Pérez-Llorcast TechLaw #16 | How AI is Transforming the Legal Sector (Max Junestrand, CEO at Legora)

So the founding idea comes from several years back. One of the co-founders' roommates, he was a lawyer. He spent four months just summarizing and putting together court cases. Why do I study law for four years just to do this thing? He expressed frustration. The best tools are going to be created when you marry sort of machine intelligence knowledge and how to build great product with domain specific expertise. That's what we're building with Legora. It's not a replacement for the legal professional, but it's an empowerment of how can I do more and better work with the time that I have. So what are the peculiarities of our industry? Everybody was saying how conservative the legal industry is and we're seeing the complete opposite. I think that we both work with code. You guys work with computer code and we work with legal code. So how do you think it's Legora being the operating system for lawyers moving from being a system that you go to solve a specific thing, right? Like you kind of need to be more integrated in all of the daily work. I like to view like every feature as a boat and then, you know, when the model rises the sea, then everything gets better. Yeah.

Thank you so much, Andy. I'm grateful to be here.

Yeah, absolutely. So the founding idea for what became Legora comes from several years back. Back in college, one of the co-founders' roommates, he was a lawyer. And he was doing these big internships, you know, working in big law, working in a court. And what ended up happening is he spent four months over an entire summer just summarizing and putting together court cases. And he expressed such, you know, frustration over that type of task. And you know, why do I study law for four years just to do this thing? And back then the early AI models weren't powerful enough to really solve those things like BERT and the early transformer. And what ended up happening is when ChatGPT came, it was the paradigm shift needed to, you know, excel the amount of use cases and the tools you could build. Since then, we were very quick to realize that the best tools are going to be created when you marry sort of machine intelligence, knowledge on how to build great product with domain specific expertise. Yeah. And so we started building a tool that you could upload a single contract into and you could have a conversation with it. And this was before, you know, chat with my PDF existed. Okay. We had that little demo, that little tool, and then we realized, wow, this is, you know, the capabilities here are so much larger, grander than this back in 2020 around back in 2023. Okay. So the first sort of MVP we put together was very simple and then we realized, wow, okay, we started showing it to people and the feedback we got was tremendously positive. And so, we said, all right, let's put everything else on ice. I dropped out of university and we started working on this. Since then it's just been an explosion both in like what we can build as the models improve but also like the quality of the software. Yeah. Before GenAI I feel like there was, you know, you ended up having 10 different tools, you know, one for comparisons, one for translations, one for due diligences, right? And now generative AI like kind of solves for a lot of those use cases. And so integrating them into a cohesive suite of features gives a lot of extra value. And that's what we're building with Legora. It's not a replacement for the legal professional or for the task, but it's an empowerment of how can I do more and better work with the time that I have.

I didn't make that much per hour, but before university I was playing Dota 2 very seriously.

Oh yeah. I mean we spend our time doing nothing else but this and we're still, you know, basically keeping up or trying to keep up. And I think to just finish on the point you made earlier, a lot of the tools didn't used to work in non-English languages and with the GPT models suddenly became possible to work in Spanish or Swedish and I think that also unlocked, you know, a completely new paradigm in terms of what you could build. Yeah. But no, coming from that background, obviously we're here to win, build great products and we're not settling for anything less. I think in the legal tech market there has been a trend where like successful startups end up selling to the much larger players like Lera or Thomson Reuters. Sadly enough, a lot of the functionality and the sort of the fire that keeps those companies going in the early age, they die out or they blow out, right? And so where that's a very good point because I think that and finally they became more rigid or less agile or less creative as initially it was they were intended to. Exactly. Right. And the way that we've approached things is how do you rethink a lot of the sort of the knowledge that you have today and how do you like truly empower collaborativeness and the lawyer of the future which we're doing.

So starting to work in the legal space, like one of the things that I had to do was very quickly get up to speed and learn. And I think we've been able to do that in a couple of ways. So one, there's maybe 10, 15 lawyers now working at Legora internally and we very quickly said, all right, if three software guys go into a dark room and try to think about what the future of legal work looks like, we're not going to come up with the best solution. So for me, I've approached everything with a very open mind. It's been about, you know, all right, like how do you take a process like due diligence or a process of document review or drafting and then you need to break that down. And I think I approach it with some level of naivety to a good extent in terms of like you're not locked down with, all right, this is how it's always been done so we need to keep it that way. But you need to honor the existing ways of doing things. And you know, one of the things that maybe has surprised me the most is everybody was saying how conservative the legal industry is and how, you know, lawyers work on the billable hour. There's no incentive to ever change and we're seeing the complete opposite, right? Like we're seeing firms and in-house teams adopt a very strong change management perspective and to think about, all right, what does this imply for our existing service lines? You know, how could we offer this type of service with a faster turnaround with a higher quality? How could we rethink the way that we collaborate between law firm and client? Because I think there's just a ton of work to be done. And historically, I think a lot of effort has gone into building great tools in this space, maybe because the technology wasn't so exciting. The coolest thing you could do was like document automation and you know, maybe that's not the...

Yeah. And I got to say it's also very large differences between the firms, right? You see that?

Some firms are still, you know, discussing whether or not to be on prem or on cloud. Yeah. And I think that that was a conversation to be had like a couple of years ago. So there's very large differences in the market and sophistication.

This is a great question, Andy. And almost all the prompts inside of Legora have been developed by lawyers. Yeah. So I think it's easier to teach a lawyer, like the domains, who has the domain knowledge about how to prompt and how to quote unquote code an LLM than it is to teach an engineer, you know, how to be a fantastic lawyer because it takes years of not only training but then practice. Most of the lawyers that are working at Legora today have some sort of technical interest, right? Like they've maybe previously been head of legal tech at a firm or they have been coding on the side. And I think the marriage between them and when you kind of put two people in a room and then you say, all right, this is the problem we're going to solve. Now come up with a solution. That surprises me every day. Just to give you an example, a week and a half ago, we decided that we're going to build out a bespoke playbook feature into our Word add-in. Like we decided a week and a half ago that we were going to build it. Last night we got a live demo of it in the office and it was working fantastically. And then it's one of the engineers who's just, I think, had a lot of Red Bull this week and was working hard and then the lawyer in the team, Robin in this case, who was working with him through the entire journey on basically encoding his, yeah, his way of, okay, let's say you have a playbook and let's say you have a contract, how do you then apply all the rules and how do you apply all the questions and if that then what, you know, so if this is breach, what are you going to do? Well, maybe escalate to partner or suggest this standard language. And the process of having done, you know, thousands of those reviews himself and then just encoding it into an LLM that we now can sort of share with the world, I think is incredible. And in many ways we're just building the building blocks for people like yourself and others to kind of, you know, expand upon it and build the next version and to apply it not only in a client specific need but a domain specific world.

Well, there's a different, and now it's you interviewing me, which is good. Means that the conversation is flowing, but I think that there it's like always, there's different speeds probably. The lawyers in the IP department or in a, with a high pace, or the departments like M&A that the pace is very high. So they are more open to new technologies than other partners. But the whole context also, I think that at the end sooner or later all partners will have to adapt. So definitely there's some differences but I think that sooner or later all partners move forward.

And maybe to give a high level overview of like exactly what Legora is and what features we have. So we've come at it from first of like a pretty general point of view and then with that toolbox we want to solve for a lot of different use cases. So the platform is built into three main features. It's the assistant which is a conversational based chat. It's tabular review which is essentially, you know, how do you fire off 10,000 parallel questions, you know, you mentioned due diligence or e-discovery of these types of tasks. And then it's the Word add-in that's supposed to serve as sort of your bread and butter co-pilot inside where you're already used to reviewing and drafting. Inside each of those we leverage a combination of models. So from very early on, actually we couldn't afford fine-tuning models back in 2023. So we said, all right, let's not fine tune but let's build an architecture in which it's very easy to swap out the models. But...

Yeah, and I think it's like you're wasting money, you're wasting time, you're burning the world's resources, you know, just improving a model, let's say 5%. And then next month a new model comes out. So I like to view like every Legora feature as a boat and then, you know, when the model rises the sea, then everything gets better. Yeah. The reasoning models are allowing us to, well, in short, solve more complex tasks. But it works like the kind of pre-plan subtasks or into a plan and then you can, depending on how big it is. You don't want to wait for 5 minutes for a response. So you might use faster models to then go and run all the subtasks. Let's say you upload 20 documents and you might run sub-models on each of those and then you bubble it all up and then give a broader response. The other thing that o1 is incredibly good at is being more exhaustive. So in the Word add-in for instance if you say, you know, my Spanish still isn't great although I studied it all throughout high school. But let's say you want to change all the wording from, you know, feminine spelling to like the, or let's say you translate a document from English to Spanish and then you need to correct all of it then it can do that across the entire document or you can say, hey, here's a playbook now run this playbook across, you know, 50 clauses and GPT-4 wasn't capable of doing all of that but o1 can because it's more capable. And I think from an application perspective, it's fantastic. Like we just want the models to get better and preferably cheaper because that means we can do even more exciting things.

Exactly. And it's not work that you'd preferably want to bill for either, right? It's like a write-off, of course. And I think it's about, well, how do we, you know, deliver more value faster and have less write-offs, right? Or to expand the terms of services that we can offer. To also touch on like the more legal specific work we can do, one of the first kind of end-to-end tasks that we did was the creation of a limited partnership agreement report. So an LPA, it's like this long, long agreement stipulating like the terms of a fund and you might ask your law firm or lawyer to, you know, read through all of that and create a report that's more digestible. Yeah. And you can also leave like a comment in terms of whether or not thing is market. And running this extraction back on GPT-3.5, we got to like 25% accuracy or like 25% of the answers were equally good or better than the manual version. And then with GPT-4o when that came out and we like redid the exercise and also with better prompt tuning from our side, we got it to being like better result. Yeah, it was equally good. And that means that this used to be a thing that an associate had to spend 3 days on putting together and now they could get a first draft in, well, 5 minutes that they can then sort of serve as a starting point. And now with the integration with Word, you know, we can also take that extraction actually draft it in the firm style.

So one of the first things we ended up building was like a citator, right? That built in citations. So when Legora was providing a response it was using the, or it was grounding that answer in the underlying content and it was showing exactly where it was grabbing it from. Hallucinations still aren't like solved. But this way of working like drastically reduces it. And we aren't yet at the point where Legora or any other system delivers the full end deliverable. And the way that you work in a premier law firm is, well, the associate, you know, the partner comes up with the plan, an associate starts with a first draft, it gets reviewed by a senior associate and then it gets reviewed by a partner. And Legora fits into that sort of staircase. It's not, oh, I give it the task, I give it the prompt, then I take that and I just send it to the client. No, it's part of the review ladder of a premier law firm, right?

Yeah. I mean, so the hot topic on LinkedIn is agents, right? And right now it's...