and kind of the rest of the C suite of OpenAI and we made our pitch. That's incredible. And did they, they wrote a check right away? Yeah. Can I ask how much, Sam, on to the company? It's the fund. Oh, the OpenAI fund? Yeah, yeah. Okay. It's the OpenAI fund. It's not individually by any of them. Is it a sizable investor? They are the second largest investor. Yes. Second largest. Okay, great. Well, I was going to ask because as mentioned, some of the most prominent investors in Silicon Valley. So I assume that maybe Sam introduced you to Elad or Sequoia Capital or Client or Perkins or Sarah. OpenAI did. OpenAI introduced to our angel investor at the time. So Sarah glow in a lot. Amazing. And then everything else from there, we were kind of doing it ourselves and things like that. And it's an interesting experience for me because I actually don't have many friends. Now I do, but I didn't have many friends that worked in tech. I didn't grow up in San Francisco. I didn't really know any of this stuff. Like I didn't know who the top BCs were. I didn't understand how you're supposed to fundraise. This was all just kind of not new to me. And I mean, thank God I had a co-founder that did. But a lot of it was, you know, figuring out all of these kind of like what the ecosystem is like, all of that together. And I think another thing too to think about with lawyers is most of the legal market in the United States or I'd say like the center of gravity is in New York, not San Francisco. And so the use cases that you need as a lawyer if you're servicing startups is very, very very different than what kind of the majority of lawyers do. It's been interesting kind of navigating the combination of tech plus law, which a lot of it is honestly centralized in New York City. That's really interesting. And it seems like most of your investors are here. Did you pitch investors in New York as well? No, we've only pitched folks here. Although we have definitely added folks that are New York based, like co-2 is an investor, et cetera. We also have some investors that are European based. So we recently added EQT. And so that was really helpful for us because, I mean, a, they're a fantastic institution. But B is helpful just because we're in now 63 countries. We kind of needed some folks that could help us, you know, at a very international scale. For someone who says they're not or wasn't that familiar with the VCC, you've raised $800 million to date. You just closed four months ago, I guess, a $300 million series E at what we are told is a $5 billion valuation. What specific milestones have enabled you to raise so much money for your series E? What was sort of the metric that everybody was looking at? So I might say something that maybe the VCC community might not love, which is I strongly believe that the best way to raise money is to just make sure your company is doing super well. And I know that sounds kind of like duh, or that should be pretty obvious. But I do think that there is a lot of advice sometimes out there about like networking and doing things like that. And to me, I think the most important thing is to spend almost the entire time on your business. And then find VCs that want to do that with you. I don't think you need to boil the ocean and have hundreds of VCs that are helping you out. I think you need to find partners and you need to find a few partners that you think are going to go the distance with you. Most things that the company have been successful so far, especially externally. And eventually like things are going to go wrong. Like it's just how it is, right? I think that if your goal is kind of going out and just making relationships with tons of VCs and that's what you're spending a lot of your time on, it's going to be really hard for you to actually find the few people that will be in your corner. And that's what you need. You need a few people that really care about your company. But also to be honest, they care about you growing. They care about you as a founder, right? And to me, that's how I've always approached fundraising. I haven't approached it as there is this art and you do as many boil the ocean pitches as possible and things like that. I've done it as 99% of your time focus on the business going well. And then spend time trying to find a few folks that you really, really think you can partner with and that will be there for you for the long run. So we can agree that it's not about the PowerPoint, it's about the metrics of the business. And we understand that Harvey is at 100 million AR. Is that correct? We reported that in, I think it was early August. So that number has obviously changed since then. If it's larger, that's great. But I'm assuming that you have something like 700 employees and no, that's not correct. Close to 400. 400. OK. So if we assume that they're getting sort of 20K per month fully loaded, how close are you guys to getting to a point where you're breaking even? Yeah. So this is a super interesting question. Actually, I think like compute costs are more expensive for us than a lot of other things. And the reason why is going back to the operating in more than 50 countries, there's data residency laws in all of these countries, right? So one of the biggest problems in, and this is slowly unwinding as the hyperscalers just give you more optionality and is open AI and anthropic give you more optionality. But for a long time, what you had to do is if you use multiple models in your product, right? You actually had to buy a bucket of compute, like there's a threshold, a bottom threshold in every single one of those countries. Even if you didn't have enough of the clients yet to support that cost, I'll give you two examples of this. Germany and Australia have incredibly strict data processing laws, incredibly. Like you cannot send financial data outside of those countries, right? And so you needed to basically have an Azure instance or AWS bedrock instance, et cetera, set up in every single one of those countries. We would set those up, but then we'd only, it's to close like three or four large German enterprises or German law firms. And it's like a one time fee and it's like a bucket and you had to buy the bucket, but the water hasn't been filled yet. And so my point with all of this is our margins look very, very, very good. If you look at this on a token base, so like if everything was just you're basically calling the API and paying for API credits, our margins would look incredible. They're worse because we have to basically spend so much on upfront compute because we're operating in so many jurisdictions at once, that will get solved over time. Like one of the main things Microsoft did recently is you can buy a bucket for a Mia. And then the minimum, like the minimum spend is a Mia instead of every single one of the countries, right? Tell me a little bit about the sales process. Let's say you've got three or four huge law firms, as you said in Germany. I mean, I realize that was a theoretical, but how do you expand? I mean, is it just word of mouth? Is it fear? I mean, is it companies? Is it law firm saying, okay, well, those guys have now reduced what they're charging their clients? We have to compete. And so everybody kind of like falls in line or like quickly, are you seeing things take off in different countries? Yeah. And I can do this in stages too. And maybe it's helpful to kind of give a quick snapshot of where we're at right now. The beginning of this year, it was about 4% of our revenue was from corporates. So 96% of our revenue was from law firms and 4% was from corporates. Right now, 33% of our revenue is from corporates. And my gut is probably by the end of the year that looks closer to 40 or something like that. And so in other words, we're selling very quickly to law firms, but we're also selling very quickly to their corporates. The reason I think that that is relevant in the beginning, what we would do is we would take public filings. So in litigation, if you use any sort of federal litigation, eventually most of those briefs, they get redacted and things like that. But a lot of them get published. They get published through something called Pacer. And so you can pull out those briefs and you can then find the partner that wrote it. And you can say, put them into Harvey and you could say, how would you argue against this? And they massively pay attention to that demo because it's relevant to what they just did. And so a lot of our sales process in the beginning was, you got to go directly to the user, which is the partner or the junior associates, et cetera. And so that's what we did in the beginning. But what was interesting about that is once we got a bunch of adoption at the law firms, right? My point here is that adoption came from not pitching them on efficiency. The adoption came from product. Like it wasn't a pitch about what is going to happen to the legal industry, a pitch on efficiency because they have the billable hour, right? You just want to go in as immediately as possible and show them the value of the product. And then they will just start using it and you go from there. That was how we used to do this. Obviously, this is like two years ago. But the interesting thing is this evolved where what ended up happening was you'd get like a large law firm like, you know, Latham is one that we talk about publicly. Latham will now go and introduce Harvey to clients. He'll go to the clients they're working with and say, hey, did you know that this is how we can do, you know, use AI to do XYZ. So what started happening was the law firms themselves would actually help us pitch to corporates because they want to collaborate in the system over time, right? Like that's mostly what the focus of the company is right now is making everything multiplayer. But I think what this resulted in is a situation where we had a sales force that only sold to law firms. But then we actually started selling to corporates and we didn't do it ourselves. Like we just recently started having a sales force that sells to like large enterprises. It actually was just through the law firms liking the product enough and going and talking to their clients about it. So moving it from an internal product to something that also lives between these two entities essentially when you say multiplayer, that's what you're talking about. Exactly. And right now our main product, like everything that we're building on the product side is how do you make this multiplayer? And this is a huge problem. I'll give you an example of why this is more complex than the general versions of this. You've seen a bunch of announcements recently from OpenAI, Microsoft, et cetera, about like shared threads, company memory, things like that. That problem is really hard to solve. You have to basically get the permissioning right. So these agents can go into the right systems, et cetera and the same systems that the human would have. It's a hard problem, but you're only solving it for one entity at a time. It's just Comcast and how do you solve that for the permissioning for Comcast? The secondary problem that we have is how do you solve that for Comcast plus all its law firms? So you need to get the permissioning right internally and the external permissioning right. And there's a concept in legal, I'm not sure if you guys are familiar with it. It's called ethical walls. What it basically is and why permissioning matters more in legal is think about a law firm in the Valley that works with 20 VCs. Tons of these law firms do that. If you're working on a deal for Sequoia, but you're also working on another deal for Kleiner Perkins, what happens if you accidentally give all of the data on the Sequoia deal to Kleiner Perkins? Huge, huge, huge astronomical problem. And so all of these firms support competitors, all of them do. The best firms in private equity, they work with all of the private equity firms. So we have to solve this really, I think, technically difficult problem of how do you do internal permissioning so you let agents work and get access to the same thing the humans have on controls. And then how do you do that externally, which I think is going to be very, very, very difficult? And if you get it wrong, you're going to have disastrous impacts on the industry. I wonder if there are also legal questions like does attorney, client, privilege apply when parties are sharing information through your platform, is your platform, something that can be subpoenaed, et cetera? Yeah. Oh, you have to solve all of these things. That's 100% right. They're solvable. I'll give you the best example of this email, like everyone collaborates in email, they send incredibly sensitive client documents through Outlook. That already happens. My point is the thing that makes it more difficult is you're trying to make it so agentex systems have access to these things. And so the control normally is you just hope, like there's nothing preventing a lawyer from grabbing a document from their desktop and emailing it to the wrong person. There's nothing that prevents that. That could happen. But very low chance of lawyers going to do that, it's a huge, you know, either the burden, the lawyers going to be thinking about it a bunch that doesn't happen very often. And then it's an agent now that you really have to enforce, like you really have to make sure that you're permissioning your ethical walls, all of those things are enforced because that agent doesn't have the fear. Hopefully, I don't think they have fear of how important this is and oh, I need to make sure that that dog that I randomly dragged from this place that normally blocks me. I'm not going to upload an email to a random person. Does that make sense? Yeah. I'm just wondering, have you solved this? It sounds like it's still in process. Oh, it's definitely in process. We're doing all of the security and the permissioning first. I think that the first version of this at scale will probably be done in December. And I think the nice thing that we have at least is because this is such a high percentage of our customer base are already corporates, they're already using Harvey. And so the security problem is much easier because they've already gone through security review. They already trust the platform. They're already uploading client data. And so if you have a law firm that's already done that security review process has gotten the permissioning right and their customer, their client, their in-house team that's gotten it, it's an easier problem. Right. Winston, how are lawyers primarily using Harvey? What are the maybe top three repeat workflows today and how is that mixed shifted over the past year? We'll do general and then I'll go into like much more specific like practice areas and use cases. General, I would say in this order, number one is drafting, number two is research and research is emerging because we just have a partnership with Lexus Nexus. So a lot of the data, it took us a while to actually get that the correct like legal grounding data. And then the third one is analyze. What I mean by analyze is a lot of what you're doing, especially as a junior associate is I want to ask 10 questions over 100,000 documents. If you think about what diligence is or what discovery is, the first step is just I have a list of 10 questions and I need to run those 10 questions over every single document. And I'd say those are the top three like general buckets of areas. And then I'd say in the beginning, we had much more transactional use cases, M&A, fund information was very popular and those are still becoming very popular and we're actually building modules that are specifically for matters or workspaces. They're just for M&A or just for fund formation. And then I think the area that's growing a lot faster is litigation and a lot of it you needed the data before you could do it. So really interested in this space, it seems like there are a number of companies that are moving in and some critics have said Harvey is just a rapper for chat GPT, I'm wondering how you would respond to that kind of criticism. Yeah. One in terms of the competitive side, the largest advantage that we have over time is two things. One, we're collecting a tremendous amount of like workflow data and what I mean by that is what are the main use cases and workflows that these models can actually do. And I think evaluation becomes a pretty strong mode because how do you evaluate the quality of a merger agreement? That becomes really really hard. What you have to do is set up evaluation frameworks and to be honest, like agentic systems that can self-eval, all of the different steps the user needs to take or the system needs to take to get there. So I think that's number one over time that becomes very, very, very difficult and something that I don't think the labs are going to be able to do or at least I hope they're not going to be able to do it faster than we can. It is definitely a speed thing. The second, I think, strongest mode is our product is becoming very strongly multiplayer. I think that this industry, because it has two sides, it has the providers of legal services and it has the consumers of legal services, you need to build a platform that is in between both of them and helping both. And so far, I haven't seen a competitor that is doing that. We have competitors that are doing what we do for law firms and we have competitors that are doing what we do for in-house, but I haven't seen someone so far build a truly multiplayer platform that is thinking about how do we make this best for both sides. I think those are the two strongest ones. And then in terms of folks saying that we're a chat UBT rapper, I think that the complexity of the use cases gets very, very difficult. And I think the reality is for 2023 and 2024, a lot of the power behind the product is honestly the model. It's the model plus a bunch of front-end work that makes it easier for the UI and UX for people to use it. If you're trying to build something that is closer to, I have 100,000 documents in this data room. I have 5,000 emails about this M&A. I have all of these different statutes and codes that are relevant to this M&A. And I want a system where I can ask questions and do work over all of those pieces combined with high accuracy. That is definitely the holy grail for us. That's what we want to build. And I think that we have created all of the pieces. And now what we've been building for the past couple of months and what we're really going to lean into hard next year is building that. And I think that once people are using that and seeing that, you'll actually have the, wow, this is a completely new way of doing legal work. And I don't want to go back. And I think that will look significantly different than a chat UBT wrapper, if that makes sense. It also sounds like maybe a company like UDIA should be concerned. I think that's a company that we've talked to general catalyst about. That's a company that's focused on the in-house piece alone. And I don't know that they're doing the multiplayer stuff that you're describing, but it seems like inevitably they would have to move in that same direction. The problem with doing it one side or the other. And by the way, UDIA is a great company and I think they're doing some really interesting things. I really like the idea of kind of like buying a bunch of the alternative legal service providers and doing a combination of that work. I think the interesting thing about professional services is I think people are drastically over estimating the death of professional services because of AI drastically. And I think this is for two reasons. One is the value that you get from someone who is doing professional services or legal. A lot of it is strategy. The problem right now is the economics of how law firms bill doesn't make sense. And I don't actually mean just the billable hour. The problem is junior associates looking at change of control provisions are 1,000 an hour. The best M&A advisor attorney on earth who could make it 70% more likely or something like that that you can close in M&A is 2,500 an hour. That doesn't make any sense whatsoever. And so my point of connecting this to the companies that are trying to do the low level legal work themselves is I think that that work gets commoditized very quickly. And the work that you want to be able to enable is how do law firms and in house teams do a combination of get that work done very quickly with software and then have the really high end important judgment work done on top. And that's why I think this is at the end of the day a software business with lawyers using the software, not you actually just become the law firm. Winston, I'm sorry to say I don't know what your business model is. Is it an outcome based pricing model? So right now it is mostly seats, but we're moving to more outcome based pricing. As the workflows get more and more complex, you can get there. But again, this is an evaluation problem. I think you want to do both. You want to do outcome based pricing for very like small things that you can make sure have the exact same level of accuracy as a human or better with very high speed. And for those you want to start doing outcome based pricing, but at the same time, the reality is you're going to want a lawyer in the loop. You're going to for so much of work. And so I think the main thrust for us still for at least the next year or two. It's a productivity suite. It is sold at seat based and it is multiplayer between law firms and their in house team. And then slowly over time, we will build more and more consumption based workflows as the systems get better and more accurate than humans in some areas. But it's not going to be like you automate an entire M&A. It's going to be their pieces of diligence that you can just have disclosure agents basically go through and automate the first pass of the disclosure schedules and then have lawyers jump into the platform and do the rest. Just out of curiosity, do your clients run more than one legal AI to avoid dependent C on you? And I don't know if potentially improve answer quality. Should they? Yeah. We have a tremendous amount of clients that basically will pick us as like their overall AI productivity platform and they'll do customization with us and things like that. The reality is we can't build everything. There are a few companies out there that have been incredibly impressed on, especially in like the patent drafting spaces, one I've been super impressed by and a couple other ones. Including these very verticalized point solutions. And the reality is we can't build all of them at the same quality that they can if all they're doing is focusing on that. So I do see a decent amount of our law firm customers, especially they'll buy us as their general productivity suite and they'll deploy that enterprise wide. But then for like some patent drafting, they'll buy a couple seats of one of these point solutions. I think eventually we will add more and more of those in, but I also like partnering with those groups and making it so you can combine them in the platform. One of the things the models are the best at is orchestration. You can think of like every point solution is a tool. It's just a tool use problem. So you can take any query or any like intent of a user and route it to the correct tool or point solution. How important is it for you to own your own data? You have partnerships with Lexus Nexus. But in an interview in the verge, their CEO was talking about protege and how they anticipate using AI. At what point do you need to acquire that data as an asset of the company? Super important to do our own data collection as well. Lexus is very aware that we are doing this. I love working with that team. They've been fantastic. I think that there's some data sources that it's very difficult to get access to and it will take a really long time. There's some data sources that's easier. It varies pretty massively country by country. I'll give you an example. In France, so there's primary law and there's secondary sources. Primary law is like what's the statute, what are the cases that have been litigated about this statute or case or things like that, right? What's the law and then what's the interpretation? Secondary sources are a bunch of experts that have written about what is the interpretation of this law and those actually in some countries are the law. It's actually funny. So France is like that and then the other one is Louisiana. Louisiana has Napoleonic code and so secondary sources are super important. My only point with all of this is there are different types of sources that are relevant for different practice areas A and then it can massively vary based off of the country and the jurisdiction that you're operating in. In some cases, it's going to make sense for us to buy data assets. In some sense, it's going to make sense for us to partner with the best providers in that space. In some sense, the easiest path is just collecting ourselves. We're going to do all of them. What percentage of your revenue is from outside the US at this point? I think it's about 65, 35. So 65% US, 35 outside. Also just because the staff popped in mind just now, I'd read somewhere that when you go into law from 90% of the firm uses the product, I'm just wondering what constitutes use? Is that opening a session? What percentage are super users inside of a firm? Our weekly active users is 72%. I think it is right now. I haven't checked it this week, but it's about that. It's pretty high in our daily active user. The difference between this year and last year is around, I think it's a 35% increase of DAU over MAU year over year. It's changing pretty drastically. The number one thing that is making usage go up, maybe it's two things. One is you need to integrate with all of the systems they already use so that you love the Microsoft suite, right? So you want to integrate with Word, you want to integrate with Outlook, you want to integrate with Lexus, Nexus data, you want to integrate with all their DMSs, you want to integrate with their billing tools, you want to integrate with their security tools, you want to integrate with all these things and a lot of that is partnering with the industry. And then the second piece is you want to basically expand and collapse your product. And what I mean by this is you want to expand all of the different use cases. So tons of different workflows, different UIs, all of these things. But then you keep the interface incredibly simple. And so the way that you collapse it is you use the models to do orchestration. What do I mean by this? You want to get to the point where if a user types in, I am trying to buy a company and I need to check for antitrust, you know, merger control guidelines in these 10 countries, they press enter. It's not a chat response. The first thing that pops up is, would you like to use the merger control workflow that is incredibly accurate at doing exactly the task you want? Then the user clicks that and there's no prompting. It just goes through and it says, upload the targets, financials, upload the requires financials. Okay. Is this right? Okay. Which countries are you interested in? Type in the countries, press go. And then it will show you, in up to 75 countries, whether you need to file for antitrust or not or for merger control, basically, in that country. So my point is, you want to build out all of these really specific use cases. And the UIs going to be different. Chat interface is not going to work for every single UI. But you can have it. So the entry point to your product is as simple as possible. And then all you do is categorize all of the different types of requests or intent of a user. And you route that category, that request to a different part of your product. And that's where we're seeing the most growth. Are VAU over MAU for people that use five products is 74%. So that's the main goal. Is how do you get it to folks can, based off of their query, they get the correct part of the Harvey product. That's the number one goal for our company on the product side. Are you seeing law firms hiring different people because of your product? And what happens to those junior lawyers who are no longer getting the apprenticeship that they might have had in the past? We could do the entire podcast on just the second question. No, the second question I care about actually potentially more than anything else at the company. And it's because that was a junior lawyer, like very recently, and I have so many friends that are. Because our firms hiring more people to, because of products like Harvey, the answer is 100% yes. The best way to look at this is like the legal cam overall, it's one to 1.1 trillion. The legal tech market is 39 billion. Tech penetration is incredibly, incredibly, incredibly low in legal. And so what's happening now is law firms are saying and enterprises are saying, wow, technology is going to be a bigger part of our day to day life. We need to hire more technologists to help figure that out. So that's happening 100%. I expected to happen way more next year, way more the year after that, and after that. And I think it's really good for the industry. I think it's fantastic. I think lawyers will have better access to technology. They'll enjoy their jobs more going to the second piece. My gut of what happens here, the goal of law firms in the next five years, 10 years, et cetera, is how fast can you train the best partners? That is going to become the goal. And I think right now that's partially the goal, partially the goal is, we hire armies of associates and we build them out a lot and that's how we make profits. But I think whether it's because a lot of things become outcome-based pricing, which definitely could happen, flat fee, or to be honest, it's because partners can actually charge more if the AI systems can't do what they do. Either way, it doesn't matter. The most important thing financially from a business standpoint for a law firm is to make sure that you are hiring and training and developing lawyers that get to being a partner as fast as humanly possible. And my thesis for this is if you can build tools that can do the first pass of an M&A, that is a one-on-one tutor for a junior associate. That's what that is. We work with a lot of law schools, like a lot of law schools and I care so much about working with the law schools and making sure that that happens. You can imagine at some point some of the training that you have, you have an AI merger that you do in Harvey, like that's literally what you do. And the system's teaching you how to do stuff that's giving you lifetime feedback, all of those things. That's an incredible training system, I really truly believe that. I believe that if you can build systems in these verticals, whether it's in the medical space or whatever vertical it's in, that can actually do a lot of the tasks, there's no reason you couldn't turn that into one of the best education platforms possible. So I think it'll be very good for the industry. It'll be rough in the beginning, like there will be some adjustment, but I think the reality is the life of a lawyer, the job of a lawyer. Just because the incentive alignment is going to go this direction for what firms need to do and pay attention to is going to be really, really healthy for the industry and really healthy for reducing burnout in all these other firms. It's great to hear that law schools are leaning in here. It's interesting. I had talked to the chancellor of UNC Chapel Hill recently and he was saying that the faculty is really divided. Half of the faculty is really excited about using AI and the other half is not. And he's frustrated because he said nobody's going to get fired from their job out of college because they're using AI. So it's really important to him that everybody be teaching it and embracing it. When some, we've kept you so long. It's been such a pleasure to talk to you. I wanted to ask one overarching question. You are more than 100 million ARR as of August, as you said, 800 million raised. How much does this business have to raise? Does the next capital event stay private? Would you think about public markets? I mean, there's going to be a lot of excitement over this company, obviously. Well, thank you. Thank you so much. Fundraising large rounds is not something that we have planned anytime soon. And by large rounds, I mean, you know, similar to the ones that we did earlier this year. We don't need that much money and we aren't burning like a crazy amount. The reason I did a lot of the fund raising I did this year is there are a lot of research directions that are going to require a lot of compute. And we are very interested in deploying that eventually. It's very much in its infancy on what will work and what won't, but we wanted to prepare ourselves for that. And then in terms of like how much capital do I think we need to raise over time, I think a lot of it is how much is it going to end up costing to do something like a company is doing a massive merger. Can you very quickly on the fly basically build an RL sandbox for that M&A where you can automate massive parts of it and then open it up to all of the professionals to go in and edit it and make sure everything's right. If that direction works, there's definitely just going to be some high like R&D costs to make that work. And then in terms of your second question of are we interested in the public markets and things like that, that's definitely what we're interested in long term. I can't give you anything close to a timeline. Sure. It's just some people seem very stubbornly opposed to it ever. One last quick question. You mentioned that penetration is really low. I'm curious to know how low is it because that obviously also underscores the opportunity ahead. Oh, yeah. A good point. Yeah. It's just like what percentage of the lawyers on earth are using Harvey right now. And it's actually a super, super low percentage A. I mean, we have a lot of users, but there are eight million or nine million lawyers on earth. So it's a lot. The second piece, I think, is the more interesting point to the penetration side is we are unbelievably early innings on how complex of work these systems can do just incredibly early. They're very helpful, and people are getting an incredible amount of ROI. But if you think about what percentage of legal work can these systems do today, it's so much lower than what I think it can do in the next five years. And yet a lot of attention is paid to lawyers who write briefs and chat GBT and are scolded by judges for doing so. I mean, I think people kind of underestimate the ability of these systems. Oh, they definitely do. Yeah, I'm just saying with like super high accuracy, and drafting is a weird one where, depending on what you're drafting, you don't need to pull in that many resources. But if you're drafting something like a motion for summary judgment, you're navigating, in some instances, millions of pieces of data to draft that one document. And so that is kind of just a next level. One quick way to think about this is think about the use case as what is the value per token? That's how complex it cost to generate that document. So an example of this is the legal fees for a merger could easily be tens of millions of dollars. The art of fact that you have after that merger is a merger agreement. I mean, there's other ones, but it's a merger agreement and an SPA state. The merger agreement is 100 pages, the SPA is 100 pages, whatever. That is the value per token on that SPA and M&A for the 150 million, 20 million, 30 million of legal fees that you had to pay to generate it. Those are the type of use cases that when I say that we're in an incredibly low penetration, it's that. We aren't at the point where you can do something like that. And the value of being able to do that accurately is incredibly high. Everything you've described sounds so dazzling. It must just exhilarate and also really terrify law firms. I mean, you'd be surprised, like, we partner with all of them, right? Like in the US, I think now we're at 51 of the top 100 firms, it is like a ranking called like the Amla. And it's not fear. It's actually a lot of them are very much like forward thinking. Let's think about new ways to work. And the most important thing that they all think about is like, it's a client business. Like, that's what it is, right? And so mostly what they care about is how can we adopt these tools and help deliver better solutions and better outcomes to our clients? And I actually think like that's how the industry is thinking about it. It isn't operating in fear. It's operating in how can we figure this out together in terms of how our legal service is going to be delivered in the next couple of years and what do clients want? Or how can law firms band together to invest in one of these companies and pick a winner? That, that we don't do. I guess that's not going to be us then. I think that's a quite a bad idea. Winston, really great talking to such a pleasure. Your enthusiasm for your company and this industry is really contagious and we're excited to watch what happens from here. Thank you. The Strictly VC download podcast is hosted by TechCrunch Editor-in-Chief Connie Loizos and me, Alex Gov of Strictly VC. Strictly VC download is produced by Maggie Nye with editing assistance by Theresa Lankan Solo and Kel. Thanks for listening. We'll see you back here next week.