As an industry, we've gone from a 'what tool do we roll out' conversation to a 'where does our harness live' conversation. And I think most organizations are starting to think about what harnesses should they own rather than rent.
Take the entire service market, it's roughly 5% software and 95% service, and that is about to change. It will make the supply of legal services much more widely available. It will make it possible to pursue strategies in litigation that we otherwise may not have thought of.
Welcome to Enterprise Aligned AI, the podcast where we interview enterprise practitioners on the reality of agentic AI adoption. What gets deployed in regulated, security-conscious enterprises? How is success measured, and how are the best leaders navigating the changes? My guests today are Danielle Benecke, who leads Baker McKenzie's Applied AI practice, a firm with roughly 4,600 lawyers, more than 70 offices worldwide, consistently ranked the world's number one legal brand. And Max Jönstrand, CEO and co-founder of Legora, the Stockholm-born legal AI platform founded in 2023, now valued at over $5.6 billion, and used by more than 1,200 legal teams worldwide. Truly a rocket ship. Thank you both for joining the podcast.
Thank you so much for having us, Aparna. It's great to be here.
Pleasure to be here. I'm glad we could make this happen. This all came out of Future Law at Stanford, so shout out to Roland Vargö and the team for bringing us together.
Totally. Legal is a hot vertical right now for AI, but both of you have been working on AI for many years before. Danielle, let's start with you. People picture a global law firm as being a buyer of technology, not necessarily building AI capabilities. And yet you've been leading the applied AI team at Baker McKenzie. You built it up from scratch. What led to that and what unique capabilities does your team bring?
So, we've been scaling our applied AI practice since 2020-2021, and it builds on a much bigger, longer AI program that dates back to the 2010s. But the real inflection point we saw back in 2020-2021 started with the GPT-3 API, and we started experimenting with that and bringing it into different service lines. And that for us was the moment where we realized we needed a forward-deployed capability. That's the capability that I founded, grew, continued to grow and lead now, and we are a group of lawyers and AI technologists that specialize in designing, calibrating, and steering AI systems that do complex legal work. The companies that we work with are large global multinationals that work across many jurisdictions and have really complex legal and regulatory landscapes. And so, what we do is we design systems that can undertake certain types of legal and compliance work to support those businesses.
Max, Legora also started in this field before we really had the foundation models. So, there's a lot of vision that went into the company. Tell us, as a co-founder of Legora, what made you start this company?
I don't think I get to take credit for those first gnarly years pre-GPT because I jumped on the ship in 2023, when things really started taking off. But with the early BERT models, and in particular there was a Swedish version called SweBERT that my co-founders were utilizing to summarize and work with court cases. It was really hard. These models were not very intelligent. They were very hard to work with, enormous context issues. And then suddenly we get GPT-3.5, right? And things just cascade from there. And so, we took the bet and we said, 'Hey, let's quit our jobs. Let's drop out. Let's pursue this full-time.' And about one month later, we started working with the biggest firm in the Nordics. A month after that, we got accepted into Y Combinator, and then we were off. And since then, not only has the underlying technology continued to develop, but the way that the scaffolding, the tool calling, the harnesses, the performance that we can squeeze out of these models, and the different user experiences that we can build around them, that has all flourished. It feels like we're running a marathon, but that we're still in mile one in many ways.
Was the original vision to summarize and automate legal work, but then in 2023, there was a breakout moment, and you realized that it could be more than that?
Well, back in 2020, it was more of a fun side project, I'd say. They were building their own text editor, they built their own semantic search layer, and all of this we just put on the shelf as soon as GPT-3.5 came. And if you think about where the vision for the company and the impact that we want to have in the world continues to develop, it's that the legal service market is so people-dependent, right? Take the entire service market, it's roughly 5% software and 95% service, and that is about to change. And I think the impact that that will have in society will be very positive. It will make the supply of legal services much more widely available. It will make it possible to pursue strategies in litigation that we otherwise may not have found or not have thought of. It will allow us to level up from gruesome doc review to the more interesting and maybe creative aspects of the work. And to Danielle's point, it actually allows us to solve some of the challenges that your clients and large organizations are facing, because at the end of the day, we're building this technology, we're applying the technology, and we're innovating because we want to go solve problems.
It's so interesting that you mentioned BERT, Max. One of the first AI-native services that we deployed to clients back in 2021 was we were taking the GPT-3 API and combining it with BERT and some other capabilities to do some scaled supply chain risk monitoring for clients, right? Obviously, a very primitive capability to what we're doing right now, but I think for us that was a really early signal about the commercial model that was possible. To your point, Max, this is fundamentally about new services that solve unmet needs. I think one of the challenges we have as an industry right now is most of the market is focused on accelerating existing services, existing ways that individual lawyers work at the task level. The real opportunity that certainly we see for our clients and that we see for the industry at large is around entirely net new categories of services, and that has been our focus from day one of our program.
Danielle, you've been at the forefront of both the legal challenges that AI creates as well as the types of legal problems that AI can solve. Most people think there's a potential for AI to do the rote work of the legal profession, but you're saying there's new types of problems and new types of work that maybe we haven't seen. Can you give us a few examples?
Yes, and somewhat counterintuitively, I think the true opportunity is around those more complex categories of work. A lot of the market sees the opportunity more around accelerating lower-level work. I think the real opportunity is around complex work that humans were simply not able to do without this toolbox that we have now. The other big thing that we're seeing in our practice is a lot of the work we're doing with clients stems from them rolling out AI across their businesses. We're in the very early innings of the shift. We had a number of clients come to us and say marketing and comms have been the first to roll out AI across the business. They're now creating massive mountains of legal work that need to be done, because in the old days they partnered with the advertising agencies and they would produce tens of pieces of content that our legal and compliance teams would review. They're now cutting deals with the frontier labs and they're producing tens of thousands of content that is highly customized down the long tail of countries that we operate across. So, one of the other early service lines that we got into is AI-enabled triage for global marketing and comms content for a lot of our global multinational clients. And we would be analyzing this content across product regulation, across IP risk, across ambush marketing, and all sorts of other legal and regulatory verticals. But we said to the clients very early on, 'This is not a scalable way to solve this problem. What we need to do is move upstream with you and start to design compliance systems that can help steer your marketing and comms agents and other agents operating across the business, so there is compliance possible in the first instance,' right? One of the big challenges we have right now is generating content is cheap, generating legal analysis is cheap, true judgment and accountability is scarce, right? And so, we don't want to get in the situation where we're in this AI arms race where we're just generating more and more content that the legal and compliance teams need to deal with downstream. We want to improve the quality of action across the company, and ultimately, I think the third-order effect of the third wave of AI in our industry is going to be all about runtime compliance steering for companies that have agents operating across their business. The first wave was obviously tools for lawyers. The second wave is now AI-native legal and compliance businesses. That's where a lot of the investment activity is now. I think the third wave is going to be about designing compliance systems that can help steer agents and business behavior in real time.
That's fascinating. Designing compliance agents that are steering output in real time sounds like it requires a collaboration between legal practitioners who understand regulation and all the nuances, and AI engineers. How do you make that collaboration happen?
We are seeing the early sparks of this right now in our industry. And to take that example earlier of the marketing comms compliance, where the original service was the downstream triage, the next iteration of that service is around designing a framework for the agents to operate within so that the content is as compliant as possible in the first place. And that starts to bring the walls down between legal, risk, compliance silos in the business as they historically have operated, into a much more collaborative environment. We're already seeing that happen across a lot of the global enterprises.
Does Legora, which is a technology provider for the legal space versus Baker McKenzie, which is a law firm, do you work very closely with your customers to start to invent the new aspects of the application?
I think Legora is one of the only technology companies where we have more lawyers than software engineers. It's very, very rare, but we do. And we've actually, I don't know if we invented the term first, but we've certainly adopted it. And it's the legal engineer. And the legal engineer is a very tech-savvy lawyer who both understands the problem that we're solving in terms of the use case, but also is adept enough to actually help drive change management and the transformation of how legal teams operate. Because if you just take this technology and you go, just as Danielle said, 'How can we improve the way that we already do things?', you'll get one type of result. But if you say, 'Well, now we have this technology, how should we reinvent the process to begin with?', then you can get something much more aligned with the capabilities that we now have. And I think in terms of new things coming, I love where this is going, because effectively all of the agents operating within an enterprise are going to have to, in real time, be able to query, 'Is this plan or is this thing that I'm now going to do compliant with our policies or compliant with regulation?' And the regulatory body that you actually need to take into account is enormous. It's too big for one single person to actually keep in their brain at all times. There's just too much context. And as we start to build these layers for agents that can operate more autonomously, then I think that's certainly a kind of cool path to keep going down.
There are all these new roles, forward-deployed engineers, which have been around for some time, and now legal engineers, who are lawyers and also engineers. But I want to ask you, Max, about differentiation and moat. This is a really important question, especially in the legal profession, because there are many different types of players. There's all these incumbent companies which have distribution, like LexisNexis and Thomson Reuters. They can obviously pull in AI capabilities into their existing services. And then there's a range of startups. Legal is a very hot area for funding startups. They're all going after different parts of the market, and certainly law firms are a big part of that. And then there's the question of can foundation models themselves start to increase what they offer by post-training, for example. For Legora, what are the true sources of moat?
Well, the space is developing at an exponential pace. And I like to say that velocity is a moat, because you have to be at the forefront and you have to be at the frontier at all times. If you look at what our product was back in 2023, it would have looked like a total joke today. If you brought out the product from 2024, it would have looked like a really poorly executed legal AI platform. And it may be even towards the middle of 2025. And with the things that we've just come out with, we are clearly moving ahead of what I'd say most of the pack is doing. And as you said, everybody's thinking about how can we win our slice of the cake here.
And we're still in mile one, right? Technology and product tends to compound. So, if you make a lot of really correct strategic decisions about where should we train models, where should we not train models, where should we build, where should we partner, where should we buy, how do we build a user experience at the end of the day, and a product that delivers more value than anything else can do on the market. And that's what Legora is obsessed about. There's, I'd say, different strategies at play. Yes, you have some of the big incumbents who are saying, 'Oh, we're going to leverage this piece of the data and our big distribution muscles, and that will be our way of winning.' And fine, that's one strategy. We see what's happening with those public stocks. Um, you know, they're maybe not doing super hot right now. And then on the other hand, you have private companies like Legora, who is in enormous demand in terms of our growth rate and where we're going. And I think that the old sort of SaaS world is also really getting disrupted by these AI-native businesses. If you look at the growth rate of some of these new players entering on the market, it's just throwing all of the old metrics and even the venture capital playbook out of the window. Because the net revenue retention, the DAU over MAU, growth in revenue, this is just unheard of. We get to experience that in legal, but the same thing is happening in coding, right? You've got Cursor on one end, which just exited to SpaceX, and you have all of these different verticals that are playing out, and we're certainly not at the finish line for the legal one.
Aside from velocity, what are the other moats?
Well, I mean, there's the degree of embeddedness with a customer. When you work with a firm like Baker McKenzie and you roll out to thousands of practitioners and they start to adopt Legora in terms of how they do work and deliver to their clients, that becomes one type of moat. When you build products like our portal that actually puts both the clients and the law firm in one single software experience, that becomes a way of driving both new types of delivery models, but which of course are quite sticky once you get them right. The way that you integrate with the rest of the ecosystem and that you build functionality that frankly is quite far away from what a foundation model would naturally go after. A lot of the software that we're building now has nothing to do with LLMs whatsoever. The amount of audit logs, permission management, trails, scheme, the amount of just things you need to solve in order to deploy within an organization like Baker or a big enterprise for that matter, in the legal department specifically, are quite extensive towards what just the horizontal models are doing. It's maybe not so clear at the surface, but the minute you step your foot in it, you realize that, oh, this pool was really shallow. If you want to upload a thousand documents from a VDR or a hundred thousand documents from a discovery exercise, and it just blatantly fails in some of these frontier models in the way that those are packaged in their very general harness. And you go to Legora and it's just like night and day. We're not positioned as we want to be the legal AI assistant for everyone. We're an agentic infrastructure layer that works with some of the biggest enterprise and legal firms in the world. I think we're naturally going after quite different segments of the market.
We've all worked with very large enterprises. There's a huge compliance burden. Would you say that Legora is a legal-specific harness?
We started to have a harness that's tuned to legal, but there's many layers in the product. We articulated this quite well, I think, with our agentic operating system that we released a few weeks back. It starts with the foundation models, and then we put those foundation models to work within our legal harness. And then we give that harness access to legal-specific tools, legal-specific context, legal-specific service areas, because not everything should be served in a chat, right? Every use case, the UX is not a chat, is not optimized for everything. And then all of that has to be wrapped in a secure and enterprise-grade system. And it has to be combined with the legal engineers. We should not forget the human aspect of this. Even if you have a tool like Claude, implementing that within an organization to drive the outcomes you want is quite extensive. There's a lot of DIYing you have to do if you take a general harness. Probably no organizations in the world can take that to the extent that we've taken Legora. And that's because we're 700 people who wake up every day and obsess about our particular problem. Whereas in these very general tools, there's maybe one PM and a few engineers.
But what is the harness? What is different about your harness than, let's say, Claude Code?
Claude Code is optimized for coding. It goes back to how do you let the model read the documents, make sense of documents, make sense of context, how does it tie together, what's authoritative between that case over there and some internal precedent that the firm might have, how do you actually weigh those two together? Which models do we use where, where do we allow it to launch sub-agents, where do we not allow it to launch sub-agents, how do we instruct the agent to actually plan? We recently introduced a new agent, and that is tailored to create a very detailed plan for the user before it sets off to do the work. Cursor does the same, right? When you're coding with it, it presents a plan before you actually allow it to go out and execute. But then the way that it plans is of course a function of the tools at its disposal and subject to its own capabilities. I think you can go very, very deep, and I could probably talk for an hour about where we want to take it.
Just to jump in on harnesses, it's important to distinguish between inner harnesses and outer harnesses. One of the things we see in the narrative in the market right now is harnesses becoming a monolithic concept, when in reality we're actually dealing with multiple harnesses if you're a big enterprise business. So we, for example, have direct access to a range of models through our cloud AI environments. We work with, within AI models, we work with Anthropic capabilities, we have access to Google capabilities, etc. We then have access to productivity layer harnesses. Anthropic's Cowork is a great example of that. You see similar capability filtering up through the Microsoft environment as well, and then you get into legal-specific tooling. Most large-scale law firms and other actors in the knowledge industry have access to multiple tools with different types of harnesses. As an organization, you also are starting to create your own harnesses. Every service line, every bespoke system that we design and deploy for our clients, that has its own unique harness. The most simple harness is probably a markdown, a folder of markdown files, right? A big piece of work that my team and I do is working with chief legal officers, chief compliance officers, general counsel, other leaders on the client side, and help them design their stack, their strategy around this technology. They're starting to think about the strategic dependency of having their harness live inside another platform. They're starting to think about what harnesses should they own rather than rent. One way that we think about it is there's a scale from generalizable to custom, there's a scale from extremely strategic to not so strategic. And I think for the stuff that is super generalized, not particularly strategic or important to your organization, totally fine to rely on a harness that lives outside of your environment. But for the very strategic stuff that ultimately turns on your expert steering or internal playbooks and internal know-how, you need to be very careful about where that harness lives. As an industry, we've gone from a 'what tool do we roll out' conversation to a 'where does our harness live' conversation. And I think most organizations are wise to that right now.
Does Baker McKenzie think about AI adoption as a top-down mandate or a bottom-up mandate? You're using many different tools, many different models. Is that by choice from lawyers within the firm, or has it been something that's top-down?
We're a services business, so everything starts with what our clients need. And in this era, it really comes down to three things. Our clients need strategic clarity, decision velocity, and risk mitigation. If you look at every piece of legal and compliance work, it's ultimately those three things. As an orchestrator of capability that serves that ultimate need, what do we need? That really drives every question about the capabilities that we plug in, whether that is our baseline cloud AI capability and multimodal access, whether that is bringing in best-in-class vendors, whether that is Legora or other players in the ecosystem. The market is moving so quickly. We need to be very agile, not bet the farm on any one thing, but just constantly thinking about what is ultimately going to serve the client need. Many players in the industry still think of this as primarily a technology shift. Technology is obviously a big part of it, but the commercial model transformation part is arguably much more profound than the technology shift. When I look to what we build and sell to our clients, the technology bit in many ways is the easy bit. The hard bit is how do you sustain that capability in a way that is commercially viable for the business and for the client, and that also meets the governance needs of sophisticated global companies. When my team is building out this set of primitives that ultimately drive all of our service lines, the commercial and the governance pieces are equally, if not more important than the technology pieces that we bring in.
I see. So, depending on client needs, you'll bring in the relevant technology, but the technology shift is actually leading to a commercial model shift. Tell us more about the commercial model shift.
So, historically, knowledge work has been about execution. Human workers executing on certain tasks. It's no secret we're five years into the shift at this moment. More and more knowledge work is about designing systems that execute rather than the downstream execution. Many lawyers and compliance professionals still are thinking of their role as shifting...
downstream. A lot of firms and other players initially thought, 'Well, we'll have the AI generate output and then we'll validate downstream.' The most important expression of expertise at this point in time is actually upstream in designing the system. The better the design, the better the output is going to be, the less downstream validation you're going to have to do.
But ultimately, changes what we're selling?
Completely. And that ultimately creates an entirely new commercial model for delivery of work. So, for example, my team, we operate off an entirely fixed fee model. Every engagement, we will design and price that around what the client needs. What kind of risk are they trying to mitigate? What kind of commercial opportunity are they trying to unlock? What is the ROI they're looking for? And then we design the service around that. Very different pricing model to the way that historically legal work has been delivered.
So, it's not about the billable hour, and the expertise, like you said, is applied in the design of the solution. And so, you scope out based on a problem what the solution is, and then it's a fixed fee, which is more outcome-based.
Exactly. And working within a large incumbent law firm, moving the ship takes time, right? My team operates in this way. That does not mean that every team across my firm is operating off of the same billing and pricing model, because it takes time for a wider industry to shift. But for the AI-native services that we sell, that is the way that we think the commercial model is best shaped.
