He invested. Max told us like, "Hey, we're going to take some time off from selling." Always what you want to hear, right? First board meeting, we raised 35 million and look these guys in the eyes and say, "Yeah, we're not going to sell for the next four or five months." The first time we ever chatted, you were in Estonia at a conference. It was like 11:00 p.m. and you were drinking at Red Bull. If we were to introduce you in a novel, that's how I feel like your character would be introduced.
Today's unsupervised learning was a particularly fun one. We had on one of my portfolio company CEOs, Max Junestrand, who is the CEO and co-founder of Legora. He is at the forefront of applying AI to the law industry. They're working with many of the top law firms around the world. They've raised over 100 million dollars and honestly one of the fastest growing AI applications out there. I was joined by my colleague Logan Bartlett and the three of us hit on all sorts of things including Max's take on the future of the law space, how he thinks about product prioritization given the rapidly improving models and what he thinks about building versus leaving to the models. And we also talked about what it's like expanding across a bunch of different global markets. This was an awesome conversation. Without further ado, here's Max.
Well, Max, thanks so much for coming on the pod.
Yeah, thanks for having me.
We've been excited to do this one for a while.
So have I. It's great to be here from Stockholm.
Yeah, we've got all our Redpoint podcast hosts today.
Yeah, exactly. This is my first appearance on Unsupervised Learning.
Yeah, we've tried to keep you out for a long time, but it was you're destined to wear a stack. We don't want to water down the quality that is this podcast.
I appreciate you having me on. We'll try and do like an NIL crossover episode.
Yeah. Yeah. You keeping me in my little box on my podcast. Well, Max, a ton of stuff I know our listeners will be eager to hear about. Maybe to start, can you just contextualize for us like where are we in like the AI meets law space today? Like what works, what doesn't, right?
Well, when we started, nothing was working. We were using the early BERT models from Google and I mean they were decent in English, horrendously bad in Swedish and this was back in 2020. So when GPT arrived 3.5 that was like the paradigm starter if you will. Since then I think we've moved from full on experimentation and trying to get stuff to work into actually implementing things that are well really taking like end to end work deliverables. Just to give one example like if you're doing a due diligence today you're not going physically into the data room you're not using F you're just taking all the documents putting them in Legora saying what you want to find and then it finds it and then based on that those findings things we generate the report. And so things are really starting to move from empty queries against like a data set to okay this is the process that we want the LLM to follow and what we do is we give an agent access to tools it then plans executes on that plan using the tools and then we get an end to end work deliverable that's actually usable.
And where do you think this is all going? Obviously these models get better every like 3 6 months. If we were to ask you to have a crystal ball and say where the legal field's going in 5 10 years, what do you see ahead?
Well, so the interesting thing with the models getting better, that's one piece, but what we're getting the most leverage out of is actually all these surrounding frameworks like function calling, tool calling, MCP. And the legal software space has been incredibly fragmented. That was like one of the early things I saw coming from outside the industry. You had one tool for translations, one tool for document comparisons, another tool for searching and another one for reviewing. And now suddenly all of this is kind of getting baked together. And you can of course also imagine a scale of sort of the complexity of legal work and at the bottom you have something very simple like just finding data extraction and at the top you have something really complex like drafting a share purchase agreement. And we've already started to fully automate a lot of the bottom quartile in this graph, right? And we're sort of slowly but surely moving up. And I think here the really interesting thing is for law firms and legal professionals to see where are they going to add the most leverage. Where do we need their specific expertise, their context, their way of instructing the models, and where are the sort of off-the-shelf LLMs just plainly good enough as they are?
I'm sure we have a lot of people listening that are aware of the legal field, but and the implications for it, but can you maybe speak to some of the elements of law and why it's uniquely well suited for AI as a potential application?
Yeah. Yeah, sure. So I don't come from the law space. I was an engineer before and when we first arrived I think the first thing that you notice is there hasn't really been that much software developed in this space. There's industry specific incentives that perhaps doesn't always align and incentivizes being more efficient and using software. So basically the hottest thing or the coolest thing that you could do was use a templating system. And what you basically have is this dilemma of law firms and in-house councils and they do very different things. In-house councils primarily work with the same stuff over and over again. It might be NDA reviews, MSAs. You're controlling the risk for the business. And when you work with the law firm, they typically do more one-offs or complex projects or things where there might not be as much precedent and so on. And I think very broadly you can sort of stitch it up in reviewing, reading, drafting, writing or researching. And in the beginning I think every piece of software in this industry was very focused on one of these. But because AI is kind of capable of doing things across this entire stack, platforms like Legora have emerged that are not solving a point solution but rather serving this entire wall-to-wall of needs. And what we've really found is law firms want to lead in this new paradigm, this new future. And for law firms that are not really leaning in I think you kind of risk two things. One is not upskilling your team in this new future and this new paradigm in terms of how you're supposed to work. But increasingly so clients are putting pressure on law firms to make this shift because they're starting to use tools like this internally. A lot of CEOs you see on Twitter and LinkedIn posting about we're going to go AI first, right? And you can't motivate getting a new headcount unless you can prove to us that you've been more efficient. I think it was actually the Fiverr CEO who also mentioned Legora for their legal stack. And I think that is to me like the most interesting thing when you start to blur the edge of what is software and what is service. Because really interestingly in the legal business software is like a $20 billion market and legal services is like a trillion dollar market.
Yeah, it's interesting. I feel like so many people thought this like hourly billing problem would make it really hard to adopt software. But I think you're totally right that if it's something that your clients are using, you need to be on top of it as well.
Yeah. And maybe clarify that point. The hourly billing problem is a sort of an incentive basically. That basically, if the pricing model is I bill for every hour I do, if you make me 50% more efficient, I'm billing for 50% less hours. But it's like, I don't know. I feel like lawyers used to go manually look up things in libraries and now they have databases. And there's definitely been examples of things that have been adopted.
And one thing that maybe it's not as clear from outside of the industry is that there's a ton of write-offs and price pressure. A due diligence used to be really expensive and now it's almost starting to get to the point where clients aren't even willing to pay for it. If you're a large PE client of a law firm, you will for sure pay for the advisory to the board when you're buying a company, but you might not pay for the contract review. And already you see the large American firms outsourcing this work because it isn't profitable for them to staff an associate who's billing 800 bucks an hour on a task like that. And it's a bit of a prisoner's dilemma in that if any of your competitors do something, you have to adjust to keep up with the pace of play. So even if you theoretically have the mindset that you're talking about, if someone else does it, then you're certainly suddenly inefficient or you're billing for things that other people won't.
That's exactly it. You have an almost perfect equilibrium of the service and as soon as somebody moves down it forces everybody to do it because there's very low differentiation on particular tasks like that.
I think we want to dive into a little bit of the AI within Legora but I guess one question just at a business level is there's other players unfortunately in the space that you're operating in. I'm curious what advantages you feel like you've had both coming from the Nordics and also maybe starting six months 12 months after some of the other players in the space. What does that advantage you as a company?
So, I think starting in the Nordics, it's almost like we were this really really small fish and we got to eat our way to become a bit of a bigger fish and then we suddenly became a crocodile in this smaller pond. And now, we're jumping across the Atlantic. We're coming to the States. We just opened in New York. And frankly, the great thing was Europe is such a fragmented market in every country effectively. So our initial market was very small and that was one of the first things that we got pushed on during Y Combinator that's like how are you going to move from Sweden and I said well we're not we're going to do Sweden we're going to win it and then we'll move. And funny enough I think one of the main advantages being the sort of fast second mover if you will is you can observe what's working not working. I think initially there was a lot of focus on we need to train our own LLMs we need to be an AI company first that's like pushing the research in this field. And coming second with also significantly less funding. I mean, our initial angel round was $50,000. And we just said, "Hey, we don't have the money to do that. We don't have the time. We don't have the energy. Let's just build an application and let's focus on the application layer and build something that at the end of the day people are really excited to use." And coming from sort of a non-legal background, it also forces you to be very humble and be very attentive to what everybody's saying. So what are the clients saying? How is the relationship between a law firm and their client developing? And I think this is also now pushing us ahead where we can move from being just an internal facing tool into really focusing on the entire relationship that exists in the delivery of high quality legal work.
I did a podcast one time with Daniel Left and he said starting in the Nordics allowed him to had he started in the US they would have made him skinny down his product a little bit into a more narrow functionality but because he started in this other pond he could be a little bit more ambitious unknowingly because he serviced the totality of the needs of the market. I'm curious if that's something you felt.
I don't think I've thought about it that way but I think if you looked back on it it kind of explains what we did, right? There's a lot of legal tech companies that are solving a very niche problem, right? Drafting, contract review, or they solve, it might be super super narrow. And even if you look at a big law firm, you've got litigation, you've got corporate, you've got the transactional team and they work very differently. And we sort of said, hey, we want to service every lawyer. I want every lawyer who's serious about doing great work and making money to use Legora just as every great designer uses Figma. And perhaps you wouldn't be that ambitious had you started here because of the sort of competitive pressure, right? I think there's also one nice thing about being able to serve enterprise in your local market to the point where you're already enterprise ready when you enter a new market. And law firms are very well connected. If you're a large firm in Sweden, you have an almost partner firm or a good friend firm in France and in Germany and in Spain and in the US. So, by starting to work with a couple of firms, you also get really great referrals because they all want to lean in and kind of do it the same way.
I mean, I think one thing that's really powerful about your product, as you just alluded to, is just how broad the capabilities are and you really are able to serve kind of end to end at these law firms. I know a lot of builders out there are kind of grappling with some problems, or thinking through what happens when I have this end-to-end product. How do I like teach people to use it or how to get started with it? And I'm wondering like what you've learned now with all these deployments you've done about like how to teach lawyers to use these tools.
Yeah. Well, I mean, I think the short learning is it takes way more work than we thought it would. One great example is Jam who comes from Baker McKenzie. He was responsible for the GenAI rollout and it was funny talking to him because in the initial traditional software rollouts he did you would be thrilled to get five to 10% adoption like that would be a great number but for any other enterprise software that is awful like those metrics suck. And now with Legora when he's doing other client deployments for us we're increasingly hitting numbers like 70 80% adoption and it's a different world when the lawyers are actively approaching the innovation team saying, "Hey, we want access. We want these tools." And that hasn't really happened in this industry before.
How do you think about where to build towards with the models constantly improving? You don't want to you want to skate to where the puck is going on some ways, but you also have to be pragmatic about what's actually applicable today. And so how do you think about that balance and at what point you sort of think about product roadmap meeting model experience?
You can think about the models, but I also think it's useful to think about these AI labs as platforms and software companies because frankly OpenAI, Claude, Gemini, they're for sure model providers, but they're also increasingly product companies. Anthropic is building out Claude to connect with other systems. Google is building out Gemini to sort of sit on top of the entire Google workspace stack. And they're not only pushing innovation on the models themselves, but also on the way that they interact with other pieces of the software and system. So the way that I've always thought about it is if it's something that the AI labs are going to build and at some point make available to builders like us, then we should not build it. And that's sort of going totally against the stream. And you want to build everything like boats so that when the tide rises, all the functionality just gets better. But a great example would be we just released a new feature called playbooks where you kind of give Legora like a set of rules and typical fallbacks and points of how it's supposed to negotiate and let's say you outline 20 different rules for how to negotiate an NDA or a MSA. Now Legora needs to go through them one by one by one to make really high quality edits. But if the models become so good that you just say hey here's a playbook in a word document or in an Excel file take this other NDA and cross reference them and give me all the red lines then you don't really need to build the feature that way. So that's something where we think it adds a lot of value building it that way today but 5 years from now maybe that's completely unnecessary.
Yeah, it's such an interesting tension where it feels like, you want to provide value to your customers today with whatever the scaffolding you need around the models while also realizing that some of that scaffolding may just be completely irrelevant in 2 3 years.
Exactly. And I think another great example would be workflows like multi-step instructions. Typically you would have these sort of node-based node builders where you have the output of one block serving as the input for the second block and you need these kind of technical builders to come in and set up these workflows. But the LLMs themselves with proper planning and then execution with tool use are fully capable of on the fly creating their own workflow or like their own plan of it. And so the only thing you need to do is really provide it with an instruction and say hey I want you to do this task and then it's going to be able to figure it out themselves given the tools that are at its disposal.
I realize there's no perfect way of predicting what the model capabilities are going to be even tomorrow, let alone 6 12 months from now. If you can let us know that would make our job a lot easier if you figure that out. But I guess at a practical level that point that you made about like what exactly can be done potentially in the future versus how you need to do it today. Do you have some framework that like, hey, if it really feels like this could be done by an LLM in the next three months or six months, we might let's hold off. Let's table it versus it feels way in the future, let's just go build it. Like, how do you think about that?
You just saw Claude and Anthropic release citations as part of their API. Now, we built citations really, really early on because it's like critical for our use case and for lawyers to be able to reference and see, okay, the LLM made this response and it references to this material and this specific text chunk. But, if the LLM providers give that for free in half a year, then we'll just deprecate our entire code. I mean, my way of thinking about is if somebody else is doing it, then let's use that.
What's like the hardest part of building these products today?
I think the hardest part is kind of to what Logan was alluding to, there's a hundred different things that you kind of want to be doing and they're all really high value and then you need to prioritize like which five are we going to go after and then how do these five tie into each other in a very cohesive platform because it's easy to build a Frankenstein monster if you just always go after like the next hot thing. And I think some other companies in our space has made that mistake of just building stuff that kind of makes sense, but the totality and the sum of it isn't really well thought through. So it's hard to plan for how your platform is going to look like and be structured just from even like a data database perspective where okay we have an organization and then we have projects and these projects can maybe be shared with a client like how is all of this going to look like and how do we build the functionality so that that fits the paradigm for what we're going for. And here as well I think the law firms themselves and the in-house teams are kind of figuring out what they want as well. So, we're all trying to figure out exactly the best way to apply the technology and then we need to align on a future that makes everybody win.
Some people put the cost structure in some ways on the client that you map, hey, this is what it costs us and therefore this is how we're going to bill you. That's not how you guys bill today. I'm curious how you think about the pricing and the value and the cost and how those three things kind of all interact and therefore what that means as model costs continue to decrease like how do you pass that on to the customer?
Yeah, this is a great question. And I was talking with Verun from Clay about it in a cab the other day. And we have the problem of running a seat-based model, right? Which I think is easy to buy. It's easy to predict. You know what you get. But then just the other week we had a user rack up 10,000 bucks in LLM costs. So I think over time it's not hard for me to imagine where you'd have kind of a platform fee and some usage element to that. But it also depends on how the LLM providers themselves develop their pricing. I think the early thoughts around all of this was LLM prices will continue to go down. And so if LLM prices continue to go down, we're willing to take a bit of a hit in the beginning to get positive and better margins over time. Now that's not really happening because the LLM models are becoming so much better, but also more expensive. I don't know if you've tried o3, but o3 is incredibly good at a lot of legal work. It's incredibly incredibly good, but it's also very very expensive. And so I do think that there's kind of an element to this where it's like do you need the bazooka or do you just need a handgun for like the task that we're doing? And increasingly so we're running classification algorithms and model pickers that kind of pick the best model for the task.
I think it's such a fascinating point because it's probably, next year that same kind of recreation. It might actually be quite cheap, but it'll also be so last year and there'll be some other model capability that's even more powerful that folks will end up wanting.
What about on like the infrastructure side? You're obviously one of like the biggest users of LLM out there today. Like how do you think about some of the gaps in infrastructure tooling that exist? Some of the stuff you've used that you found helpful.
The most interesting thing to me is MCP where we can give Legora access to call on outside tools. We can provide a set of tools that are already good and let's say we wanted we give Legora access to red line and document, right? And then that's a tool that the LLM can use. But the really interesting part becomes when our clients can provide tools as well. So you'll have Legora access a client-specific CRM or a client-specific knowledge database or a set of templates or it's able to push notifications to the client's emails or something like this right and that sort of very very quickly expands the universe for what's possible.
Since we're a venture sponsored podcast I have to ask you every VC's favorite question which is how do you think about what the moats are in AI applications?
The moats of course look very differently across industries. Take Figma. I think the moat in Figma is that it's kind of the system of record and it's the platform where both the designers, PMs, marketers are collaborating but also where they're directly collaborating with their clients. And I think for us there's a lot of interesting directions to explore there. And we currently integrate on top of a lot of our customers own data. We integrate with outside databases for things like case law legislation and coming
Regulation and a lot of it actually sits in the system that you'd like to work in. I think right, it's sort of a taste-based, I like working with this thing. I think if you grab the average lawyer in big law, they spend 80% of their time in Microsoft Word, Outlook, and iManage, their document management system. And I think that's about to change.
