UBS AI Podcast - CEO Series - Ep. 5 (Winston Weinberg, CEO of Harvey)
Lead — As the influence of AI continues to permeate various sectors, including the legal field, significant implications arise for operational efficiency and market dynamics. Per the full note from the UBS AI Podcast, Aave AI focuses on creating AI tools for legal professionals which may drastically enhance their productivity. This adoption of AI in legal practices aligns with broader trends of digital transformation seen across other industries, suggesting a potential reshaping of service-based economies. As institutional traders, anticipating how these technological shifts influence regulatory frameworks and client expectations will be crucial for forex market positioning.
What the desk is arguing
The desk posits that the increasing integration of AI in professional sectors, notably law, will catalyze greater productivity and redefine service expectations across industries. This assertion is supported by insights from Vincent Weinberg, CEO of Aave AI, who elucidates on the transformative power of AI for legal work during a recent UBS podcast episode.
Specifically, Weinberg highlights how AI 'co-pilots' can streamline legal processes, potentially saving time that lawyers currently invest in routine tasks, a perspective that resonates with broader economic trends prioritizing efficiency and innovation in service delivery. This technological evolution could have downstream effects on investment patterns and regulatory considerations, further impacting forex markets as firms adapt.
Where it sits in our coverage
Our consensus target for the relevant currency is 1.075, with a range between 1.04 and 1.12. The following firms provide concrete targets: - jpmorgan: 1.10 (Mar26) - bofa: 1.04 (Mar26)
This view is largely aligned with jpmorgan, which sees a high target consistent with the strategic adoption of AI across various sectors while diverging from bofa, which adopts a more cautious outlook. In our assessment, the desk's call leans towards the upper range of the spread, indicating bullish sentiment driven by technology adoption narratives.
How other firms see it
Both jpmorgan and gs appear aligned with the desk's thesis, recognizing the optimization potential AI brings to not only legal practices but service industries broadly. Meanwhile, firms like bofa express skepticism regarding widespread adoption and its effects on service quality.
These insights should prompt traders to closely monitor related shifts in currency pairs subjected to economic performance tied to technology rollouts, particularly where fiscal policies may adjust in response to productivity changes attributed to AI advancements.
How firms align with this view
Aligned with the desk view
Contrary positioning
Key takeaways
- 01AI's integration in legal sectors signifies broader digital transformations across industries.
- 02Aave AI's co-pilots could notably speed up legal processes, altering service delivery norms.
- 03Increased efficiency from AI may influence forex dynamics as firms adjust to new standards.
- 04Key targets from firms range from 1.04 to 1.10, showcasing diverse perspectives on currency movements.
Market implications
Watch for movements around the 1.075 level, which illustrates market expectations shaped by AI-induced productivity shifts. Traders should assess sentiment around tech rollouts linked to performance improvement forecasts.
Risks to this view
Should regulatory bodies impose stringent controls on AI technologies or delays in adoption emerge, this could significantly alter market outlook and undermine the bullish sentiment currently prevalent.
Hello, and welcome to the latest edition of our UBS AI CEO series. I'm Ulrike Hofmann-Borchardt, Head of Global Equities and CIO for the Americas for UBS Wealth Management. Today, we are diving into one of the most fascinating intersections of technology and professional services.
How AI is transforming the legal world. In past podcasts, we have focused on companies enabling AI. This is the first time where we focus on applying AI to create value in the real economy.
I'm joined by Vincent Weinberg, CEO and co-founder of Aave AI, a company that's quickly become one of the most talked about startups in legal AI. Their mission is bold but simple, build AI co-pilots for lawyers to make the practice of law faster, more precise, and more accessible. Vincent started his career as a practicing lawyer at O'Melveny & Myers and has seen firsthand how much time and energy legal work can demand.
Together with his co-founder, Gabriel Pereira, formerly at Google Brain and DeepMind, he bridges the gap between frontier AI research and the day-to-day realities of the legal profession. In our conversation, we'll explore the origin story of Aave, the challenges of bringing AI into law, what adoption looks like inside top law firms, and what the future of the profession might be in a world where Aave AI is deployed at scale. So, let me kick it off with your background and how you decided to start Aave AI together with your co-founder.
Yeah, so I actually met my co-founder my second year of law school and we definitely didn't have a plan to do a startup. We just became very close kind of initially after meeting. We met in San Diego.
And I do remember the first time I talked to him, he was running an education startup and talking about how you could kind of use a lot of different modeling to teach people how to get better at chess. And you take that, basically all the properties that you've learned there, and you translate that to just how do you teach people to actually do personalized AI tutors and things like that. And this was before the launch of Chat2BT or anything.
But so I definitely was interested. I think that really what made us come up with this idea was there was one day in early 2022 where he showed me GPT-3. And GPT-3, there was a launch of it in 2021 and it wasn't very good.
And there was an update to it in 2022. And this was completely public. Everyone could have access to it.
This isn't early access or anything like that. And I started using it actually at work to a degree on not using confidential client data but doing things. I was on a landlord-tenant pro bono case and I was just doing kind of like local California statute research.
And I remember I put a bunch of the prompts in there and the answers were actually pretty close to right. And I kept messing around with prompting. And what my co-founder and I started doing actually is what's now referred to as chain of thought prompting, which is basically here's the scenario.
If you have A, then go to step A.1. If you have B, go to step B.1, et cetera. And we came up with kind of a long prompt that was for landlord-tenant law.
And we went on r slash legal advice, which is basically a subreddit where people just ask questions and then ask, who can I sue? Every single time it ends with that. We grabbed 100 landlord-tenant questions and we basically ran this chain of thought prompt reasoning over it.
And we got the answers and the outputs and we gave them to three landlord-tenant lawyers. So 100 of these. We didn't say anything about AI, literally nothing.
We just said, here's an answer that a customer or potential client asked. And here's the output. Would you send this to them with zero edits?
And 86 out of 100, two out of three or more of the lawyers said yes. And we took those and we literally just cold emailed at the time, his name is Jason Kwan. He's the general counsel of OpenAI.
And we emailed him and Sam just kind of these results. And we just said, hey, did you know that it was this good at legal? What was the response?
Yeah, the answer was, let's jump on a call. And we jumped on a call and we kind of talked about what our pitch was. And at the time, we actually were looking more at consumer law.
And so we didn't know tons about unauthorized practice of law. We didn't know tons about the ownership bar rules like 5.4, etc. And because we had initially gone into landlord-tenant and we looked at this, we thought, oh, maybe we could do something that helps folks with landlord-tenant, family law, things like that.
And then we actually ended up talking to a lot of folks about the regulatory environment and we decided that wasn't the best idea. But originally, the idea was actually more consumer-facing. And then what happened was we were kind of talking to OpenAI at the time as well.
And we decided, okay, we're either going to do some version of consumer or hopefully these models get better and we can actually go upmarket to big law, right? And so we ended up raising from OpenAI. And a couple weeks later, we actually got access to GPT-4.
And I remember I went back into my room for like 18 hours straight. And I tried to do all the things that I couldn't get GPT-3 to do. And it worked significantly better.
And we kind of presented all of our results to OpenAI and we kind of changed our thesis where we said, oh, wait, we can actually start making tools that we could sell big law. We didn't need to do kind of the lower level work. So OpenAI was the first investor?
Yeah. OpenAI was the first investor. We actually didn't do like a full fundraising process.
We literally just went to OpenAI. So this is sort of a testament that I think also OpenAI sees that, no, we might need more verticalized versions of these models with Harvey AI versus just what the frontier models can do. At a super high level, I think there's two problems to solve for kind of like all of these verticals in general.
And I'd say vertical application layer companies, but also just if you want to get the models to work in a vertical. The two problems I see, one is a product problem and one is an AI problem. The product problem is basically how do you give all the context to the model that a human would need to complete the task?
That sounds very simple. It's actually incredibly difficult. Like if you think about the context that you need to give the model so that it can do an LBO, that is an incredible amount of context, right?
And I think that like a lot of the advantage is actually how do you build a system that can get all of the context and pull all of that context out of a user, combination of external data sets, internal data sets, precedent, their emails, all of these different things. That's very, very difficult to do, especially as a general matter, right? It is very difficult to have a system that can reach into every single source of context and pull that context out.
That's problem number one. Problem number two is now that you have all that context, how do you make it so that the answer is correct, right? And the answer here could be a 500-page merger agreement, right?
It can be a very long answer or it could just be a chatbot answer, right? And I think that getting that answer correct is also incredibly difficult. And a lot of the reason for that is it's all processed data.
So if you think about like one of these large processes like an M&A or litigation or something like that, it isn't necessarily done in something that is reason, right? You can kind of think of this as if you had someone, the person you know who has the highest IQ and you threw them into the middle of like a very large merger, right? They would figure out how to do that merger faster than somebody with a lower IQ, but they wouldn't automatically know how to do that merger, right?
There are certain rules, there are ways, there are processes that aren't just kind of learned from general reasoning, right? And then the other part of this is the majority of legal work is actually based off of precedent, right? So if you're doing an M&A, a lot of what you're negotiating is actually just what the market rate is for a certain term, right?
And that data doesn't exist, right? In private equity that data is literally held at the law firms or it's held at the private equity shops. It's not even public, right?
And so really what you're doing as an application layer company is you're solving those two issues. One is how do you make sure that you get all of the incredibly complex context into the model to do a task? And the more verticalized the tasks, the harder it is to do that.
And then the second piece is how do you get the right answer, which again, it isn't just general reasoning. It's actually process and it's industry-specific, I mean, it's task-specific process data. So it seems like there needs to be more structure that needs to be imposed on this models from A, the context side, but then also there is some problem solving on the individual task and case that is at hand.
I think like long-term, the companies that are going to be successful are the companies that find verticals where the context that goes into the model is a lot and from a bunch of disparate sources. That's very important. Otherwise, if it's just all the context you need is just on the internet, right?
This becomes a problem if like web search gets really good, you're kind of done, right? So that's number one. And then the second piece is the completion of the task isn't necessarily a reasoning problem.
It is a process and industry-specific problem, right? And that makes it so that this data isn't on Reddit. And so hallucination, right, is a big problem with these large language models.
By using those two things and infusing the domain knowledge, you get to much higher accuracy and precision. Yeah. So you can think of this as the nice thing about having verticalized tasks is the more verticalized the task, the easier it is to evaluate it.
So if you think about the problem of how do I do evaluation across every single use case known to the human race, which is what ChatGBT is, right? Like that is very difficult to do evaluation on. And there's a billion benchmarks and all these things.
But those are becoming pretty useless for verticals, right? If you think about a more simple problem of, say, I'm trying to build a tool that does automatic merger control guidance, right? And so I upload a target's financials and then, you know, my financials as the acquirer, and it tells me in 50 different jurisdictions what I need to file for antitrust, right?
I can break that problem into 500 different steps, sub steps, right? And I can create an evaluation framework that is evaluating each part of that subset of the task, right? That is much easier to evaluate at scale than something that is every single task known to a human, right?
Even in a domain, right? It's impossible to build evaluation benchmarks that are incredibly accurate for every single legal task, right? We're trying to do that, and we have something called Big Law Bench, which is attempting to do that.
But our evaluation frameworks are much, much better if it is a verticalized task. How do you then solve that problem? I love this example, concrete examples, right, the merger filing, all the regulatory filings that need to take place.
Do you build then agents for different subcomponents? Exactly. So you can take, you build agents for the general processes.
So a general process would be case law research, right? And case law research falls into a lot of these subcategories of tasks, right? So if you're trying to build a, you know, write or draft or build a platform that can do a motion for summary judgment, you can think of a motion for summary judgment as a combination of all of these general tasks, right?
General task one is case law research, right? General task two is drafting. General task three is probably reviewing all of your documents and summarizing them in a discovery, right?
And so you can build agents that do all of those different disparate tasks, and then you combine those agents to do the task that is draft a motion for summary judgment, right? And we're starting to do this actually with LexisNexis. And so my point here is you can think of all of these tasks as more general to more specific, and then the more specific you get, all you're really doing is chaining a bunch of more general tasks together.
So it's not like maybe what some people may think is that you have a certain prompt and you just give it the prompt and the outcome is then sent to a law firm or a corporation. You really have, it seems, a very intricate tech stack, multi-agent framework that is built on top of these foundational models. Well, that's what law firms are, right?
If you think about what a law firm is, the input is an email from a partner. Well, technically the input is an email from a client to the partner. And then that gets basically forwarded out, right?
And the partner breaks down the task. And like a very good partner, a very good manager, what they're incredible at is, hey, you know, XYZ person does this, the other person does this part of the task, the other person does this part of the task, and you're going to go out, get all of that data from all those disparate knowledge sources, some internal, some external, basically compile whatever the sub-answers are, send that big package of items to me as the partner, and then I'm going to go through those and put those into an answer to the client. That's actually what the law firm is doing as well.
So when you now, maybe let's talk about just the go-to-market strategy, the M&A is one of the use case, but there's so many other filings, but what are the typical entry points for you to engage with law firms and also corporations? Yeah, it's definitely changed. We actually used to speak more to litigation partners in the super, super early days, and the reason for this was there was more publicly filed documents.
And almost all of our demos from, for like the first year basically, where I would go find something that a partner had filed in federal court, grab that, and use that as an example. And the best one always for a litigator was, how would you argue against this? Always.
And then rate this out of like a 10 or something like that. And the reason that worked is lawyers are just incredibly busy, and if you show them something that they recently filed, they pay attention. And they pay attention for two reasons.
One, I think it's because it's close in their memory, but the second reason is because they want to argue with you. They want to see the answer, they're going to read every single word in that answer and try to argue why that answer is bad. And if it's not related to anything that they've done, they can't do that.
And so you instantly get them super focused, which is just interesting because we did have kind of litigation partners at first using it, and then we went through a phase where I think actually the larger percentage of folks using it was transactional, just across the board, right? So M&A was a large one, securities was a huge one. And this is still kind of early days, maybe like year two, and now it's kind of across the board.
One of the main things we needed for litigators was a lot of litigation needs external sources, like case law is the number one, right? And so we needed to actually have a data partnership with a large case law provider before I think litigators would trust us to actually do case law research through Harvey, right? And you've seen a lot in the press about these kind of like case hallucinations and things like that.
Most people are using Chattoobee Tee for that. Yeah, and that's probably a good reason to say like, look, you need actually to build your own tech stack that talks to context, that talks to the specificities, right, of this particular case, maybe just to try to talk about value capture for your customers. Yeah, so I think a couple answers to this.
One is, surprisingly, I think it's around 21 or 22% of our users are partners, and they're using it for very different things, right? So I think a lot of the times when you talk about AI, it's just efficiency, efficiency, efficiency. They're not using it for efficiency.
They're using it for pretty similar use cases to, I mean, extended, but similar use cases to what I showed them originally, which is ideation, right? It is, hey, this is a new law that came out. How does this apply to a large hedge fund, right?
And I'm about to go into a meeting. There is this case, they just filed a complaint, draft, a 10-page summary of different arguments how we could respond to this complaint. It is more ideation than it is just efficiency, right?
I mean, a lot of them aren't doing kind of like large doc review projects. On associates, it's much more efficiency, right? So it is, you know, how can you make it so that you can do disclosure schedules faster?
How do you make it so that you can do just like normal M&A diligence faster? How can you make it so that you can, you know, do 10K filings faster, things like that, right? I think that the other interesting thing is how do you measure efficiency or the ROI with a law firm versus an in-house team, right?
And it's very different. So we are more and more, we are starting to have law firms that do ask us to run studies with them that are on hours saved, but that is an interesting conundrum for them, right? Because they build by the hour, right?
And so I think that the way that they're starting to look at it is twofold. There are some firms that are starting to say, we're going to go to more fixed fee, right? And, you know, the big four already does this for like the vast majority of their work is actually fixed fee just across the board, right?
And so I think I am seeing more law firms doing this, but I also see law firms say, okay, this is going to basically help automate the work that we don't make as much money on. So an example of this is a lot of law firms, they'll go to private equity and they'll say, we'll do the X, Y, Z type of lower level work at basically a loss. So that when you do your next merger, your next LBO, you'll go with us, right?
And so a lot of law firms have been justifying it basically through that, right? And so they don't need to actually switch from the billable hour. They can justify the ROI on using AI there.
In-house it's much easier, right? If you save in-house folks hours, they aren't billing anyone. So efficiency is great.
And we do have an increasing amount of in-house using it for things that they would send to external counsel, but it's really the low end work. And I think that a lot of law firms have seen this work as AI is going to start cannibalizing it anyway, and they're starting to move away from it anyway. So it's closer to the work that they would do at a loss.
So you maybe start off with these efficiency use cases on the low end, but then you've obviously had tremendous success within law firms, also within corporations. How does it evolve? Maybe talk us through like the typical customer journey and the realization of, oh, this is actually real ROI here.
We need to do more. Yeah. So on the law firm side, the nice thing is a lot of law firms now are starting with enterprise wide.
And so they're basically, you know, we just announced Latham, I think a couple of weeks ago, and they launch it to everybody at the same time. And the large advantage of that is it is still a massive investment on just training, right? How do you use these tools?
And even a tool like ours that is verticalized, there are still tons of limitations to it, right? And it is really important that you train folks on what those limitations are. And when a firm does one of these massive rollouts, everyone learns from each other.
Like I think people underestimate how powerful it is for everyone to be using a tool across the office, like across the hall from you. It is a huge thing. And I talk about this a lot, but there is also a little bit of like the transitive property for use cases.
In other words, you will have someone who's in securities and they will find this really cool use case for analyzing a thousand different documents and doing X, Y, Z with them. And then they'll have someone in litigation and they'll tell them about that. And the person that's doing white collar investigations or something like that will also say, oh, I could kind of like translate that to something in my practice that's super helpful.
With in-house teams, it's interesting too because we've started expanding outside of law. So there are a lot of in-house teams that we will sell to their legal team and then we will also sell to their tax team and their compliance team. And what happens from that is if you think about a lot of these workflows, these workflows for in-house teams don't just involve lawyers, right?
A lot of legal workflows involve tax folks, involve compliance folks, sometimes they involve sales folks if you're doing contracting, right? And so for them, it's a little bit more of you land and you expand to actually other verticals that are adjacent to legal. So it's interesting how has then the pricing or the contracts, how have they evolved with your customers?
Because it makes a lot of sense to me that you have network effects, right? And therefore enterprise wide license agreement may actually be the most accretive. But how has that evolved?
Where do you see that going? So we definitely eventually are going to have a consumption-based pricing element to our platform, 100%. I think that you don't want to do consumption-based pricing until your minimal viable quality is at a certain point.
So what do I mean by this? If you can get the outputs on a bunch of disparate tasks to be 80% of the way there, seat pricing makes the most sense. Because it's hard to pay for an output if the output is 80% of what you want it to do, right?
And so I think you need to get to the point where you have systems that can manage an entire project from start to finish before you're charging for consumption in our space. And if you look at the other spaces where people are charging for consumption, you just get the output, right? Ticket resolved, right?
Or something like that in the, like Sierra, right? And so for legal, it's just harder, right? I don't know how you would charge for litigation settled or something like that, right?
Maybe a part of the fee, right? Or if you have a higher success rate, because ultimately, right, that could also be part of the KPI. 100%. And I don't think the tools are there yet.
I will say the thing that we're really excited to work on as a company is we're starting to go into these, what we call workspaces. And what a workspace is, is how do you build an environment that is based off an entire client matter, right? And so that client matter could be a litigation, that client matter could be a fund formation, that client matter could be an M&A.
And I think there, you can get a lot more interesting ways to do pricing and packaging because it's project-based. And I think that you're actually doing a larger percentage of kind of the upfront document processing and work. Maybe let's talk a little bit about competition.
I think every single company on earth is competing with OpenAI and Anthropic. And I think just as a, and Gemini to a degree as well, I think that just to a degree of as the models get better indirectly, you need to just provide so much more value on top. Your platform has to, right?
And so I actually think of my number one competitor long-term as just the models get better and you need to make sure that your system is, provides significantly more value than the underlying models, right? And then I would put the other buckets of competition into, you know, there are other legal AI startups. I think like as we go into other verticals as well, there are other tax AI startups too.
And I think those have been growing really, really fast. And then I think the third bucket is self-built, right? This is waxing and waning.
In the beginning, it was a huge, like kind of everyone was, we're going to self-build ranging from we're going to build GPT-4 ourselves in-house to we're going to buy a dedicated instance on Azure and kind of build our own chatbot. This is starting, I think, to change. And the transition that I'm seeing is instead of we're going to self-build from scratch, we really care about a vendor allowing us to build on top of them.
And we do a lot to support that. So if you think about kind of, you know, as these products get better and as our platform gets better, a lot of the differentiation of a law firm is going to be how can you take what differentiates you today and make that better with AI or put that into an AI system, right? And we do a lot to allow law firms to do this, right?
So we make it so that they can build workflows on top and then actually just populate those workflows into another Harvey users instance. So if you have an AT&T as one of our customers and you have a law firm that supports AT&T, they can build a system on top of Harvey. And then that system actually shows up in AT&T's instance, right?
And so we're supporting that and we're doing a lot more of that long-term too. And so I really see those as the three buckets of competition. There's the foundation models, which I actually think is the most important to pay attention to and make sure whatever you're building is just a massive amount of value on top of those and that they don't kind of snowball through you.
And then the second one is just other legal AI and professional service AI companies. There it's just move as fast as you possibly can and execute better than everyone else, hopefully. And then the third bucket is self-build.
And I think the best way to compete with self-build is to not compete with self-build. It's actually to help them self-build. And how about confidentiality?
There's a lot of data that really is very private. Is that another vector of differentiation? Yeah.
We spend an incredible amount of resources on security. And I think that we're starting to work with large banks, other financial institutions. And I do think the way that you architect your system with security in mind is actually like a massive differentiator.
I think that that can go away over time, right? Because other folks can do the same thing. I think security can be a pretty large differentiator.
And I also think that one interesting thing that we've ran into as we've worked with more in-house folks is we've actually had the situation where sometimes a law firm will say, hey, our client doesn't want to use AI, right, for a particular matter. And then we'll actually find out that that client is using Harvey currently. And so there is a little bit of back and forth, you know, and these are like large organizations in terms of kind of navigating what is the company okay with, what is the law firm okay with, which ultimately is what is the company okay with.
And is it right to think about that you have a fully sandboxed environment for all your client data, so there's really no leakage at all? Yes, 100%, right. And that's something we did from the beginning.
And one thing I learned a lot about very, very quickly is data processing in different countries. So we're in 56 different countries now, and I know the data processing laws in basically all of them, or at least roughly. And so one of the things we did really early was actually set up dedicated instances in all of those countries so the data never leaves.
And how do you think, I mean, this is another differentiator, right, just even being able to navigate the different laws and jurisdictions between different regions, countries, states, and so forth. How do you think, if we now think about Harvey AI being deployed at scale, like, say, five years from now, how is this going to change the legal profession, you think? Yeah, so I actually think the most important thing that it's going to do is increase speed.
And if you think about the most important legal tasks, speed is really important. And it's interesting because we think of legal as a very conservative industry, and a lot of folks in companies think of them as the cost center and the no people, right? But actually, the most important function that a lot of law firms and your internal legal folks have is how do you prevent problems down the line and move things along faster now, right?
If you think about a large merger, the main thing that stops a merger is time, 100%. It's time, right? And so if you can actually do disclosure schedules, you can do all of the diligence much faster, it's actually going to make your company better, right?
If you can buy, if you can look at companies faster, if you're a private equity fund, and you can actually do diligence on 10 times the companies in the same amount of time, you're going to make better purchasing decisions, 100%, right? And so I actually think the main thing that we're going to start seeing, especially five years from now, is we're still going to talk about efficiency, but we're going to talk about performance and speed, right? And if you can do a lot of these tasks faster, if you can prep all of your documents internally before an internal investigation or something like that, that has so much value to a corporation.
And this is a great point because we are thinking so much about loss of labor, right? But it could also create much more activity. And that obviously requires more labor in totality, even though the proportion of AI within a specific task is higher. 100% agree.
And I look at this from two sides too. I would look at this from the corporate side, which obviously, if you can do more mergers or at least look at more companies and make better decisions before purchasing a company, that's fantastic for you. And for the law firm side, I look at this from two sides.
One, I think they're going to keep getting increasingly levels of legal work. And also, think about how much legal work there's going to be about navigating AI, right? That's going to be a massive, massive new industry for law firms.
But I also think about this from the associate side. I'm 30. I was a junior associate three years ago.
And I think about this as, how amazing is it going to be to work in this profession if when you're 35, you can be the lead person on a deal, or you could be first chair at a litigation, right? That is what people want to do. I mean, I know almost no one that went to law school because they really, really want to be in data rooms, or they really want to do doc review for discovery, right?
What they want to do is be in high stakes situations and work with clients, right? And the problem is, right now, how it works is you spend so long doing discovery, doing data room work, and you should do some of that. I'm not saying don't do any of it because it does teach you a lot, but you spend so much time doing that, that you don't actually develop the client-facing skills until way later on in your career.
And I would argue that the most fun part about law and the most valuable part about law is the client relationship. And if you can develop those skills earlier on in your career, it's going to be a way better profession. We will attract better people to the profession, right?
We will always attract the best folks, and I think it will be overall not good for the industry. Sounds very compelling. Let me maybe talk a little bit about where else you want to go, since you've already painted that great a picture for the legal profession.
What else other professional services lend themselves, I think, to similar type of AI amplification? We think of it on like a task base or a project basis. If we go back to the M&A example, which I've given a hundred times...
It's a good one. But if you think about like a large M&A, it's not just lawyers, right? And so you're doing commercial diligence, and then you do tax diligence, right?
And so it's not that we are going to open up an entirely different vertical and just focus on that anytime soon. It is more we are going to add different verticals and specialties to project-based systems that we're building, right? So if you're building a system that helps with fund formation or M&A or whatever it is, for completeness sake, can you add these other verticalized solutions on top to make it so that the platform can do the whole thing from start to finish?
And so taxes is obviously a clear one. Compliance is probably the other one. I would say that those are the two big main ones.
So it's legal tax and compliance. Would it also expand more broadly into consultancy services? We don't have immediate plans to do that.
And I think it's also important that we stay just very focused since we're still so young. But I do think there is a world in which we kind of involve these other verticals and especially from a task completion standpoint. So everyone that is involved in an M&A, everyone that is involved in a fund formation.
Yeah. Communication is certainly also part of that, right? Exactly.
So I can think of this as a lot of, go back to that question that is basically, can you take all of the context and give it back to the model that you would need for a human to do the job? A lot of this context is in a similar spot, right? So the context that folks need to do tax diligence is actually usually provided in all of the documents just for a general M&A, right?
And even for comms, right? So there's a lot of pieces of this work that is actually just like a summary of what happened in the M&A. And maybe let's zoom out a little bit to AI more broadly and where we are.
I'm curious, do you need more capabilities for the large language models in order to do what you're doing right now? Or this is good enough. Now it's really just about building the agentic framework on top.
Yeah. I think the reasoning ability of the models in just kind of like if you think of it as like a raw IQ, we don't need much progress. I think it would be great and it would be helpful for some of these use cases.
But at the end of the day, I think that if you just think about what is the percentage penetration of legal tasks that Harvey can do right now, it's low, right? If you think about it from all of the different tasks that you can do from start to finish, it's a low number. That number can get maybe 50 times higher, maybe at least 10 to 20 to 30 times higher without the models getting any better at all.
And it's just how do you make sure that you give all the context to the models and then teach the models to do the right reasoning steps, right? And again, those reasoning steps are not necessarily something that just, you know, someone with a super high IQ would know how to do. It's process.
It is actually industry-specific process. Yeah. So we don't need to get to AGI in order to get productivity enhancements that we see in some of these pilots.
No. There are a lot of features that I want to get better. And so, you know, like anything that is making it so that tool usage is better, that is incredibly helpful for us, right?
And so there are different functionalities of these systems, I would call it systems instead of just models, that are really important for us to improve on. It's just not necessarily, we don't need GPT-6 to have a significantly higher reasoning IQ than GPT-5. Yeah.
Makes sense. And how do you think about software more broadly? You have obviously proven out that there is a lot of scope to using AI in the legal profession, but how do you think that's going to evolve for other areas of software like customer service, go-to-market, IT?
So I think that one of the main things that we're going to see a lot more focus on is that thing, basically what I've said in the beginning, which is how do you get more context to the system? And I actually don't think a lot of companies are as focused on this as they should be. And I actually break this down to a simple problem that I think we underestimate, which is humans aren't very good at communication, right?
Like if you think about the folks that you work with the best, it's you send them instructions and they send the perfect follow-up questions to clarify what you definitely missed in your instructions and you work with them and then they can go on their way and they can come back, right? I think that a lot more work needs to go into that first part of how do we extract basically the context or the intent out of the user. And this just goes to say that I think you basically have two problems.
You have the legacy systems that have a lot of that context, but those companies might not be as good at what do you do with that context, right? So once you have all the context, can you make the AI systems give you the correct answer? That's going to be hard for them.
And then you have companies like us where we're much better at getting the right answer once we have all the context, right? But it's hard for us to get the context because it's locked in old ways of doing work or whatever it is. And so I think there's going to be this pretty large battle between those two.
And what at least we've seen in our industry is we integrate, right? So we integrate with all of the large DMSs. We integrate with case law, with LexisNexis.
And so we're not trying to just kind of automate that problem. What we're trying to do is say, hey, you have the context. We have the AI on top.
Let's partner together. And I think you're going to see that in a lot of industries, but you're also going to see a lot of blocking that, right? We see what Salesforce did to Glean with Slack usage, right, blocking it.
But I think that's the number one problem that needs to be solved is the legacy companies have the context. The AI companies know what to do with the context. What happens in that middle area?
But wouldn't you think that over time, these next generation companies would also just gather the context because of course you're already operating in this environment. So the question is, what is the value of potentially a context when the world is... It's just faster.
Yeah. So who's faster? Are the legacy systems faster at figuring out what to do with the context and to get the right answer given the context?
Or is the new folks on the block faster at building the context, building those datasets, or potentially rebuilding some of the systems of records that have existed in the past, but AI native, right? And so that's another world that we might see. We might see some of these AI companies say, wait, I don't actually need to rebuild the system of record because in an AI first world, a lot of the features of a system of record are irrelevant, right?
And so it might be faster for them to rebuild the system of record because they can just take the pieces that are necessary now. Where do you come out? I think it really depends on the industry.
So I'll give you an example. In ours, there's also the goodwill value of the data, right? And so LexisNexis, Relix, the parent company of LexisNexis was founded in 1818, right?
And so they have a lot of goodwill, right? And going back to M&A, I'm going to just say a billion things about M&A. In M&A, there's something called a goodwill valuation and there's actually like a way to price it, right?
And I actually think that for a lot of datasets, whether it's internal or external, there is like a goodwill valuation attached to the company that provides that data. And for the internal datasets, something like a workday, the goodwill valuation they have is people trust them from a security standpoint, a compliance standpoint, right? All of those things.
The goodwill that Alexis has is people just trust that the data is complete, right? And so it just depends on how powerful that goodwill is. And depending on an industry or a particular site, it's really powerful if Relix was found in 1818, right?
For other companies, maybe not so powerful, right? For internal datasets, I would argue that for internal data, it might be less because you can gain the trust of the company over time and they can give you that data, hopefully. I mean, a lot of SaaS companies are built on internal datasets.
I agree with that. I think there is an argument that if you are the application layer company and you can provide such a better experience, now it might be worth the really annoying process of switching, right? And 10 years ago, it wasn't.
Like if you have a 10 times better, you know, sorry, if you have like a two times better mousetrap, which is what you used to have, or maybe 1.5, right? Change management, like actually taking all of these documents, putting them in a different system, all of this. I don't know if I want to do that.
I don't know if I want to pay the big four a lot of money to do this for me. And so maybe I don't want to do it. But if the better mousetrap is 10 times, 20 times, 40 times, I would argue a hundred times better.
It might be a different story. Couldn't be a better finish. A few questions, just lighting around.
One thing law school didn't prepare you for, or one thing that law school prepared you for. Okay, okay. I think actually a better, yeah, I'm going to do the one thing that law school prepared you for, because law school definitely didn't prepare me for running a company, for sure.
One thing it really helped with that I think we have a unique situation in our company because we have a lot of lawyers is attention to detail. And it's incredibly important. And I don't think tech pays actually that much attention to it.
And rightfully so. Like I think that there is some amount of attention to detail that is a waste of time. And it is to show discipline or because that's how we've done it for the past 50 years and that's how my generation did it, so you have to do it, right?
That I don't think has much value. But I do think there is something about reading every single sentence, checking, like your brain immediately saying there's an error here. There's something about that training that you have in law school.
I mean, I remember I would lose points on an exam if I didn't double space, 100%, right? And like that level of attention to detail, it does help you when you're running a company. And my leadership style is I read everything.
And I think normally folks would say that that takes a long time. But I actually think I'm a very, very fast reader partially because of law school and partially because of the legal profession where it forces you to be really fast at reading and pay attention to detail. I mean, there are some partners that without doing like a, you know, a control F double space, which is by the way, anyone listening, the trick to getting double space on everything is you do, you basically do control F and then you search double space and then you can tell if a sentence break has a highlight or not.
It's faster. Anyway, I knew like there were partners that you could print out a page and they could instantly be like, that's not double spaced or that's, you know, that's an em dash, not a hyphen or whatever, right? And they could do it really fast.
And I do think that there is some value to having that attention to detail. And I like to read literally everything and it makes it so that my decision making process is really fast. And so maybe I spend a lot of time upfront reading everything.
But the flip side of that is I make decisions incredibly quickly. And I don't think they're based off of quote unquote, God, they're based off of I've read every Slack channel. I've read every document.
I know what's going on. And so I have like a preconceived notion of like what the right answer is here before actually thinking about it for that long. Most overused phrase in law and in AI.
You probably deduplicate pretty well too. In AI, it's definitely agents. And I think it's just that we're all over the place with what we define as an agent.
And there's like the RL definition of agent, which nobody uses anymore. And then there is kind of just everything in between. I think that's definitely the number one used.
In law, I hear a lot of prompt injection. And I'm not sure people know what that actually means. But we kind of go through a lot of these buzzwords that end up people kind of picking up.
If you weren't building Harvey, what would you be doing? So my plan out of law school, actually my plan to go to law school, I was not a very good student before law school. And then my first year of law school I started was the first year I started working really hard.
But I wanted to go to law school because I had an internship at the U.S. Attorney's Office in the Eastern District of Louisiana. And for everyone that doesn't know what the U.S.
Attorney's Office is, it's federal prosecutors. It's assistant U.S. attorneys. And they kind of handle incredibly complex, very high stakes cases.
And it is, I would argue, the most fun job in the entire legal profession. You work with the FBI. It's incredible.
And it is a very hard job to get. It's like harder than working at the best law firm on earth. And I went there and I met all the, based off of all the people in the work, I was like, I need to go to law school.
This is what I want to do. And my plan was always to work at a law firm, become a U.S. attorney. You almost entirely had to work at a law firm first.
And so you do that, become a U.S. attorney, do that for maybe five years or something like that, and then go out and try to start my own firm. I really wanted to build my own litigation firm. And this has been happening a lot in California.
There's a lot of boutiques that are really, really good. So like Hecker Van Ness is a super famous one. Houston Hennington is a very famous one.
And I wanted to go out and do something like that. So maybe I'd be doing that. Awesome.
What do you think is the most exciting AI application outside of law? I mean, I think coding is going to be the one that has the most impact on the world in the short term. I think that just the amount of software that is going to be able to be built and incredible software once these coding systems get better is going to have such a massive effect on the rest of the economy.
So I would say coding. Coding. Super.
Thank you so much. Winston, such a pleasure to have you here today. Thank you so much.
To our listeners and clients, thank you for your time and engagement. If you have enjoyed this episode, you may subscribe to UBS Market Moves, available on Apple Podcasts and Spotify for additional content. Until next time, where again, we decode the future of AI layer by layer.
UBS Chief Investment Office's investment views are prepared and published by the Global Wealth Management Business of UBS AG or its affiliate, UBS. This material has no regard to the specific investment objectives, financial situation or particular needs of any specific recipient and is published for informational purposes only. As a firm providing wealth management services to clients globally, UBS AG and its subsidiaries offer both investment advisory services and brokerage services.
Investment advisory services and brokerage services are separate and distinct, differ in material ways and are governed by different laws and separate arrangements. In the USA, UBS Financial Services Inc. is a subsidiary of UBS AG and a member of FINRA-SIPC. For information, please visit our website at ubs.com forward slash working with us.
For a full legal disclaimer applicable to the independent investment views produced by UBS, please visit our website at ubs.com forward slash CIO dash disclaimer.
Sources & References
How we cover this story