Lawyers and the AI revolution

0:02Hello, I am Danny Fortson and this is the Times Tech Podcast. Once again, I am writing solo, but fear not, Katie Prescott, my fabulous co-host who's been out for a few weeks now. She's going to be back next week, I promise. But in the meantime, we're continuing. We're soldiering on. For this week, we're talking AI. In the past few years, there's been ominous predictions of a, quote, white-collar recession. As companies start replacing more and more of their human staff with AI, Dario Amadei, the CEO of Anthropic, friend of the pod, he recently predicted 50%, 5-0 of entry-level white-colored jobs could be gone within five years.

0:47And you're certainly seeing bits and pieces of that out here where you see the big tech companies growing very fast, but also cutting staff. and there was a really interesting note that Satya Nadella from Microsoft just put out talking about the irony of being a fast-growing tech company while also he's cut 15 ,000 jobs this year and people were really kind of freaked out by it because he's like look we're doing better than ever record profits record share prices things going like gangbusters and people need to do more with AI basically he can do more with less it was a real just kind of moment for you One of the largest companies on the planet to be like, this is the irony of running a successful company.

1:31What the arrival of this new technology is that we can get rid of a bunch of people and we can still do quite well. So, you know, pretty terrifying for anybody in a desk job. And it's clear that some white collar jobs will be hit harder than others. But what's on the chopping block, you might ask? What is kind of like the front lines? One possible area. The law. So much of being a lawyer revolves around combing through huge mountains of documents, making sense of them, and this is something that AI tools are particularly good at. Goldman Sachs predicted that 44 % of legal services could be automated in the future.

2:12So we'll see change coming for that profession. And at the heart of this change are companies like Luminance, which is another UK-based startup like our friends from Limbic last week. And they complete tasks that once took lawyers hours. They can do it in seconds. And they're really centered on contract review and really focusing on lawyers inside big companies who can spend days and days and days and days just going through contracts. They basically become an expert on a company's data and then use this to mark up and analyze incoming contracts, flag risks, make sure everything aligns with the company's goals, etc.

2:48You can think of them as kind of a synthetic in-house legal brain on steroids. So it's a really interesting idea. And Luminance is run by their CEO, Eleanor Lightbody. Luminance just raised$75 million at a$400 million valuation. They're growing like a weed. And so I want to explore this kind of brave-ish new world of legal AI and the shock that is actually reverberating through the law industry as we speak. How long before we don't have lawyers anymore? How long until we don't have lawyers? I think, sadly, for your listeners, we are quite far out from that one. But how many lawyers is a different question.

3:33Machine learning has been around for a long time, and AI actually has been around for a very long time. And I think that lawyers were, or I know that lawyers were using machine learning tools since they were on the market. But the rate of adoption has massively accelerated with the instruction of these generative models. And why is that? I think because previously, Luminance and whoever was building technologies for law firms or for lawyers were slightly capped with what they could give their end customers. And so I think that what they were doing is that they were producing results that were 2x beneficial to the customer.

4:14But I think to change behaviours and to really get people using technology on a daily basis, you almost have to see a 10x in productivity. And that's what the introduction of LLens into these systems created. is that you almost like cross the chasm and suddenly it was like, okay, this is no longer nice to have and I'm only going to use it on certain situations where like it's not feasible to hire hundreds of lawyers to I want to use this every single day and I can see more and more value the more that I use it. Yeah, it has to feel almost like magic for people to be like, yeah, okay, I'm going to use this now.

4:48Yeah, we always talk about the hook. What's going to drive someone to keep on coming back? What is it that it did today that it didn't do yesterday? What value is it providing you that you might not necessarily have thought about? And actually, the legal profession is a really interesting one because a lot of the work, as you said, it's text heavy. There's a huge amount of information that you can start learning from. But actually, a lot of the work that we focus on in-house legal teams, a lot of the work in-house legal teams are faced with on a daily basis is pretty repetitive. And they were coming to us and saying, oh, we've got mountains of work that we're having to hire many paralegals to do.

5:26and many kind of, you know, first year, second years to help out with, but they're not getting to the work that we want them to get to. And so I think that's what's really drive the usage. Originally, so two years ago, it was like, I want to focus on the low value, high value work. Now, obviously that's changed and it's being used for a much wider array of use cases. But that was the start. Well, speaking of the start, can we go back to the beginning? Because you guys have been doing this AI thing before it was cool. It was trendy before. yes so what year did the company start and what was the idea then we're celebrating our 10th year anniversary 10 years 10 years danny and i haven't as a tech startup that's like you guys are like a dinosaur that's crazy and we're still here which is and we're growing at you know 5x ari over the last year so we're boss of the number four which is really exciting 10 years ago it was founded by two mathematicians from the university of cambridge who were to speak very openly i think quite fed up of conversations that they were having with friends of theirs who were lawyers who were complaining about the hours that they were putting into huge pieces of due diligence work and so they looked at each other and they thought technology is going to help humans get through data faster like that's inevitable and they created the first situations of luminance and took it to a law firm here in the UK called SlaughterMA one of the most reputable law firms and Slaughter's were so impressed with the first platform that it not only became our first customer and gave us access to their data so we could start training but also became one of our first early investors and the first five years the team really focused on selling to law firms for a few reasons the first one was is that they knew that data was absolutely crucial in this world and this is before most of the world knew you know today we talk about how valuable data is but three years ago no one really knew how valuable or no one really thought to that so we're getting access to this very very rich data set which for legal we can talk about this later but it's super super important because you want to build a specialized system that understands whether a question is right or wrong in the legal context the second thing is is that because lawyers were using it it was getting trained on the fly.

7:40So every single time a lawyer was kind of logging in and verifying things or teaching it, our systems got smarter and smarter. Then of course, you know, the instructions of transformer-based models, generative models, we were able to adapt very, very quickly because we've always been in the AI space. And about three and a half years ago, we were seeing that many more in-house legal teams were coming to us and saying, what you're building is really useful actually for us like this is not just an m &a platform but we want to use you across all of our contracts whether that's creating contracts whether that's negotiating contracts or whether that's finding key information within your contracts and so we expanded out to work with in-house legal teams and that's the product that's just flown off the shelves i would say metaphorically because obviously it's a sass platform there's no shelf Exactly.

8:29There's no shell. Flown off the website. And we've welcomed the likes of DHL, Allianz, Coke Industries, MD and many more. So that for us is really what we're focused on. And that's building these legal brains within organizations to understand the organizations, understand the legal departments, understand the legal user. I can help automate and augment that every single interaction that a team might have with a contract on a daily basis. But going back to that 10 years ago, were you guys like a voice in the wilderness? 10 years ago, there was a lot of work on educating the end user. And there was a lot of skepticism, much more so than there is today.

9:11And so the teams really had to focus on what was the value and what was the future value of this as well. Because the end users, today you go in and you can ask any questions. And even if it hasn't seen anything, it will give you an answer. whereas 10 years ago if the systems hadn't seen something you'd have to train it on the fly so it's a lot of like okay how do we get this system to be all bespoke and smarter to you as an organization how did the arrival of chat gpt change things because obviously it broadly speaking woke the world up to ai everybody's like oh my goodness but luminance by that point had been around for seven-ish years, did it allow you guys to do new things with the product or show you the way?

9:54And how much of it was just the awareness of, oh, we are now in the AI age, everything is going to be AI. So customers were all of a sudden like, we need to figure out what we're doing here. I think it's a combination of things. The first thing is that ChatGPT was the thing that woke everyone up to it, but we had been using transformer-based models for a year or 18 months previous to that so our customers had been interacting with generative models through our platform for a while before that chat gpt helped spark everyone's imagination and in one sense that was incredible because the things that you could start doing with and it's not just chat gpt you know whether it's gemini whether there's so many different models that you can use for different tasks and actually some are very good at certain tasks but some are better at other tasks and it's actually how you orchestrate them that gets you the best outcome for your customer especially in legal when like accuracy is the most important thing so it worked the everyone's imagination to it what we could do with our platform obviously the value increased more and more to our end users and what's incredible is like even just this year the questions that have been asked on the platform have increased by 40x i couldn't quite believe that number i was like 40x in the space of six months like wait what has increased by 40x the amount of usage within our platform over the last six months.

11:12Whoa. Yeah. Well, I think I was trying to figure out why that is. And I was like, maybe it's like to do with the tariffs and people wanting to really understand quick. And actually, that was only like a subsection of the questions that were being asked. But I think it's very representative of the growth that we've been on for the last few years. On the other hand, though, which is quite an interesting point, which is suddenly everyone thought that they were maybe AI experts. And as you said, it was everywhere in the news that legal was going to be disrupted. So you had a lot of a lot of people playing around with models who were building quite specific solutions for very specific use cases and so there was again another educational piece that we had to do to say okay like there's a difference between like a generalist system and it's probably quite useful for certain tasks like giving you a first draft where you don't need to be legally accurate or giving you some advice when you don't need to necessarily trust the citations or trust the case law that's being used but actually most people now especially from what we're saying from our customers they want something that's going to capture that end to end they don't want to have like one legal platform for one thing and one legal platform for another thing they want to have something that's super robust that helps for any different use case that they might be faced with with the contract And are you building on top of other LLMs like ChatGPT, like Claude, like Gemini, like Llama, etc.?

12:37Because there's this universe of these kind of general purpose AIs. They call them rappers out here, where you just slap a new rapper on it, and all of a sudden you have a legal AI, and you have a medical AI, and blah, blah, blah. And there's a whole universe of companies doing that. Are you guys in that world as well? We're not a rapper at all, but we use lots of the models that you just spoke underneath the hood. So for us, we use some open source models. We use foundational models. We use some models that we actually built ourselves. And we've got access to some of the most rich legal data out there.

13:16So we can train all of these models. But that's only one piece of the puzzle because large language models are non-deterministic by nature. and so what you'll find is sometimes you will ask the same question four times you might get six different answers and it's actually the way that you orchestrate the models together to check each other's homework to make sure that like there's probabilistic consensus in the answers that are being given to the lawyer because you know in an industry where creativity is really important like it might not matter that you might you know not have consensus in the answers that you're getting back that actually might be a good thing but a lawyer as soon as they don't trust the answer that they're being given, you'll see that rate of adoption.

13:56Yeah, massively declined. The reason I ask is because there's a recent case, which you may have seen, of this company, WindSurf, out here, which is one of these AI coding assistants. And opening, I was like, we're going to buy this company for$3 billion, but WindSurf was built on top of Claude. And then Anthropik's like, you know what? We're going to cut off access to our API. And all of a sudden, it's like, hmm, well, what is this company now? And then Google swoops in and hires a CEO and the top team. And then the company was like, what are we now? And then it got acquired by another company.

14:32It was like this crazy succession of events. But I think it kind of opened a lot of people's eyes to like, well, what do you have in this world where Claude, ChatGPT, LGBT, these companies are throwing billions of dollars to create these really powerful models that you can either build on top of or, you know, next week, Anthropic might be, you know what, here's our new legal AI. It's the most powerful, blah, blah, blah, blah, blah, blah, blah, blah. I'm wondering how you navigate this world where literally every week, it's like, oh, this new model is more powerful than anything previously seen as of a week ago, you know and it's just constantly evolving like that well I think that's really really key I think it's not being too reliant on one model as I said we have like all these different models that we use because actually one model is very good at text summarization but another model will be actually very good at calculating different costs within a contract another model is actually much better at understanding the jurisdictions and so we have like this leaderboard in Cambridge where we have all the different models and what we're like testing them for and as soon as there's a model update that comes out we will check it against everything else and the key thing here Danny is that you've got to not just test for what your product's doing today you've got to test for like a lot of the conversations I have with our team it's like cool of course these model updates are going to hopefully you know improve whatever we're looking at today but like what can they do that we can't do like what are we building thinking about what's to come and that's like a really I I think that's a hard thing for businesses, but like the really exciting thing, which is, you know, it's almost like chips, like you always wanted to build much smaller chips, because you know that they're coming.

16:14It's like, with the models, you know, the models are becoming more and more powerful. So you need to make sure that you're building use cases that as soon as they're there, you can be at the forefront of delivering something that like people are like, wow, I didn't even know that this could even exist. And how do you navigate this idea that models lie, they manipulate, they make stuff up? I mentioned it on an earlier pod this woman who asked chad gbt to help her choose she's a writer she's like i'm gonna make a proposal to a literary agent can you read my stuff it's like yeah i'll read all your articles blah blah and then it turns out like chad gbt said it read it it actually didn't read anything it made up the content of which it read that it then analyzed it was such an elaborate web of lies and then when it got caught it like doubled down and said no no i read every word and then it was like you know what actually I didn't I lied it was such a kind of colorful example of these things are very good at being confidently wrong and how do you deal with that as to your point in the legal world like you can't afford that the first time a lawyer is like oh this thing is making things up or this thing is just wrong or lying you know it's down tools yeah absolutely so I think the two elements of that the first one is getting the different models to check through its own homework and so like if you've got eight different models all saying the answer is the same then you can pretty much confidently think that the answer is the same but if there isn't probabilistic consensus we actually have like on top of all of our models an orchestration layer and it's almost like the supreme judge who basically looks through the answers and goes you know what I'm actually not sure that you guys are like correct on this and so i'm going to hand it back to the human there's a huge amount of work that we've done so that like if the air is not sure in the platform it will color code something yellow and it will say this or this answer like so there's like an ai judge judging the other exactly and that ai judge is not an llm got you the other thing is you mentioned earlier it's like well this thing has to be like 10 10x better than anything that's come before and i don't know if he saw there was a recent study by this company, Meter, of AI coding assistants, because coding is another area where there's been a huge amount of investment and hype and excitement.

18:34And it found that it did a study of human coders using AI coding assistants, and it found that they're actually 20 % slower than people who are not using assistants. And probably because it has something to do with them, like double-checking or triple-checking, or maybe they're just still trying to figure out how best to work with it or whatever. But it wasn't like this kind of night-to-day revolution of like, oh my goodness, now I can just do twice or 10 times as much work or whatever it may be. Does that ring true at all? Do you see any aspects of that as people start adopting these things? Look, I think first of all, there's a lot of hype.

19:12So cutting through the noise and trying to understand what's happening. We always talk about return on investment. What's the business case? why would someone adopt this rather than it just be kind of an egotistical thing or something that you're like this is great but like I actually can't can't quite communicate what the value is when it comes to things like coding like we obviously use AI to help our developers but we've seen that with our experienced developers it will have a huge amount of impact but with our junior developers that's where it just causes a lot more time to get code out there and do triple checking code review and so i haven't read this report and i'd be interested to read it but for us there's definitely a there's a difference between who's using it and how they're using it and then again to check tests and to like write pretty like mundane code super efficient but anything that you want to be like super like slightly more creative or anything that you're introducing to maybe like a web of complexity then yeah someone that knows that code base better is still more efficient.

20:16Are there any parallels with, you know, junior lawyers versus very senior lawyers using your software? Weirdly, I see the inverse. Because with junior lawyers, a lot of us, you know, regardless of whether you're a lawyer or not, you learn through repetition. And you learn through insights. And so the great thing about junior lawyers using Luminance is that the AI will tell you why another user has done something in the past and how they've done something in the past. So actually it can upskill them at much faster rates. And also it then allows for businesses to have assurance that like institutional knowledge stays within an organization.

20:55What do you mean? I mean, like AI is not going to leave you. Let's hope it's not going to leave you unless you stop paying for it. And then we'll cancel your subscription. We'll take all the learnings away from you. But it's there and it gets more and more powerful. And it learns why you as a business have done things and why the user have done things. And so should someone leave, someone can pick up where they left off very quickly and get that in-depth understanding. And it also doesn't sleep. It has that longitude and latitude context and memory. So I always think that in legal, it's even more powerful to a certain extent.

21:30Is it kind of like, I think one of the best businesses on the planet is like a therapist. because if you start therapy you spend the first six months sharing all of your family history and your deepest darkest secrets and then this person knows you and then they're like then they can advise you their institutional knowledge of you is set and there's a whole lot of work involved with that and you're like i don't ever want to have to find a new therapist again is that kind of the luminance model you're like the ai legal therapist of like once we know your organization you're never going to want to leave us because we're just going to get smarter and smarter about you well i would hope that um unlike a therapist where it might be a love-hate relationship that you might have a love-love relationship with the luminance and that you're you wouldn't entertain thinking about what was next and therapy always has an end point well you would hope that therapy has an end point whereas with luminance yeah many many years so yes to a certain extent but it's beyond that because it doesn't need six months to learn who you are like all you have to do is upload your existing contracts and upload your playbooks and your precedents and then it gets a very good understanding from day one on what you've done why you've done it now of course if you as the end user are seeing new complex contracts that the ai has never seen before every time you interact with it the more that it learns from you and obviously it has that reinforcement learning into it but unlike kind of legacy softwares where you you always have to wait six months to get started this people can try like totally for free and tomorrow they can start seeing value from it.

23:01I want to talk about humans and staffing because this is the other big question, right? Everybody's freaking out. Like there's never going to be a human smarter than a machine ever again. On some like, you know, I guess IQ level, that's probably true. But I think it's a little facile to be like, well, therefore, if that's true, then we're all going to be put out of a job sooner or later if it has to do with anything to do with like kind of white collar or kind of thinking quote unquote jobs. How do you see this playing out? Or what are you seeing already with your clients that are using this, using it more?

23:36As you say, the 40X increase in the last six months, which is, that's an astonishing figure. Like, how do you start seeing that trickle through these organizations? I obviously think about this quite a lot. And I still come back to this idea that humans, we love to make work for ourselves. We have traditionally time and time again. And so it's not necessarily clear what that work might be. But I do, I back ask humans in finding a way to not sit on a beach 24-7 if we wanted to. I suppose what are we seeing in practice? Well, it depends on the company. But, you know, you've got lots of companies who are saying, I just haven't had enough time to do the job that I was hired to do in the first place.

24:20It's like, you know, if you think about a business, it's people, contracts and processes. And a lot of the contracts have rich data and like rich insights into how that business is doing, how it's operating, where it should go next. And actually lawyers and people adjacent to those have this very unique opportunity to be business enablers, to be strategic advisors. But if you speak to most people, they're like, I'm just spending time churning quite similar contracts because that's what I really need to do. so it's opening up this whole new area of work that is actually probably more interesting and why some of our customers got into law in the first place second to that I think we start to see roles blend and change so yeah you might see less paralegals but you also might see more legal data analysts you might see many more people be introduced to the environment to think about okay I've got a legal AI platform here and I also have a CRM here how do they talk to each other like and how do I make sure that they are you know and I've got agents now I've got my legal agent and I've got my API agent and I've got my my search agent how do those all fit in and how do I make sure that we're delivering value for the business so I think whilst there might be reductions in some areas I would see and I have started to see kind of either roles change and morph into other roles or totally new roles be introduced any totally new roles come to mind vibe coder i'm serious like we've just hired our first ever vibe coder and i never thought we hired a vibe coder we have hired a vibe coder and like you advertise luminance looking for a vibe coder or is it not that cool i i can't remember exactly what the what exactly the spec was but i do remember speaking to our product team and i said look let's just try something unique and we found this person who didn't have an engineering background, but he had this incredible roster of things that he built using LLMs.

26:24And we were like, this is really cool. Let's just see, you know, it's a small experiment in the grand scheme of things. Let's hire him and let's totally separate him from the engineering teams because we don't want to get them too distracted. Actually, some of the stuff that he's built is now going to make its way into product. And so - And this is him basically telling the AIs what to build in natural language effectively. Exactly. Exactly. Right. But without a computer science or kind of a coding background or anything like that. And so I was like, wow, this is a new role. This is a new, now let's see how easy it's going to be to get his ideas into the product.

26:59That's the second thing. That's a separate thing. But yeah, it's a whole new generation coming into the workforce. That's fascinating. The other thing is hype. Like I'm here in Silicon Valley. I'm old. So I was here for the dot-com boom and bust. That was kind of how I started my career. And so I've seen, in a way, I've seen this movie before. This new thing comes along. It's going to change everything. There's just like a deluge of money and hype. And then fast forward two years, nine out of 10 companies are gone. Most of them went bankrupt spectacularly, et cetera. You have companies, rivals like, say, I don't know, Harvey, that are worth$3 billion.

27:45I think your last valuation was$400 million. Is that right? Yeah. How do you see this playing out? Yeah. I think there are going to be some massive winners and there are going to be some massive losers, like anything. So I think that there are some companies who are literally providing a huge amount of value to their customers and who are being tested and rigorously tested before people are buying them. And so that in itself shows that there must be some tangible benefits to them. But I also think there's a lot of recent examples of companies who have been super overvalued and didn't have much AI to their products at all.

28:21And there was a lot of people around it rather than AI. When I think about the market we're operating in, valuations don't necessarily signify the potential future of the business. We're very focused on the growth and the fact that our AR over the last year screw 5X and the fact that, you know, as I said, usage is just like totally through the roof. But that's what we really focus on, like driving what we know to be the most important things, which are product innovation and customer acquisition. So going forward, five years from now, what does Luminance look like? And what does the, like the legal industry look like?

29:00Is it fundamentally different? Yeah, I mean, five years forward is a very long time scale to think about, because I think things are moving at such rapid rates. The one thing I think about is, at the moment, every company in the world will receive an email, they'll receive a contract, they'll click into the contract. Hopefully they're using Numits and they'll press a button and the AI will rewrite the contract to be more aligned with their positions. But that's going to change. The kind of softwares that we as humans are so accustomed to using, I personally think is going to, that platform is going to be radicalized with agents.

29:35And so you're going to have agents talking to each other during most of the negotiations on your behalf. And somehow as an end user in the human, the way that you're interpreting and analyzing the outputs might not be in a contract form, might not be through email, might be in a totally different format. I don't know what that format is before you ask, because if I did, we would be very much pushing it today. And for that, I think that will take time because, again, lawyers are creatures of habit. So I don't think that's going to happen tomorrow. but I do think that that's on the horizon and for Luminance I like to think that we've done a very good job with legal you know we've grown very very fast in the space and what a lot of our customers are now coming to us to ask for is there are so many use cases that sit adjacent to legal compliance procurement HR and we're building these modules to help empower those teams and to start helping essentially automate the work that legal might do that sits alongside compliance alongside risk alongside finance whatever it might be creating like an outsourced legal workforce or insourced insourced exactly ai kind of brains that are specialized for different departments so that was eleanor light body of luminance and i guess the thing that it left me thinking about is my job, a lot of my friends' jobs, how much of their work could be done, could be easily automated, because we all think we're special snowflakes and that no, no, no, no, I'm very creative or what I do cannot be replicated.

31:08But I think what we're finding is a lot of these things can be. And that is at once, as I said, totally terrifying because the obvious follow-on, and I think a lot of what companies are doing is like, yay, we can just get rid of humans. But as we saw with Klarna, they got rid of a bunch of people, but then they brought a bunch of people back because people actually wanted to talk to humans in certain contexts. So I think it's really a question of like, if you can extract 20, 30, 50 % efficiency today from what you were doing two, three, five years ago, the question is, what do you do with that extra time, that extra money, that extra resource?

31:50Does that mean you'd simply need less people or will these technologies because they free people up to do something over here then they can do new stuff over there and i do think and maybe this is kind of i'm fooling myself but i am optimistic that ai will overall be a job generator but i think we're in the phase of the cycle where you know the tide is going out it's taking jobs right now but I do think it'll create new ones. That's my hope anyway. Otherwise, you know, God help us all. Anyway, next week, Katie Prescott is back. So please do tune in and until then, take it easy and we'll talk very soon.

32:34Bye-bye.