‘The Future of Legal AI Is Collaboration’ – Harvey

In this Law Punx blast, Harvey’s co-founders Winston Weinberg and Gabe Pereyra set out how the future of legal AI will be about collaboration between law firms and their clients, and where there will be a ‘multi-player’ approach to AI platforms.

They also discuss the current state of AI adoption in law firms versus in-house teams, the importance of knowledge sharing, and creating collaborative AI systems that enhance workflows. The conversation also touches on billing practices and the need to attract talent to the legal industry as it adapts to new technologies, as well as how AI enhances what junior lawyers can do.

Please press Play to watch inside the page or go to the AL TV Channel, where there are over 190 other videos. There is also an AI-generated transcript below.

—

Legal Innovators UK starts this Tuesday, 4th Nov, then Inhouse Day on the 5th, and then our new Litigation Day on the 6th Nov. 

—

Takeaways:

• The future of legal AI will heavily rely on collaboration between firms and clients.
• In-house teams are currently a year behind law firms in adopting generative AI.
• Collaboration tools must address both internal and external needs in legal practices.
• Knowledge sharing is becoming increasingly important for law firms to maintain competitive advantage.
• Law firms need to develop workflows that can be shared with clients effectively.
• The demand for fixed fees may increase as collaboration improves.
• The legal industry must adapt to attract new talent and retain existing professionals and AI can help there.
• And also we look at the analogy of foot soldiers vs special forces in terms of legal skills and the law firm hierarchy in the age of AI.

—

AI Transcript

Hey everybody, Richard Tromans here again at Law Punx. Today we have Harvey, a company that you have no doubt heard all about. But we are very, very lucky to have with us Gabe and Winston, the co-founders of the company. They are going to present their ideas and we’re going to have a bit of a chat. In fact, I’m not even going to introduce your idea. I’m going to let Winston kick it off. What is the big idea?

WW (00:33.179)

Thanks so much for having us. We’ve been talking about getting on here for a bit, so thank you for giving us the time. I think the main idea that we would like to chat about is something that I haven’t seen as much talk about in the industry overall. And that’s, we really believe that the future, especially next year, and for 2026 and I think onwards through 2027 as well, is really going to be collaboration between law firms and their clients.

And I think even maybe taking a step back, I actually don’t see tons of discussion about how in-house folks are using AI. And so maybe as like a little bit of a starting area, I can walk through what Gabe and I have seen in the past three years in terms of adoption and kind of those changes. So the first thing I would say about in-house is I’d say they’re about a year behind in terms of adopting generative AI compared to the law firms. And it was just interesting because

I think when Gabe and I started the company, we definitely thought that in-house would move maybe at a similar pace to law firms. And what we found is actually law firms started adopting this in kind of early 2023 and really a lot last year. think 2024 was kind of the liftoff period. And 2025 has kind of been, it’s transitioned from do we adopt legal AI at all to what are we going to do? How are we going to adopt it?

And that’s kind of where law firms are. I would say that in-house is actually about a year behind, where they are kind of how law firms were in 2024. So right now, we spend about 50 % of our time with law firms and about 50 % of our time with in-house teams. Last year, was, oh, god, what do you think it was, Gabe? Like, 90-10? Maybe something like that, maybe 80-20?

And I think the transition is actually very similar to what law firms went through, where they are doing what law firms did in 2024, where they’re starting to say, okay, we do want to adopt Gen. AI, and what is our plan for it, right? They aren’t necessarily at the stage of, okay, every single department is going to buy some Gen. AI tool, and 100 % that’s what we’re going to do. But they’re actually starting to think about that and starting to think about the long-term effects on their orgs, how much they spend on outside counsel.

What is their tech stack? And so I think that’s been a massively changing environment for us, especially the last six months. And I think we’re going to see a lot of this next year. And maybe another point on the broader ecosystem. The word I would use for this is collaboration. The alternative word I would use is multiplayer. And if you look at what Microsoft is doing, what Anthropic is doing, what OpenAI is doing with all of these products, they’re trying to do this, right? And I think the…

Richard Tromans (03:27.695)

What do you mean by when you say multiplayer? How do you mean?

WW (03:31.118)

Yeah, so if you look at what they’re basically releasing for enterprise and OpenAI and Microsoft specifically in the past couple of weeks, they’re releasing features that basically allow you to collaborate internally with AI tools, right? And sharing threads, things like that. And it’s definitely new, and it’s pretty low level to start, but you can see how this grows over time. And I think the interesting piece of this is

If you think about the problem that law firms and in-house teams face with AI, they face the same problem that normal corporations have, which is how do we make our products multiplayer across internal orgs? But then there’s a second issue, which is how do you make it multiplayer for internal orgs plus your outside counsel? Or if you’re a law firm, the reverse, is how do you make it multiplayer, not just for all of your associates working on the same matter?

But how do you make it multiplayer for your clients that are also working on the same matter? And so I think that this is going to be maybe the summary of those two points is one, we’ve definitely seen in-house folks start to really think about how they’re going to adopt generative AI. And that’s changed a lot. They’re probably maybe a year behind. But I would say that it’s ramping maybe even faster than law firms now that they’re starting to catch up. And then the second piece is

I think if the future is going to be collaboration or multiplayer between a law firm and an in-house legal team, you really have to solve a problem that these general systems aren’t trying to solve, which is how do you do external collaboration, not just internal? And internal is going to be a nightmare, by the way, with permissions and ethical walls and everything. But then you have this second issue of a lot of the in-house folks we talk to is, well, we’ve bought Harvey, our law firms have bought Harvey.

How do we collaborate in the same system together? And that’s actually probably where we’re spending maybe like 90 % of our time is figuring that out. Gabe, I don’t know if you wanna add to that.

Gabe (06:03.98)

Yeah, a lot of what we’ve been thinking about to Winston’s point is I think if you look at the evolution of

in the past three years, how people have thought about these Genai products. It kind of started with a co-pilot, which was some tool that is embedded in, you know, maybe a Word document or a code editor if you’re programming. And then now it’s kind of evolved to agents where we’re starting to see these more autonomous AI systems that you can delegate tasks, they can go and use other tools. We have some programmers internally that orchestrate tens of agents to solve tasks. And so I think

kind of the next step in what Winston’s talking about is if you think of the big problem law firms are solving today, it’s you have thousands of highly trained, super specialized attorneys working on tens of thousands of different client matters or different client data. so thinking about how you deploy AI in that setup is I think going to be this like really interesting technical problem where to Winston’s point, I think a lot of the cloud providers, model providers, are starting to solve the problem of how do you make sure that these agents only access the data that they can see? But the whole reason you have the legal tech industry is you can think of a data room as kind of like a Google Drive or a shared file system, but you need this completely unique ethical walls and all this legal. ..like information architecture on top. And so that’s a lot of what we’re starting to think about. And we can maybe share more of kind of this matter OS concept that we’re thinking about of how do you build all of the guardrails, the process where these law firms can start thinking about not just an individual lawyer interacting with an agent, but how does a deal team or a litigation team work with a team of agents on a client matter? And then at the firm level, how do you deploy thousands of these agents on all of your client matters. And then to Winston’s point, the last level is how do you share that with your client? So now you’re working with your team of lawyers, the agents and your client all in this shared workspace. And I think there’s a bunch of very interesting both data privacy, security, confidentiality, privilege, all of these both technical and legal data problems you need to solve. There’s, I think, all the questions around product of what are these interactions

And then there’s questions around AI, how do you make these systems better at the task they need to do? Because most of these tasks are not gonna go into the public foundation models. And then what does that collaboration look like with law firms and their clients? Because it’s a very like asymmetrical relationship, right? Clients don’t wanna see everything that a law firm’s doing. And so you need to be really smart about which parts you share.

And I think the big opportunity we’re seeing for law firms is, to Winston’s point, these in-house teams are now starting to think about this technology and adopt it.

But I’ve been talking with a lot of CIOs and they’re really seeing demand from in-house teams to act almost as implementation partners and really help these in-house teams think about, hey, this is how we’ve been using Gen.AI for the past year, two years, three years. These are the parts of your internal process, how you do contracting, things like that. that we can help you build workflows and we can share those. And so I think we’re seeing kind of a bunch of opportunities both on for us how we can build some of this infrastructure and especially for law firms on how they can take this and work with their clients in new ways because we think this collaboration is going to be kind one of the things that defines how law firms differentiate.

Richard Tromans (09:55.151)

Yeah, interesting. No, I just want to ask a question that I the I mean, something that some people have been talking about, you know, is more so much more focus on knowledge sharing as well. I mean, there’s several points that you both made together. I mean, the one thought that came to mind is, you know, a law firm has a platform, its client doesn’t have Harvey, how does that work? Another law firm, the client is working with is a long-term user of Harvey, how does that work? Does that change the multiplayer? They’re no longer a guest, they’re an equal partner, as it were, at least technologically. And then you’ve got the aspects of sort of there’s a creative tension there around knowledge management and knowledge control. As someone said to me the other day, clients work for a particular law firm, they effectively train the law firm on its problems, on its legal issues.

WW (10:49.168)

It’s literally what they’re doing. Yeah, yeah. Yeah, exactly.

I mean, one point I guess to what you were saying, like we’ve talked a lot, I mean, just the industry overall of like, say these models get really good, what is the differentiator of all the law firms? And I think one of the really basic ones is actually like, if you are a bank and you’re trying to open up in a new territory, one of the massive values of a law firm is they’ve done it for other banks.

And so I think actually that data sharing with clients already is a huge part of law firms’ notes. If you’re a law firm and you work in the private equity industry and you have all of the different data points of what types of private deals, which terms they agree to, it’s actually a massive advantage. And it’s part of the reason why private equity goes to you when they’re trying to negotiate one of these mergers or M &A or whatever, LBO, whatever it is.

I think that that is something that my guess is law firms can actually distill more and more readily share with clients. And so that mode is actually just going to increase over time.

Richard Tromans (11:59.267)

Yeah, essentially, so you’d see so far the AI doesn’t break down the proprietary information in some ways, it actually makes it easier for them to share, which actually almost accentuates the ability for them to tap their knowledge.

Gabe (12:16.008)

I was going to say like one thing that I would add that is an interesting problem we’re starting to talk about is I think exactly what you said of the value of these law firms is I’ve worked with this specific client for two decades and I’ve done hundreds of millions of dollars of legal work. How can together we build a model that encodes that knowledge? I think where there’s a bunch of interesting technical problems but also IP problems is what Winston’s talking about is how do you start separating which of that knowledge

is the clients, which of that knowledge is the model? Because what you can’t do is train that model between a law firm and their client, and then share that with another client. But there is parts of it where some of it is this is my expertise of how banking legal operations works or how private equity fund formation works. And then there’s some of it that this is sensitive, this is the work that my client is doing. And so I think there’s a bunch of really interesting conversations here to start figuring out what does that look like. But that to me seems like a very big opportunity for.

Richard Tromans (13:23.383)

Yeah, and it goes back to your earlier point about ethics, data control, and so forth. But when you say model, because it’s a term that people misinterpret sometimes, when you say it’s a model between that lawyer and that client, you don’t mean a language model, do you? You mean something else, or maybe you do mean a language model.

Gabe (13:39.628)

Yeah, I think when I say model, just generically, like I think of most of these systems as you give it some input and it gives you some output under the hood. could be a language model. could be an agent. could be a non LLM GenAI system. But I think eventually people will start thinking about these systems the same way you think about Google search where you have a search problem, you type in a query, you get a result, you kind of don’t care what Google’s doing behind the hood. And I think these systems will become that way. But there’s still a bunch of technical challenges of thinking about if that system is training, how do you prevent that from data leaks and things like that? And so I think there is a bunch of interesting problems to solve around that.

Richard Tromans (14:21.453)

So it would be almost, mean, we talk a lot about playbooks for in-house lawyers, you know, who get a high volume of similar documents, but would it be like a sort of very intelligent kind of self-learning playbook that both the law firm and the client would kind of share? You know, kind of mutual, but, as you said earlier, there would have to be some kind of partition because obviously the law firm wants to retain some of its personal IP. Otherwise it kind of why have a law firm.

Gabe (14:35.608)

Yeah.

WW (14:46.766)

Yeah, and I think this is going to have to be negotiated in engagements too. Like that’s my gut of what ends up happening over time is like, say you’re one of these massive companies that has a panel, you might put terms into your panel that like certain data they can use to train their own system, like on the law firm side and certain data you can’t, right? So that might, my gut is actually that’s how it will end up being resolved is it’ll be done engagement by engagement.

Richard Tromans (15:14.095)

That actually is a really interesting point as well. Sorry, guys, just one thing about panels, because panels have been floating around for so long, that they’re so, even now, quite ineffective. I was just thinking, wouldn’t it be amazing when, rather than sending off a request for proposals, that they literally went, here is our data, here are some of our playbooks, these are some of our workflows.

Whichever one of you five law firms can do it best or work with us best. We’re not even going to get into the fact that Bob or Mary is a better lawyer, right? We’re talking about your systems control, your systems design, which one of you has got the best design we will work with. know what I mean? So AI becomes a real defining.

WW (16:02.798)

I was gonna let Gabe answer this one, because we’re already kind of working on stuff like that.

Gabe (16:06.264)

Well, we’re actually really seeing this. So I’ve been like working with law firms to help them build workflows and pitch them to clients. And we’re seeing like huge demand for exactly this, because I think to your point, this is such a good way to showcase, this is how we’re thinking about this technology. This is, you know, how you can use it effectively. And I think the appetite we’ve seen from clients to have these law firms help them think about how to build those systems, I think you will start seeing more things like this. And I think these clients will start to expect that.

Richard Tromans (16:40.472)

So it doesn’t mean we’re not giving anything away. How do you do that? How do you build these workflows so that the law firms can share them?

WW (16:40.686)

Yeah, I think it’s

Gabe (16:48.93)

Yeah. So at the, at the product level, it’s kind of what we were talking about earlier of we sell Harvey to a law firm and to a client. And so we are just building product features where you can develop a workflow in Harvey and share it in your clients. Instance of Harvey. then at the process or operational level, I think it looks pretty similar to how law firms work with clients today. They have a good understanding of here’s all the problems I’m solving for you.

And I think what we’re seeing the really forward thinking law firms develop is they’ve been building a bunch of workflows internally to make themselves more effective. And then they can start mapping those to, know, here are the customer problems that I want to solve. And so Playbooks is a common one where maybe I do a bunch of NDAs for private equity and I want to create kind of some Playbook workflow that can help automate parts of that. And so we’re seeing a lot of that.

Richard Tromans (17:44.176)

really interesting. mean, it kind of connects to the sort of infinitely everlasting debate about, you know, will in-house lawyers take on more work? And my gut feeling is, is that they don’t want to take on more work, but they do want the external families to be more helpful.

WW (17:57.168)

No, 100%. And I think that’s the point of it being collaborative. And I also think like, it depends, we’ve seen that it depends on the kind of like the category of in-house. So there’s some in-house folks that are basically just a law firm internally, because they just have so many. And I think those folks are more okay taking more work inside, but then there’s like asset managers who they 100 % want to keep paying law firms. They just want to collaborate with them using AI. And then I think the other point to what Gabe was talking about,

WW (18:26.584)

If you think about where this goes, like an aggregation of workflows can end up being an entire matter. So say you’re doing an &A, you could basically have it so a law firm sets up here are all the &A workflows. And that’s like one general area. And then it’s not just &A, it’s private equity &A, right? And then it’s private equity, so that’s like tier two. And then tier three would be it’s private equity &A for Blue Owl or whoever it is that we’re actually working with. And so you can get down from the, here’s a type of work that we do, to an aggregation of that work, to just further aggregations, basically. And I think that gets really interesting, because how do you put that into a product? How do you make it so a law firm can basically spin up a &A matter that has all of the workflows that they use for that particular client in there?

And then what do you do from the client side so the client can see some of that basically?

Richard Tromans (19:24.737)

Yeah, I guess that could happen with litigation as well. know, different similar cases going down the same route, same courts, that kind of thing. And certainly where you get that sort of complex, but routine flow, like insurance companies who are just constantly being the same. Obviously, know, Mr. Smith and Mr. Jones have crashed their cars at different places, but it’s the same thing again and again and again.

WW (19:28.046)

Yes, every area.

WW (19:37.764)

Yeah.

WW (19:46.858)

And again and again, yeah.

Gabe (19:47.854)

Oh, I think actually, well, it’s interesting. So for kind of the banks and private equity firms, we see a ton of interest for the transactional workflows for a lot of the Fortune 500 insurance companies, especially. There’s a lot of interest in how do you do this for litigation for exactly the reason you said. And then I think one thing that is interesting is I think there was a fear of as these models get better, do these in-house teams try to in-house more of their work?

WW (20:03.94)

Yep.

Gabe (20:15.628)

And we’re actually hearing the opposite where I was talking with the CEO of a large bank and he made the really good point where most of the work that we use outside counsel with variable costs. And so we actually don’t want that in like our firms operating structure. And so what we want is better ways to collaborate with our law firms and make them more effective. But our ideal is to leverage them because the things they’re doing are so complex and they’re not part of our daily operations.

Richard Tromans (20:45.359)

you know, I’m going to say billable hour, just do that very quickly. I mean, it’s going to be with us till the end of time, I’m sure. But perhaps there are more areas now where because there’s more collaboration, there’s more understanding on both sides about the benefits of data workflows that some of the clients just go, you know what? We’ll just give you a flat fee because we really value what you’re doing here. Or maybe it’s on a sliding scale and we’ll give you this, this and this. But we can just knock off all this whole hourly stuff because we value.

what you have and that’s what we’re going to pay for. Yes, there’ll be a bunch of variable work on the hour around the edges. But you know, the truth is we know what we’re paying you for. We’re not silly, you know, and you’ve invested in technology. You’ve put the time in to bring this data together. You’ve got a bunch of very smart people who are constantly updating their understanding and knowledge. That’s what we pay for. And you know, maybe that could like help shift more that way.

Gabe (21:48.396)

Yeah, I think this one, how we think about it has been evolving. think we hear a lot of people say, great, like everything will just switch to fixed fees. But I think the more we’ve learned about the industry, it’s just way too complex. And I think there’s a bunch of like human psychology. was talking to some partner who was like, I always offer fixed fees and I’m like, Hey, this is just going to cost X upfront. And they’re like, actually I’d rather just like pay as we go. so I think there’s a bunch of.

Richard Tromans (21:58.637)

No, not anytime soon.

Richard Tromans (22:17.906)

It’s a two-way street. Absolutely, it’s a two-way street. Let’s not labour the billable hour point, but just broadly, think the collaboration bit, it’s interesting because people have talked for years about collaborating more. But mean, perhaps AI is the bridge that makes it happen.

WW (22:18.362)

But then it ends up being more. But then it ends up being more, but it doesn’t matter.

Gabe (22:41.154)

Yeah. Well, the thing I wanted to tack on with the billable hour point that I think maybe becomes interesting is I one of the challenges with the billable hour is. It’s hard to quantify the hour of work where some of those hours of work feel like I shouldn’t be being built for this, where it was okay, this person was redacting this document. And then some of those hours are, actually came up with a novel way to structure this transaction that is worth infinitely more than I’m being paid for that hour. But I think what actually I imagine will happen, which will be really interesting is as these models start getting better and we start building more of these benchmarks.

You can start more accurately quantifying the units of work. And I actually think you start seeing this like convergence of here is what an hour of associate work looks like. Here’s what an hour or a, you know, whatever unit of compute of model or agent time looks like. And then you can start separating, okay, these are the things the models can do. Maybe that starts going away in terms of like, we build those, but then you also see this expansion where it’s like, here’s the things the models can’t do. And it’s actually much more valuable. So I think you’ll actually really nice like teasing a part of like where the value actually is because I think our bet is in the next 10 years you’re not going able to replace the work that these partners do in orchestrating these processes and doing all of that. And I think what you’ll see is their time will actually become more valuable because these systems will be much more leverage.

WW (23:51.47)

And that’s even worth more. Yeah, yeah, I think that’s right.

Richard Tromans (24:17.027)

Yeah, I’ve always felt that my gut instinct is that the top partners will be making 10 times what they do today.

WW (24:22.49)

I think that’s right. And we already see this happening, right? This is happening before AI. Like this was already becoming the trend. And like we’re, you at this point, we are a consumer of legal services, like a pretty large one actually. And I think like one of the things that I’ve felt, at least when I receive a bill is the Delta between what it costs for an associate to do a task versus a partner is actually too low.

Richard Tromans (24:34.095)

Hmm.

Richard Tromans (24:44.591)

Mm-hmm.

WW (24:47.12)

So what I mean is, if you’re reviewing XYZ terms for so many hours and that associate is billing $1,200 an hour, and then the partner who is actually like giving you the massive insight on how the industry works and things like that is $2,500 an hour. It actually doesn’t make sense to me. Like I’d be, the associate part is not as great. I would rather pay the partner like a significantly more amount of money. And I think what you get from that would be, that would be very good for the industry is that kind of forces law firms to focus on developing partners more.

And I think that’s actually like, you basically get this incentive alignment where because your best partners are actually like worth so much to the business, it is even a larger incentive on the firm to basically train folks to become partner as fast as humanly possible. And that’s like the best thing for the industry, right? Like, I mean, I was an associate three years ago. And if we can basically create an industry where the entire goal of a law firm is to create partners,

That’s literally what their goal is, to create partners to service clients best. That makes it so it is the most vibrant industry. That is the industry that everyone wants to go into. That makes it closer to TV shows versions of lawyers, which is a lot of why people go to law, let’s be honest. I think that would be such a net positive for the industry.

Richard Tromans (26:13.679)

No, totally. And you could argue, well, I’ve got two analogies came to mind. Number one were special forces where at least in the UK, everyone is either an officer or non-commissioned officer. don’t have, as far as I understand, let’s say change of roles, but you don’t have privates in the SAS, right? You’re hired into those roles or chosen into those roles because you’re considered to have a high level of understanding and autonomy.

WW (26:19.984)

Yeah.

Richard Tromans (26:38.689)

and independence and you can make very, important decisions by yourself. There’s no point having your private standing around going, I don’t know what to do on these complex missions. I I was also thinking about barristers and there are junior barristers in the UK, they do get trained, but the goal of every barristers chambers is to have a whole bunch of senior barristers. Like in gaming terms, you might say like every single unit on your side is like an elite unit.

WW (26:59.269)

Yeah.

WW (27:09.476)

Yeah, you want you want it to be more like elves. feel like elves and games are always the race that has like the most like you don’t have as many units, but they’re all very, very strong. And so maybe that’s the takeaway is we want to turn all law firms into elves.

Richard Tromans (27:18.167)

Exactly!

Exactly. you just, mean, I mean, this is funny thing about law firms now, isn’t it? It’s that you’ve just got so many foot soldiers and they’re very, very, very, very hardworking, very bright people. They wouldn’t be there otherwise. There’s always. Yeah.

Gabe (27:24.301)

Okay.

WW (27:29.828)

Yeah. Yeah.

WW (27:34.99)

But it’s damaging to the foot soldiers. I think the point is, if you think about, I’ve said this before, but if you think about 20, 30 years ago, sorry, even 10, most folks, it’s like, what did you want to grow up to be? Lawyer or doctor, right? And I think that has somewhat dissipated. And I do think we have, as an industry, we have a little bit of a problem with talent attraction right now.

Richard Tromans (27:46.319)

Hmm.

Richard Tromans (27:50.681)

Hmm.

WW (27:59.6)

and I think we will in the next 10 years. And I think if it looks a lot more like Navy SEALs, et cetera, special forces, I think you’re gonna just attract more people into the industry, which is fantastic. And I think this is gonna happen in other industries too. And I think that’s a net good.

Richard Tromans (28:06.255)

This is. Fantastic. Wow, this could keep going forever. We’ve really struck on a topic I love as well. But we got to leave it there, unfortunately. But we’re to have to have you back on more punks and on artificial intelligence. Thanks, Gabe. Thanks, Winston. And see you again soon.

WW (28:31.662)

Thanks so much.

—