In Conversation with Saam Masshad and Rami Karabibar, Co-founders of EvenUp

This interview was conducted in Spring 2023. It has been edited for brevity and clarity.

Asher Noel (AN): Even if it is in the news a lot, 5X-ing [5X increase in] revenue over this past year is really incredible in this environment. How have you noticed, how are you guys thinking about the application of all these new language models to the law? What are the most important things to know about legal tech from someone who just uses ChatGPT?

Saam Masshad (SM): There's definitely an answer to this. I'm just curious about what we can say that would be interesting to your readers without being too revelatory.

Rami Karabibar (RK): The comment that I would send, let me know what you think about this, is that I think it's a really good idea to have a legal team that can work on the legal side of things. Let me know what you think about this. Specifically in our use case, there's a lot of mundane manual review that is better than by a machine, than by a human. And so when you put yourself in the shoes of a case manager or paralegal that has to sift through thousands of pages of medical documents every month where you're looking for the same information time and time again, and you're summarizing that information in a specific type of format time and time again, up to a hundred times a year. The way I'm going to put that is it's not exciting work that folks really want to be doing. It's better done by a machine than by a human. And so I think the first pass of a legal draft will always be done by a machine. I don't think it can go all the way to fully automating the whole piece of work anytime soon. But at least that first draft can get you a lot of the way there and save folks a lot of time in the process. So I don't know if I can add anything else to that, Saam, but that's my view.

SM: I think the way I’d say it is that there are two parts to this. There's the input function and then there's the output function. The output function is ChatGPT, but the language model that it generates. And then the input function is the prompts and then the things they put under the prompts. I think that area is going to be really, really interesting for the law because it's a problem that hasn't been looked at as aggressively as a human typing into a chatbot. The input is often as important if not more important than the output that's generated by some model. So really focusing on that is going to be a really important arena in the law in particular because to get to the right output, you need to parse through a ton of information, extract the right information, and then put it in some appropriate manner to a chat or whatever other model that you're what other language model that you're using.

AN: Okay, so a lot of that resonates with prompt engineering, so to speak, for the law.

SM: Yes, that's partially part of the story.

AN: And then, as Rami mentioned, a lot of summarization, a lot of probably image recognition and processing, just very broadly at a very high level. So not like a UI path, which is just image recognition, but also a lot of synthesis. So there's been commentary that GPT-5 is writing lease agreements at a level better than first-year lawyers. Is that consistent with how you see all forms of not just criminal, but any crime law really developing?

SM: I think it's going to be really a function of the training data that GPT-5 is going to be subject to. So there are areas of law where there is widespread data of good legal output, and then there are areas of law where there is not widespread data or available data of good legal output. And being very parsimonious about the data that you use to train your language models for any specific area of the law is going to be incredibly important. If we're talking about corporate law, and we are not in corporate law, but you can imagine that there being a lease, you know, a contract drafting model for Goodwin versus Cravath versus some other big law fund. And the distinction between all of these is going to be, these large law firms are going to be the data that they use in order to feed, to build their models. You can think about whoever builds the best model is the law firm that has the best underlying data.

AN: That seems like it would really advantage the existing, most prestigious, largest law firms.

SM: Assuming that they actually do everything in time, which they haven't.

AN: Or the platform that has most of them as customers first, which really prioritizes first mover.

SM: One thing I'll say about that is these law firms will never allow you to enter into their law firm and take their data and share it with anybody else. That will be something that none of these law firms will ever, ever agree to. You could think about the unique advantages of some of the language models being unique to some law. Especially when it comes to domain-specific knowledge and specific techniques in writing or some conjectures they may make or strategies that they may undertake in preparing certain agreements. There will be some idiosyncrasies at the law firm level that will not be allowed to escape the law firm.

AN: It would almost be as if there were firm specific models. Like Cravath has their own model and they use that to do everything internally. Even thinking about tailoring models to each firm in the crime space?

SM: We're in the personal injury space. We're very, very different. The most important thing for us in driving value for our customers is understanding medical records. In a weird way, we are in a legal tech space, but to generate a legal output, we need to be able to understand medical records. The understanding, incubation, summarization and analysis of medical records and drawing the nexus to certain legal concepts is what makes us work versus what would make a large corporate law firm work. They are very, very different spaces, so it might be helpful to draw that distinction for your readers.

AN: That's very helpful. At the same time, you all still come to predictions about damages, come to predictions about evaluations, how these different pleas might fare in front of a judge. Those, I'd imagine, are probably similar engineering tasks to some degree if the court opinions are all public or on some database that's accessible by a broad number of people, right?

SM: There are decisions from the bench that are reasoned, where there's a decision behind it, and then there are verdicts. Most personal injury cases are derived from verdicts. And verdicts have no underlying reasoning. There is a court reporter that is sitting in the back that is then taking notes or looking at the court records after the fact and summarizing them. And then that is what is published on jury reporter websites or publications. And that's the information that we have to parse through and use in order to assess what a pain and suffering amount would be for a back sprain in Boston.

AN: Okay. So it's mostly material produced by reporters, as opposed to a corporate law case in the Court of Chancery by the Vice Chancellors, which is more unstructured and more error prone, less trustworthy, but probably a harder engineering problem and therefore more valuable when it's solved. Which is great for EvenUp.

SM: That's precisely correct. And then the other interesting aspect of this is that there is a morass of unearthed data, right? One or two percent of cases go all the way to trial. There is a rich set of settlement data that can be interesting indicators of what injuries could be worth as well. So those are the two data sources that we can look at in order to describe the damage. It's a two-step problem. Step one is to identify the value drivers by looking at a set of records. And then number two is to map those to a constellation of data points that are responsive. And those data points, again, can come from settlement data or they can be derived from jury verdicts.

AN: How was it that you two found personal injury to be the space that you both wanted to dive into? Was it top down? There is all this unstructured data. It's a hard engineering problem. We don't have Cravath’s database. Or what was that process like? I guess personal entrepreneurial history to some degree.

RK: I'll take the first pass. We all have personal tie-ins to the space. So the three co-founders, I used to work in mobility companies, Waymo and a food delivery business startup before. He was on the other side of these claims on the defense side and used to work at CHOP. And our third co-founder, he has the most personal story, but his dad was in an accident that left him with a permanent disability. And so that's his tie-ins to the space. The way we thought about personal injury is we almost thought of it as more like an asset class where you have more personal injury cases in the US every year than there are house sales, just to give you a sense of magnitude. And yet despite that, ultimately it was a gut call on what was the right number to actually settle these cases for. And so as you think about, hey, there's 20 million injury cases. There's no great stance on this, but approximately 20 million in the US every year. And because of the nature of settlement data and how 99% of these cases are settled, a lot of the time these injury attorneys are using their gut to come up with a number. And we knew it's only a matter of time before there's more transparency in this space. And that's really a big reason why we called ourselves EvenUp to begin with, is we want to make sure that these cases are won based on their merit and not on the opacity of the space or on, again, how overworked a paralegal might be in the document review phase of what we talked about before. That's the premise here. Big space, personal tie-ins, and one that, as you think about it, inevitably will be a much more transparent market than it is today.

AN: Do you think that transparency will carry over probably to a lot of other asset classes within the law? And then, to follow up, how quickly do you think these transitions are going to happen? What are the big things that we should watch for?

RK: My view is, especially in our domain, client interactions are just really important. What I mean by client interactions is having an attorney or a paralegal or a case manager walk a plaintiff through the process is always going to be a thing. I can't imagine that going away anytime soon. And do I think the manual document review of coming up with template letters, is that going to stick around for the next 10 years? Probably not. So I think that's the biggest change, but I would say I'm pretty skeptical on, hey, we're not going to need paralegals anymore, we're not going to need personal in-duty attorneys anymore. I don't think that's going to happen. And the analogy I like to paint is folks have probably said similar things about real estate agents at some point in the past, given the transparency of data in that space. But again, it's such a personal journey that these folks are going through, and I think this applies to other types of litigation as well, but you're going to always need to have somebody to speak to.

SM: What I would add to that is, if you think about your life and what is significant in your life, you know, birth, graduation, getting your first job, death of a family member, and maybe getting injured and dying, right? And getting injured is a traumatic event, and thinking that a robot on a standalone basis could handle that part of your life for you, I think is a little bit naive. But I think what is not naive is to believe that the mundane, manual data extraction summary analysis that can be computerized would still be done by a human. And I think that is the aspect that we're focused on and worried about. So in a way, we're trying to make a lot more human by allowing people to focus on the customers, on the personal injury victims, and that's around the paperwork and just the morass of work that surrounds the extraction of small little details to get a particular outcome.

RK: And to what Saam mentioned, I can't speak for all parts of law, but for personal injury, a lot of folks want to, I mean, they go into this space to be advocates for their plaintiffs. They don't go into it for a manual document review. As Saam mentioned, I think they want to be there with their clients. That's what gets these folks jazzed up, as opposed to how excited can a human being be that I'm going to read 100,000 pages of medical documents in a year? It's just really painful work. As we mentioned earlier, it's thankless work as well, but it's also one that's better done by a machine than having somebody just have to do this every day for the rest of their careers.

AN: Are there any last words you might want to say to a college audience or to students, whether it's just, okay, show up on time, or, you know, exercise, anything you'd want to share with college students or anything broadly?

RK: Anything I want to share? I don't think it comes to mind immediately, I guess. It's like, there's so much... I mean, I'm sure these folks get overloaded with advice, so I don't think I have anything unique to add to that equation. I don't know, Saam, if you have anything.

SM: I don't think we've entered that phase of our careers where we can start giving inspirational advice to young people without being sure we're giving them good advice.

RK: Actually, I know one that we heard, these are not my words, but like a person who runs a close to a billion dollars of ARR type of software business, and his words were, what, Sam? Like, even when you're going from, like, zero to one million of revenue, one to ten, a hundred, like, ten to a hundred, a hundred to a billion of ARR, every day is a fist fight. And it seems unattainable, but it doesn't actually change. You just have to keep going at it, and it'll eventually work out. He has a little bit more colorful language, but we're just trying to keep it PC for the audience. Every day is a fist fight. You just have to keep going. Whatever feels impossible at the moment, he would have never thought he would have a billion dollar ARR business, but you just gotta go there one step at a time.

AN: Terrific. Thank you so much for the time.

SM: If you don't like this Insight stuff, just message us… That's our advice to you. Do software.

The Harvard Undergraduate Law Review is thankful for Saam and Rami’s time and expertise.