What happens when generative AI becomes core infrastructure for law firms rather than just a productivity tool? In this episode, hosts Jen Leonard and Bridget McCormack are joined by Gabe Pereyra, President and Co-Founder of Harvey, one of the most influential legal AI platforms in the world.
Gabe explains how leading law firms are using AI to rethink profitability, training, and scale—without abandoning professional judgment or trust. The conversation explores the future of the billable hour, supervising AI-generated work, building trust in AI systems, and how lawyers can be trained faster using AI-powered tools. Together, they unpack what it really means to practice law in an AI-first era.
Key Takeaways
- The Billable Hour Evolves: AI optimizes legal work without eliminating traditional models.
- Trust Through Governance: Human-in-the-loop systems remain essential.
- Faster Lawyer Training: AI can accelerate skill development and learning curves.
- Scaling Expertise: AI enables firms to grow while maintaining quality and judgment.
Final Thoughts
This episode underscores a critical shift in legal practice: the future of law is not about replacing lawyers, but about redefining how expertise, judgment, and scale work together. As AI becomes embedded in legal workflows, lawyers who learn to supervise, orchestrate, and collaborate with AI systems will define the next era of the profession.
For law firm leaders, the takeaway is clear—AI fluency is now a core professional skill.
Transcript
Jen Leonard: Hi everyone, and welcome back to AI and the Future of Law, the podcast where we explore all of the interesting developments in the world of artificial intelligence and how they impact the legal profession. I’m your co-host, Jen Leonard, founder of Creative Lawyers, and joined, as always, with the wonderful Bridget McCormack, president and CEO of the American Arbitration Association.
We are thrilled and delighted to be joined today by Gabe Pereyra, who is president and co-founder of Harvey. Welcome, Gabe. Thanks for joining us today.
Gabe Pereyra: Thanks so much for having me.
Jen Leonard: Thank you so much. We know you’re a busy person. Thanks for making the time to come chat with us about Harvey and the future of the legal profession. Before we dive in, we have an opening segment.
One of the things that people tell us about the podcast is that it inspires them to be creative and push the boundaries of how they’re using AI in their lives.
So we like to ask our guests for an AI Aha!—something maybe unexpected or delightful—that they’re using AI for. So we would invite you to share your AI Aha! with our audience.
AI Aha! Moment
Gabe Pereyra: Yeah—love this question.
I have something I’ve started doing kind of a lot in my personal life. I mean, obviously I use ChatGPT and all these products a ton. But I found kind of one thing that’s missing from them is it’s still quite hard to get context into them.
And so one example that I’ve started doing is I have a long Google Doc where I log all my workouts and kind of how I’m feeling. I’m recovering from knee surgery, so I’m doing PT. And I found it super useful where, like last year, I did this every day of the year.
And then I can take that whole doc and put it in ChatGPT and be like, “What workout should I do?”
And these models are getting so good that it’ll be like, “You’re doing too much passive stretching. You’re not doing this one PT exercise. You’re overdoing it this day, which led to your knee being sore the next day.”
And then it can recommend workouts. And so every day I’ll be like, “What workout should I do based on all these factors?” And so it’s like a running workout log and journal. I found that to be incredibly useful.
And then you can do this in a bunch of different areas, like for your diet. And then obviously in the company for one-on-ones and all these things. But I found that was kind of a really useful use case that’s kind of like a superpowered version of the ChatGPT memory.
Jen Leonard: Very cool. Bridget uses it a lot for similar things.
Bridget McCormack: I do, actually. I’m doing something quite similar. Lately my husband and I have been doing our fitness and nutrition. We’re both having big birthdays coming up, and we’re trying to figure out the best fitness and nutrition routines.
And it’s so helpful because once it has all this data about you and information about you—and we also upload all of our health documents because we want all of the benefit of all of the feedback—it’s really incredible.
It’s like having the best personal coach in the world, but also really kind and gentle when, like, you know, my workout’s kind of lame. It’s like, “It’s okay, I’m going to do better tomorrow.” So that’s very cool.
Main Topic: Interview with Gabe Pereyra
Bridget McCormack: It is great to meet you, Gabe. Thanks so much for coming on the podcast. We’re really excited to talk to you.
We’ve been following Harvey and really interested in learning more about you and how you ended up leading what I think is the buzziest legal tech company ever.
And you’re not a lawyer. You don’t have a legal background, as far as I can tell. How did you end up leading this pretty impressive legal tech company?
Gabe Pereyra: My roommate—and now kind of co-founder and CEO of Harvey—Winston, was a lawyer. And I think before starting Harvey, kind of for maybe the decade before it, I had done AI research.
But the goal of doing AI research was always to start an AI company. And I think I was relatively agnostic of the domain. For me, it was kind of, “I think this technology is going to be really important. Where can you apply it where it will have a very big impact?”
And I think before Harvey, I thought maybe it would be personalized education or something else.
And then when Winston and I were living together, I was working at Meta on their large language model team. Winston was doing antitrust and securities litigation at O'Melveny. I showed him the models. He started showing me some of his workflows—kind of the tools he was using.
And I think that was the light bulb moment.
Since that, I think it’s just been this perfect application of this technology. I would say maybe the only other place that it’s an even better application is coding. But I would say coding and legal have just been these magical domains for applying large language models.
Bridget McCormack: And I feel like it’s maybe ahead in coding because the people who are building the models—that’s the most important use case they imagine.
And when everybody turns their attention to legal, there’s just as much upside. And so it sounds like you had, through Winston, a window into just how ripe for disruption the legal industry is.
And I know people say that about lots of industries, but I think they usually mean you have to climb up the tree and find some fruit on the higher branches. But in legal, I feel like it’s kind of just like a piñata on the floor that people have been walking over for like 200 years.
And so—so much upside potential. Well, we’re glad you found your way.
In case there are folks in our audience—probably not many—who haven’t heard of Harvey or don’t know sort of what it does, can you just give us an overview? Describe, in a nutshell, what is Harvey and what makes it different from other AI solutions within legal.
Gabe Pereyra: The way I would think of Harvey now is we are building a platform that lets large law firms—and now actually law firms of any size—or in-house legal teams deploy generative AI at scale.
And I think a lot of the problems we’re solving are when you’re deploying this to thousands of lawyers who are working with tens of thousands of clients on secure client data. A lot of the product and platform problems you need to solve are not just, “How do you build a good product for an individual lawyer?”
It’s all of the security, data, privacy, governance, ethical walls—all of the regulatory challenges—everything you need so that a large law firm can use this with confidence.
And the way the product started was, I think, how a lot of these AI products started, which is: how do you build a really good productivity tool for an individual lawyer?
That was kind of three years ago when we started the company—three and a half years. We were really focused on taking these models, giving them to lawyers, and making them useful.
And a lot of the problems we were solving were: the models don’t have access to case law.
They don’t have access to the internal documents of a law firm.
And once we solved kind of all these immediate problems, then you got into the much more interesting problems, which are: I want to collaborate with the rest of the lawyers on my deal team. I want to collaborate with my clients.
And when you think about that at a large law firm, a lot of the problems are: I don’t have just one deal team. I have a hundred deal teams. I have a thousand deals going on simultaneously. How do I track all of these?
And then I think what’s starting to be really interesting is the real value of what we’re building is not productivity. It’s helping these law firms think about how they build the future of law firms.
And a lot of the value of the product is going to be: how do you make these law firms more profitable, not just how do you make individual lawyers more productive?
And then similarly for in-house teams—for these massive Fortune 500 in-house teams—a lot of their challenges are: how do you deploy this technology at scale, and how do you get better, faster, more robust legal outcomes at scale?
And so when we think about the product, that’s a lot of the problems we’re solving.
Jen Leonard: Gabe, the question of how do you make the law firm more profitable is the perfect segue to the biggest question facing law firm leaders in the face of generative AI.
Everyone sort of recognizes that efficiency is the enemy of the billable hour, which is one of the two pillars of success of the modern law firm. And people have been talking about this since time immemorial.
But I remember the 2008 Great Financial Crisis, when the leverage shifted to the clients and the rise of legal operations and automation and electronically stored information and all of this. And people predicted the demise of the billable hour—but we still predominantly price legal services this way.
But now we’re really thinking about this in a new way. Bridget and I present to lawyers all the time, and it’s very different now—the way that partners are responding to generative AI.
Do you really anticipate law firms changing their fee structures? And if so, what do you think the time horizon is for that change to happen? And what will it look like, do you think, on the other side?
Gabe Pereyra: Yeah. I’m actually somewhat skeptical that this is the death of the billable hour.
I think it gets a bad rap because people focus on the obvious—like the incentives are somewhat misaligned, and this isn’t the best way to build value. But I think a lot of people miss why this system is so good.
The biggest problem the billable hour solves is: here is a super standardized way to price services that are very difficult to price.
And when we talk with a lot of these large law firms and partners, for example, an issue they usually run into with fixed fees is they’ll go in and say, “Hey, we can actually do this—just pay us a million dollars upfront.”
And then the client will be like, “Actually, let’s just bill it and see how it goes.”
And I think the issue you run into—and this is what we’re starting to talk with firms a lot about—is every practice area at a law firm is almost its own different business. The way you run fund formation is very different than how you run transactions or mergers, which is very different than how you run litigation.
And so figuring out a general way of pricing that is better than the billable hour, that works for all of these, is incredibly challenging. I’m not sure that’s what completely happens.
Because if you think about what you would need to do, you’d need some way to go into any practice area and say, “This is how much this is worth to the client.” And the reason legal is so challenging is that’s quite hard to do.
What we are starting to see is just hybrid billing.
And so what we’re seeing, for example, is on an acquisition or a merger, you can say, “We’ll do the diligence fixed fee,” and then it’ll be billable for other parts.
But when I think about how these law firms are going to get more profitable, I think there is actually a huge amount of optimization you can do within the billable hour framework.
So, for example, when we talk with large private equity firms and their law firms about fund formation, in a lot of cases all the first-year associate work gets written off just right off the bat. It doesn’t even matter what work they do.
The private equity firms will say, “That work—I’m not going to pay for it.”
And so if you can go and say, “Well, here’s an agent or here’s a system that will do all of that work,” you can reallocate that associate time. That is actually making you more profitable.
Some of these projects are capped. And so if you go over the cap, you’re losing money.
The realization rates aren’t 100% because you argue with the clients and they’ll say, “I’m not going to pay for that amount of legal research.”
And so there’s actually this huge amount of wiggle room within that framework.
And so I think the big challenge—and this is what we’re working with a bunch of firms on—is what is the right business model for every practice area.
And I think where it’ll be super interesting is you’re going to see a bunch of different solutions depending on the law firm size, the industry, the country, the region, the clients.
And so I think it’ll be interesting.
To me, the question is: what part of this can you turn into software? What stays billable hour? What makes sense as fixed fee?
But I don’t think there’s going to be one simple answer to this, because legal work is just way too complicated.
Jen Leonard: Do you have a sense—in practice areas where clients refuse to pay for first-year associate work across the board—do the partners have a sense of how those associates would get trained in those practice areas?
If it would be taken over by an agent, for example, or software—what are they thinking there?
Gabe Pereyra: This is something we’re thinking a lot about.
We get the general version of this question a lot, which is: as these AI systems get better, how do we train the future generation of partners?
I think the part where I’m super optimistic is that typically the work getting written off is the lowest value in terms of educational value.
The work getting written off is the work where clients are saying, “I don’t even need this associate that went to Harvard Law to do this task. I shouldn’t be paying that rate.”
And it isn’t the type of task that, when you become a partner, you look back on and say, “Oh, that was the thing that really trained me to be this very valuable partner.”
And so we already see this with our engineering team, where increasingly these models are getting better at coding. But we’re seeing the opposite of what people expect.
It’s not the case that our engineers are not able to grow and learn. They’re actually learning faster. They’re able to work at a higher level of abstraction and take on more complex projects because they’re able to delegate a lot of this lower-level work to the models.
And the skill that’s really valuable for these engineers—and it’s the same for partners—is not doing the day-to-day work.
Most partners aren’t doing the day-to-day work. Their value is they know how to orchestrate a team of associates to do the work. They know when to review it.
That’s the same thing we’re seeing with our engineers, where they’re able to delegate the low-level work that they would normally have to spend a bunch of hours on to these agents.
But the skill they’re developing is: how do I coordinate a team to do that work?
And so I think what’s actually going to happen is you’re going to be able to train associates sooner, and you’re going to be able to push associates up into doing the things that humans can uniquely do.
And that’s what makes you a valuable senior associate or partner.
Jen Leonard: You know, Harvey is valued now at an $8 billion valuation. So it’s being priced more as a labor substitute than just a mere software tool.
So how should a law firm managing partner be thinking about the way that law firm hiring happens? Firms are hiring for associates sometimes—like banking and clerkships—sometimes three or four years before they need the associates to be doing actual work.
So how should law firm managing partners be thinking, if they’re saying, “We don’t necessarily need the Harvard Law grad to be doing this work,” on that continuum, when the technology is changing so rapidly?
Gabe Pereyra: One point I would make is it’s not clear to me that our valuation implies labor substitution.
There are software companies that are significantly larger than us that sell software that aren’t labor substitution. Salesforce, for example, is a multiple-hundred-billion-dollar company, and I don’t think people view it as labor substitution.
It is a tool that massively augments your sales team.
I would actually say that’s a good analogy for how to think about Harvey and the software we’re building, where it will create a huge amount of leverage for these law firms.
And so I think the same way that before technology like Salesforce, there was just a limit to how big a sales organization you were able to build—because you weren’t able to coordinate all of the humans to work—and Salesforce helped solve that problem.
I would think about that as similar to what we’re doing.
To me, the question isn’t, “What are these associates going to do?” It’s: how do they work in this new way so they can use this technology to get more leverage?
I expect what you’ll see is that there will actually be much larger law firms, because I think this technology will let you scale law firms much more the same way Salesforce did with sales organizations.
But I think a lot of the challenge is going to be that you need to learn to work in this new way, more than this being a simple swap-in replacement for any type of labor.
Bridget McCormack: Let me pivot to how you foresee the legal profession growing its trust in the products that you’re building and others are building.
Just by way of background, at the AAA we have built the first AI arbitrator. It’s really twenty or thirty—or forty, depending on how complicated the case—different agents that operate across an arbitration process.
It’s trained on a lot of data. It’s deployed for one very specific use case to start with.
And, you know, in my view, the question is not whether the technology can do an excellent job. It can—especially a documents-only dispute that’s pretty contained to the type of data it’s been trained on.
But the question is: when will lawyers trust it?
And I assume you must confront that as well. Do you have thoughts about the trust cycle—where we’re sort of starting out—and when we get to a place where we’re not talking about hallucinations anymore?
Lawyers are talking less about them now, but for a long time that was their first response to why they didn’t have to learn about this technology.
Do you feel we’re doing well on trust? And do you have any predictions on when people will trust AI dispute resolution—or when trust won’t be a barrier for adopting some of the products that you’re building?
Gabe Pereyra: That’s a great question.
We’ve definitely talked with some folks about arbitration and things like that. I think one thing I would separate is that, to me, there are two ways of using AI that imply very different levels of trust.
One—which is kind of how Harvey is used—is human-in-the-loop, the same way that a partner would work with a team of associates.
The partner is never directly sending associate work to the client without checking, and I would still think very much that that is how Harvey is used.
It helps these large law firms or in-house teams do some percentage of the work, but the work of Harvey is always tracked. In the same way, you reduce hallucinations with Harvey is the same way you make sure your associates or your team don’t make mistakes.
And I would say in that sense, actually, in the past year, we haven’t really had hallucinations be a blocker for using this technology.
Now it’s become the same as: I need to structure my law firm so that mistakes don’t get to the client—whether those mistakes are hallucinations or human mistakes, it kind of doesn’t matter.
And so I would say we’ve actually—not we as in Harvey, but we as an industry—largely solved that trust problem.
Now all of the debates are about how do we add these systems better, not should we use them or not.
And then if I understand correctly what you’re talking about—which I think is a much higher level of trust—is how do we directly trust the legal output of these systems.
And that, I think, we’re not there. That’s not quite the problem we’re solving for, which I think is a much harder problem to solve.
But to me, in the long term, it’s the same way you learn to trust self-driving cars, right? At the end of the day, the way you trust a self-driving car is just its track record.
These self-driving cars have now driven millions of miles. You can look at statistical results of their performance, and at some point you can build confidence and say, “For this set of tasks, this performs on par with a human, and I’m comfortable with this making a binding decision.”
But I think that’s a really interesting problem. It’s kind of not what we’re focused on.
We’re much more focused on how do we augment humans with these technologies. But I think there will be low-level work—like, for example, NDAs—where the systems in the next couple of years will get to the point where you’ll probably just be able to say, “I want to enter into an NDA with this company.”
It’ll take all your historical data and their historical data and years of precedent, and it’ll be like, “We’ve generated this. You’re both bound by it.”
And it’ll be good enough that you don’t even need to check it.
I think for very complex work—that’s what most of the law firms we work with focus on—I don’t see a world, even in the next decade, where for an antitrust litigation or a very complex acquisition you’re able to trust these systems end to end.
Bridget McCormack: I should be clear—our generative AI arbitrator does have a human in the loop, because I think that is important.
And I kind of agree with you in both examples that it comes down to governance. If the agents you’ve built are auditable and you have a trail that you can show—a record you can show—I think that is how we get there.
But it’s going to take a little while.
Gabe Pereyra: I strongly agree with the auditability and explainability point.
This is something where we’ve actually learned so much from lawyers and law firms. If you think about all of the expertise they’ve built around doing complex arbitration, mergers, litigation—how do I make sure no mistakes go through?
They’ve built these incredible systems of: this is how you do citations, this is how you show your work, this is how you decompose the work product.
A lot of how we’re building these systems is inspired by that. And I think that’s going to be a really important thing going forward.
Bridget McCormack: What about some of the other macro changes we’re seeing to the profession and how they are either impacted by AI—or not?
So, for example, we’re seeing regulatory shifts in states that are interested in new models of legal businesses. That wasn’t a thing a few years ago. Now it is, and I think we’re going to see it in more states across the country.
The interest from private equity in law firms is stunning lately. I get a lot of calls from people trying to understand things about the market that weren’t happening even a year ago.
What I’m seeing right now—and it does feel to me like we’re seeing a big shift in how the business of law operates—is that it’s a result of both of those forces, and maybe some others.
AI, to me, just accelerates that change and shift. But I’m interested in your thoughts, because I’m sure you’ve given all of this a lot of thought.
Gabe Pereyra: I think that’s exactly right.
Our general sense is that most of these trends were already happening before generative AI got very popular, and then, in different aspects, AI accelerates them greatly.
Even products like ChatGPT raise some interesting questions. When we started the company, we actually started looking at consumer law, and we wanted to use generative AI to massively expand legal access.
We talked with someone who represented LegalZoom when that case went to the Supreme Court. Then we did this deep dive into all of the regulatory issues in the U.S., and the conclusion was: you can’t use these models to give legal advice.
So you can’t actually build a product and say, “Hey, you’re getting evicted. Here’s a product that will help you figure out what to do in this legal situation.”
And now with ChatGPT, this is kind of happening ad hoc. There are pros and cons to that, and I think it’ll be really interesting to see how this gets regulated.
To your point, this is starting to change. There’s Arizona, Utah—I think there are some states where they’re starting to allow some forms of this.
I think this is something where, in the future, we want to use what we’re building to help consumers. But right now we’re still much more focused on law firms and enterprises.
That feels like the right place to play, because you sidestep a lot of these regulatory issues. And the best place to work through this is with law firms themselves.
On the private equity side, we’re starting to work with a lot of private equity firms. I talked with a personal injury firm just yesterday that has a private equity backer, and they’re looking to scale really aggressively.
So I think, tied to the regulatory issues—especially in the U.S.—there are a lot of things startups or tech companies have been able to do that law firms have historically been prohibited from doing because of regulation.
You can’t invest outside capital into a law firm in the U.S., for example.
I think it’ll be interesting to see how those things change.
And similar to my point earlier about business models, legal is just so complex. There’s so much surface area.
It’ll be really interesting to see how regulation plays out in different countries, how private equity thinks about this. There are so many different angles—consumer legal services will look different, enterprise services will look different.
There are companies trying to build AI-first law firms. I think it’s good to see all of this innovation.
At the end of the day, the goal—whether you’re providing services to clients or to companies—is how do you use this technology and all the things around it to provide better services.
And so I think it’s exciting to see how much opportunity there is.
Jen Leonard: Gabe, our team works with a lot of different law firms of all different sizes and practice areas.
And they’re all grappling with these issues—supervising AI-generated work product, training new lawyers, thinking about pricing models, talking with clients.
Some of the early guidance emphasizes the duty to supervise both the attorneys you’re working with and the outputs of the models.
Really sophisticated lawyers who understand the technology may be ahead of the game, but a lot of lawyers are struggling because they’re trying to understand what the technology even is—and then it changes so quickly and gets better so quickly.
You started to touch on this a little bit in your exchange with Bridget, but as the technology sort of approaches legal general intelligence or legal superintelligence, how does a lawyer reasonably supervise the AI and its output?
How would you encourage lawyers to think about that—to stay abreast of how good the technology is, where its limitations are, its jaggedness—and still continue to be expert in what they’re substantively expert in at the same time?
Gabe Pereyra: I think the analogy that works really well for a lot of this technology is thinking about how you do this with humans.
If you think about how law firms solve this problem now—if I’m a firm working on a really complex merger—typically the partner running that deal will be working with other partners who are more specialized than they are.
But you’re still able to figure out ways to coordinate those parties and get them to work together.
We get questions like, “These models are a black box. How do you evaluate them?” And my counter to that is: so are humans. We don’t really know how the human brain works.
Partners are able to do these really complex things, but you don’t really see that much of what they’re doing—besides the fact that they produce a document and can explain it.
These models can do that as well.
What we’ve seen law firms do really well is take all of the expertise and systems they’ve built to coordinate teams of lawyers and start thinking about what parts of that they can repurpose.
I think a lot of the work for law firms and foundation model providers will be: how do we continue to build these systems in a way that’s explainable?
The thing that’s really nice about legal is the work product itself is language.
It’s not like other domains—programming, mathematics—where you could imagine writing software or theorems that are so complex no one can understand them.
At the end of the day, the language you produce in legal needs to be understandable by the counterparty or the court.
So there’s this built-in constraint that even the best systems will need to be understood by a judge or by the other party you’re negotiating with.
And downstream from that, the same way you check the work of junior partners and senior associates, there’s a lot of that structure you can leverage.
In the longer term, you might get systems that can do certain processes end to end—like automatically negotiating NDAs.
But I would say, as a society, we already have forms of superintelligence in different domains.
The best example is the stock market. It’s a system that can price assets better than any human. And yet we’re still able to regulate it, understand it to some degree, and control it.
So I think there will be parallels there as these systems get better at legal reasoning.
But for legal, we’re still a ways off from that.
Jen Leonard: Bridget mentioned hallucinations earlier—the obvious hallucinations of cases and quotations that raised early concerns.
Now some of the conversations we’re having are about reasoning hallucinations and whether the juice is worth the squeeze if a partner has to read an entire case just to make sure the AI accurately summarized what the case stands for.
So for a non-technical audience, could you describe whether and how AI tools are mitigating some of those reasoning hallucinations?
Because of their language-based nature—and because practicing law is fundamentally about arguing over what language in a case means—this can become a conundrum for lawyers working with large language models.
Gabe Pereyra: That’s a super interesting problem.
For a simplified explanation, the reason you get hallucinations is that if you use language models out of the box, they’re not connected to data.
They’re trained on a large corpus of text, but they may be missing important information. So they’ll produce an answer that sounds confident, but it’s wrong.
As the models get better, the reasoning improves, but they still make logical mistakes.
What’s concerning is that the way they make mistakes is non-intuitive to humans. They don’t make the same logical errors that humans do.
If you think about how partners solve this with associates, the best partners have worked with associates for decades.
They’ve built a mental model: this associate doesn’t understand this concept, these are the mistakes they tend to make.
People are still building that mental model with language models.
One thing that’s been really promising—and honestly pretty amazing—is the lawyers we’ve hired on our product team.
The first was actually my brother, who worked at O’Melveny. We’ve since hired about twenty other lawyers to help build the product.
If you talk to them, their mental model of the reasoning errors these systems make is almost perfect.
They’re not surprised by the errors anymore. They’ll say, “If you use this prompt, it’s going to make this kind of mistake in these situations, for this reason.”
You can develop that same mental model—but it takes training.
And to me, that’s the biggest solution. It’s less a technology problem and more about learning how to use the technology.
It’s the same way we learned to use Google search.
Before, you’d go to a website and it might be wrong. Over time, people learned: this is Wikipedia, this is the New York Times, this is a random blog.
You develop instincts for credibility. That’s what’s happening here.
But then on the product and technology side, a lot of what we’re doing is changing how these systems reason. The way people train these systems now is that they’re starting to learn to reason more like humans.
So what works really well is, if you want to do a task like, “Hey, can you write me a research memo on the pass-through defense,” or defeat class certification, or write some complicated litigation topic—if you just ask the model, it’ll kind of come up with some knowledge it scraped from the internet.
But if you look at how an associate would actually do this, they would Google the topic. They would find the seminal cases. They would read those cases and summarize them.
Then from those cases, they would find supporting cases. They would look at their client and say, “Okay, I need to find other clients similar to this.”
They would use Westlaw or Lexis to do that search. You’d see them go through all of these steps to find the information, and they would link to where they found it from credible sources.
That’s what these reasoning models or agents are doing.
So a lot of what we’re building is: how do we build systems that reason like a lawyer?
There are two benefits to this.
One is that the models start to think more like humans. When you mimic how humans solve these problems, we—and not just us, but the foundation model companies—have found that this is just the right way to solve them.
The second benefit is explainability.
Now when you get a mistake, it’s not just, “Oh, it generated a bunch of text and I don’t know what happened.”
You can say, “Here is the reasoning trace.”
You can see, “Oh, it pulled the wrong case,” or “It summarized the case incorrectly.” Now you understand why it came to the wrong conclusion.
That’s exactly what partners do with associates.
They’ll say, “You looked at this topic, but you didn’t find the seminal case. Everything downstream is wrong because this case says something critically important.”
As these models get better, there’s a lot of this that can continue to improve and be solved.
I think one of the biggest shifts with this technology—not just for lawyers—is that early on there was this intuition that it needed to be 100% correct to be useful.
People have moved past that.
I use this for my writing all the time. Even though it’s not perfect, it’s like a 60–80% draft. It gives you structure, which is incredibly valuable.
The reason law firms and engineering teams are structured the way they are—where junior people do the drafting and senior people review—is because it’s exponentially less effort to review work than to generate it from scratch.
What we’re seeing partners do now is give a task to an associate and give the same task to the model.
They can review both quickly.
In some cases, they’ll look at it and say, “That’s correct. I already knew the answer, but I needed someone to collect the information and structure it.”
In other cases, they’ll say, “That’s way off. That’s the work I should delegate.”
And I think that will continue to improve.
Bridget McCormack: What if it improves so much that we don’t really need lawyers anymore?
What’s your advice to lawyers who have thought of their value proposition as drafting beautiful motions and briefs—activities that the models are going to be able to do better than lawyers going forward?
Is there still a need for lawyer supervision when there’s no way they can really compete with what the models do?
Is it like self-driving cars, where in twenty years we’re all going to think it was outrageous that we let humans drive them?
Are we also going to think in twenty years that it’s unbelievable we let one human with one little tiny brain try to work their way through a complicated legal problem?
Gabe Pereyra: Yeah.
The way I think about it—both for engineering and for lawyers—is that people overly associate the job with producing text.
As you get more senior as a lawyer or a software engineer, none of your value comes from writing code. None of your value as a partner comes from writing documents.
For us, as a software company, we’re not bottlenecked by junior engineers. We’re bottlenecked by very senior people who can create abstractions, come up with product ideas, scale complex systems, and understand very complex business problems.
I think that’s going to continue to be the big gap that these models won’t be able to close.
On the legal side, the value of a partner is not that they can write the best briefs. They do have deep technical expertise and they supervise teams—but a lot of their value is understanding the client, understanding the business implications of decisions, and understanding trade-offs.
I think that will continue to be the case.
Even as these systems get better—and to our earlier point—you can automate things like NDAs or arbitration, but you still need to design the systems that govern those processes.
So I think about automation differently. There are ceilings for different domains.
With something like driving, once you build a self-driving car that’s good enough, there’s nowhere else to go. That task is done.
Programming and legal are very different. I think they have infinite ceilings.
There is always more code to write. Over the last few decades, we keep inventing new programming languages and new ways of writing software.
Even if the models get very good, that doesn’t stop.
It’s the same with mathematics—you just go to higher and higher levels of abstraction.
Legal is the same. We’ve barely scratched the surface of how to build legal systems.
It’s easy to get stuck thinking, “I’m a junior person, I need to write NDAs or LPAs.”
But the real goal is: how do we elevate people to think about building better legal systems and creating better outcomes?
I think this technology will keep pushing people to higher levels of abstraction. I don’t think there’s a cap in domains like math, programming, or legal—these are fundamentally complex systems.
Jen Leonard: I totally agree, Gabe, about the senior lawyers—I just wanted to go back for a minute to those junior lawyers, whose development now is sort of commercially supported by the business model.
I read, I think when we were researching for the interview, that you’ve talked about flight simulators that are used to train pilots as a model we can look to in order to turbocharge the way we train associates.
So could you talk a little bit about how you envision that sort of replacing the model that we have that takes years—and that I think Bridget and I would both say is not an ideal model anyway? What do you think that would look like?
Gabe Pereyra: Yeah.
Already people are using our product—and these AI products more generally—as personalized tutors. I think these tools are incredible.
I use our product, ChatGPT, all the time to ask questions about how to do my job better. When
I’m programming, I’ll ask, “How would I do this in this programming language?”
We see lawyers using our product in the same way—“I’m working on this fund formation. I need to draft this LPA. How do you typically structure it for funds of this type?”
As the models get better, they’re going to get better at teaching you how to do the things they can do. You can get to the edge of the model’s knowledge much faster.
The next big opportunity is that this works really well for coding because most coding knowledge is public. It works less well for legal because most legal expertise lives inside law firms.
So what we want to help firms do is take all of that internal expertise and build systems that are unique to them—systems they can use not just to do legal work better, but to distill the firm’s expertise into their more junior associates.
The first step is putting all the feedback, all the practice guides, all of that institutional knowledge into these systems so they can be used to train associates.
The flight simulator idea is the next step.
A lot of the foundation model companies are thinking about what are called reinforcement learning environments—RL environments.
In software, that looks like a code base with unit tests. An agent writes code, runs it, deploys it, gets feedback—basically a simulated environment that mirrors real-world work.
For legal, the equivalent is a client matter.
You could say, “I want to run a simulated fund formation, a simulated litigation, a simulated merger,” and go through the process end to end, getting feedback from a simulated client and simulated partners.
What’s interesting is that this infrastructure trains agents—but it also trains humans.
These two things converge.
I know there are already people thinking about this—we’ve talked with folks at Stanford and elsewhere.
I think we’ll see powerful educational tools that dramatically accelerate learning.
More broadly, my intuition is that in a lot of these jobs, you learn a lot the first few times you do something. But over ten years, there’s a lot of repetition that isn’t high-value learning.
I’m very optimistic—not just for legal, but for all knowledge work—that we’ll be able to train humans much more efficiently and get them to the edge of what AI can do faster, so they can focus on what humans uniquely do.
Bridget McCormack: Jen and I have been kind of surprised by how quickly lawyers have started using generative AI tools.
It’s a small-c conservative profession. We’re trained to look for risk everywhere. And yet there does seem to be a transition happening—even at the individual lawyer level—that surprises both of us.
In 2023, at least, it was a long slog in presentations. But now we’re seeing people adopt pretty quickly.
You talk to lots of lawyers. What do you think explains this faster transition compared to other technologies?
And within the profession, what distinguishes lawyers who successfully adopt AI from those who don’t?
Gabe Pereyra: The simple answer is that this is the first time the technology has been centered around language.
Before this, legal technology wasn’t language-based. You had tools that didn’t fit your workflow, and you had to force your work into them.
With generative AI, it’s all language.
That’s why legal and programming have seen such rapid adoption.
Programming went through something similar. Before generative AI, the main tool was an IDE—it was basically a word processor with autocomplete.
Now you can ask questions and generate entire documents.
One of the most compelling things we saw early on was senior partners using this technology.
There was a perception that this would be useful for junior lawyers, but not seniors. Early on, Gordon Moody—then a senior partner at Wachtell—joined Harvey as an advisor.
He’s one of the top transactional lawyers in the world, and he was already using these models extensively.
When someone with that level of expertise and a deep understanding of the models uses them, what they can do is incredible.
We’ve now seen this across top transactional and litigation firms.
What distinguishes those partners is how much they use the technology.
People sometimes talk about AI as something you “swap in,” but it’s much more like a personal computer or the internet.
Some people are vastly more skilled at using those tools than others, even though we don’t always notice it.
The skill gap between a first-time user and someone who has used these tools for hours a day over multiple years is enormous.
The only way to build that mental model is to use the technology constantly.
You can give people guides and best practices, but it’s like teaching someone to use an iPhone—you only really learn by using it every day.
That’s what this technology will become.
Legal, programming, and other complex language domains are just perfect applications for it.
Jen Leonard: Well, Gabe, thank you so much for answering our questions and walking us through how you’re thinking about the future of the profession.
You certainly arrived at exactly the right moment. We’re living through unprecedented change, and we’re really grateful to you for sharing your perspective.
And thanks to everyone listening for tuning in. We’ll look forward to sharing our thoughts with you on the next episode of AI and the Future of Law. Until then, stay well.