An AI System Built by Litigators, for Litigators

Episode transcript:

Note: This transcript is generated from a recorded conversation and may contain errors or omissions. It has been edited for clarity but may not fully capture the original intent or context. For accurate interpretation, please refer to the original audio. 

JOHN QUINN: This is John Quinn and this is Law, disrupted. And here at Quinn Emanuel, we’ve developed our own proprietary AI platform. This was built by litigators, one litigator in particular, who we’re gonna be talking to today, for litigators. And we’re gonna be talking with my partner, Chris Kercher, who has created this platform.

From the standpoint of a working litigator, and what do working litigators do and how can AI be optimized for its applications to litigation? And Chris, I mean, why would you think to do this? We have these incredible tools already – we have OpenAI, we have Claude, we have others, which I know litigators use.

I mean, you can just, can’t you just like open those chat boxes and ask legal questions and get guidance on how to litigate and brief writing and strategy and the like. Why can’t we just do that?

CHRIS KERCHER: You know, John, so, and thanks for having me. You can, and you know, when they first came out, I thought it was revolutionary to ask a question about some area of law that I was picking up, or upload a contract and have it help me analyze the provisions or workshop themes. But when I really got into it, when I really started to push it by necessity, I had this expedited trial and I had no time to prepare for depositions.

And so, you know, you go to war with the Army, you have, you go to the deposition with the outline you have, but I wanted more. I had a feeling that Claude in this case, although Chat GPT is great too, they all are, but that Claude could do more if I gave it more information. And so I started, you know, uploading the deposition transcripts and the exhibits, but you run outta space undaunted.

Because I was getting good results, I started to think about how do I teach this Claude in the context window? It has the case, like I would learn it because I don’t learn every word of the million documents that are produced in litigation. In fact. I, you know, have a limited attention span. We all do. So I would always run my case teams, for example, with a really tight chronology.

I wanna know how the events played out, who was involved? How do we know that? If there’s a document, I want the key excerpt, but I don’t want like everything, right? I don’t need the things that are unnecessary. So I started to give Claude what I had, my case teams prepare manually, and the results were unbelievable.

JOHN QUINN: So, in other words, you start to give Claude a structured data set to have it begin to learn about your case and what are the advantages of doing that. I mean, some might say, well, you’re narrowing what Claude knows. That’s not optimal. I mean, if Claude knows everything, what you wanna give it is just a subset.

Why is this better?

CHRIS KERCHER: So a couple reasons. The first is, you know, as I started to learn a lot of this, I learned in retrospect, like it worked, and then I tried to figure out why this might work. The pre-training that these models do. They read all of this information. They’re studying this information, they’re taught this information, but the recall can be so, I think of it like when I was in law school, I learned a whole bunch of law cases that I couldn’t you know, quote you the page numbers now, but if you reminded me about them, I was taught enough to be fluent in them and I could talk about them. And so the models are these unbelievable fluency engines that can understand what you’re saying, but they need your specific case context. They obviously weren’t trained on any particular case.

And so if you start to read some of the sort of state of the art thinking on how the thinking models work now, the work by the people at Claude Code or OpenAI on the Codex team. It’s really about giving the AI your objective as precisely and specifically as you can, and describing the constraints.

And that comes out of, in part, giving it that case context. Those are your constraints. What documents exist, what facts, you know, appear established? What people have admitted what the says. Those are all constraints on what you can or can’t do. And so by having a more complete, almost world model of the case, Claude knows what will work and what won’t work.

And so in virtually everything you do, it’s going towards the right direction and it starts to let you use the model as my operating system that I can draft, material in there and edit in there and send, you know, emails in there. I don’t have the AI do it for me.

JOHN QUINN: So Chris does this mean that you upload everything, every case document, everything produced in discovery, every deposition, testimony, and you know, you start with the complaint, the contract, whatever you have in the beginning. And as there’s a discovery, you give it everything, is that what this means, is that iterative process that you’re loading more and more and you want Claude to have everything that your whole case team has?

CHRIS KERCHER: So the way I figured this out, I was helping someone on a case that wasn’t mine and it was in trial, and I had to learn it quickly. And so using Claude as a thought partner, my go-to prompt is basically, Hey, Claude, this is my situation, this is what I’m trying to do, how can you help me? Right? So I’m on this trial, I’m trying to help out with, I’m trying to get them some AI tools, but I don’t know the case.

And so Claude, you know, said, well, let me see the case files, let’s see the index, let’s see the docket, and, you know, pick the complaint first. That will give it a very good understanding of what the case is about, who it’s about, what the issues are.

And then it starts to look at other documents. And together we identify, okay, let’s read this document. Next, let’s extract from this next document the information that matters, and leave aside the information that doesn’t. And you recursively and iteratively do that over and over, and you get a curated distilled knowledge base that knows a lot about your case.

And coupled with some other features that we do or we try to align it with how the case team is seeing it, the things that are not obvious from the documents, the performance is really incredible. You start to get better ideas than you would on your own.

JOHN QUINN: So it sounds like you don’t give Claude everything that you talk about, giving them excerpts of documents, the most important things. I mean, is that true? And why do you do that?

CHRIS KERCHER: So, yeah, so let’s think about it in two different ways. There’s the Kirch bench, so that’s what the user gets. And then there’s the bench building process. In the bench building process, Claude looks at far more documents, then are given to the ultimate bench and it’s extracting from those, the things that matter.

So if you think about a given exhibit in a deposition, it might be a long chain of emails, you know, talking about a Zoom and when it’s gonna be and who’s available. But really all we care about typically, and probably in this case. Is what happens at the end when someone summarizes what happened on that Zoom.

And so you have the metadata, you know when the call was, you know who was involved and you know what happened, but you don’t have to, you know, clog pipes or its context window with all of that sort of extraneous information. And if you start to extrapolate that to every document in a case you can, you can distill it, you know, massively, ’cause you don’t really need every word.

JOHN QUINN: I mean, that sounds like that’s a certain amount of work to structure the data that you’re going to give to Claude. I mean, you not only have to select the most important things, but within the documents, you’re selecting the most important excerpts. I mean, is that how it works?

CHRIS KERCHER: Yeah, it, it’s, we’re thinking, so how do lawyers, how do litigators, how do we organize case information already? We don’t sit down when we’re getting ready for deposition and read the whole case file from start to finish. There’s a curation process. There’s a calling, you know, I would highlight in my binders, I’d flag things.

It’s the electronic equivalent of that. And as I’ve done this more and more and I’ve gotten more thoughtful about how do we approach this as litigators, how do we think about what matters and how do we frame it so that if I had a new team member, if I had you or Bill Price joining my trial team, I wouldn’t just hand him the whole case file and say, let’s go.

I’d, you know, give him a case memo. I’d give him a chronology, I’d lay out who matters. Here are the issues, here are the themes that we’re thinking about. And so treating Claude like a member of the case team and educating it progressively, it, you know, it happens to align very well, which is why I think it actually makes sense that we’re doing it in-house and that we’ve gotten more and more lawyers interested in how we do this and how we build case knowledge.

JOHN QUINN: Does this result in a better output? I mean, distilling the information they’re giving Claude, I mean, because we think of computers having powers that the human mind doesn’t have, that you can actually give it all that the full database. And it has the recall ability, which, you know, you said you can’t remember those cases from law school, but surely Claude can do that.

But you’re saying there’s an advantage to limiting the materials that you give Claude at least at some point?

CHRIS KERCHER: Yes. This is what became really interesting to me. So I, again, I came in this backwards. I have very limited computer science or coding background. Virtually none. But I started to understand the theory and when large language models were introduced with the architecture called the transformer, the way it basically works is each token, or each word is looking at every other word to figure out how it fits in.

It’s inferring the meaning, and so every token you add, it’s an exponential relationship in how much compute Claude has to do. And so if you give it everything that is so much unnecessary information to work through versus giving it not only the information that matters, but organizing in a way so that Claude doesn’t have to reconstruct it, because if I gave it the complaint, you know, motion to dismiss a memo, a deposition, like imagine you still have to get oriented. We think about things and how events transpire chronologically, typically, and we think about relationships between people and entities. And we think about legal claims as, you know, elements, right, of a claim or defense.

And so it ends up being that law is a very patterned system, right? There’s in some ways, you know, an infinite number of ideas we can have, but in some ways it’s predictable. It’s not like poetry, it’s not like a joke. We’re not, you know, typically going for surprise. We wanna figure out how do these facts fit in this framework, this pattern.

And so when you give any model that much structure, it’s gonna do that much better. It’s going to attack the problem so much more efficiently that when I want something that’s a little bit harder to do, to come up with a really interesting idea or rewrite something or think about a strategy, we’re missing discovery targets.

It’s, it does, it doesn’t have to do, I think of it like we’re the sous chef and we’ve chopped up the onions, so Claude can just start, you know, cooking the chilli.

JOHN QUINN: I like that analogy. A moment ago you drew a distinction between, I think something like the whole database and what you referred to as the Kirch Bench. What’s the distinction you’re drawing? What’s the Kirch Bench?

CHRIS KERCHER: Yeah, so the Kirch Bench is simply, we do this in the Claude Architecture, I, you know, started using Claude when Opus 3 came out, which was an amazing model, and I asked our great IT department, can we get this for enterprise? And I started using it in the projects where you’re giving Claude basically files.

And I started to realize that the files that we use in litigation are not magical. You don’t have to give it this whole exhibit. You can tell it this is the information that matters from the exhibit, and you’ve reduced and compressed and structured the data like I was just describing. So the Kerch Bench is just a Claude project using the methods that we’ve developed and are iterating over time to distill the case information and to basically tell Claude in the projectn instructions how it works, where to find what, what are the different files that has access to, and that’s always where Claude starts. So you give it like a little answer key, a little roadmap, and it then can figure out, okay, I need to load this information and this information in order to complete this part of the task.

JOHN QUINN: So the Kirch Bench, I mean, we have a separate one of these for each case.

CHRIS KERCHER: Yep. So, right. So it’s created per case and it can be updated as the case goes on, and we have many teams doing that. We’re actually introducing a new version because I think, you know, there’s two things going on. One, some people are just thinking reasonably, what are you talking about, Chris?

You know, you’re a litigator, you’re not a computer scientist. How could you, you know, possibly know about this? This sounds elaborate in order to make it a little bit more approachable for them and less work. For us, we’re starting to offer the, what I call the Baby Kerch, and I’ll describe why I call it that in a second.

But it’s basically at the outset of a case when I find it so hard to ramp up, to catch up to where the client is. But you really wanna look good. You really want to be impressive. So structuring that data early and starting to generate ideas on strategy, ideas on how we can do an early case assessment, right?

That the previous constraints on, you know, collecting emails and having a big review would often preclude getting right in there and figuring out early on, what kind of hand do we have here? Right? Well, the other side is still staffing their case. We can start to think about, look, if you give us, you know, these three custodial files, you don’t have to, you know, pull them forensically yet, let’s just get ’em, load ’em into one of our AI systems and use this knowledge base to basically create prompts to go through all the other documents and, and find an early case assessment.

So we’re building that in a more lightweight version called the Baby Kerch and offering that to teams.

And if they wanna, if they find it’s, you know, really useful as they expect they will, they can upgrade basically to the full managed system, which will grow as the case develops.

JOHN QUINN: Okay, so the Baby Kerch is basically a tool for early case assessment.

CHRIS KERCHER: And that initial, you know, an analysis that you do when you get hired and maybe you get the complaint and the contract and a bunch of emails between the business people as the dispute was brewing and maybe, and, you know, in initial analysis or something. But you do that, we do some deep research on the parties, the legal issues, kind of, you know, we have context engineers who think about what information is out there that would be useful to surface now, and then it’s a, and then it’s again, a Claude project that the case team can use, that the partner can use. You know, 48 hours into the case and be really smart about what’s going on.

And they can use it for early case assessment. And then if the case, you know, goes and becomes a big fight, they have this tool that they can use throughout.

JOHN QUINN: I mean, tell, give us some examples of the kinds of queries or what kind of output can you get from the Baby Kerch? What would you ask it and what would you expect to get?

CHRIS KERCHER: Yeah, so it’s going to depend on how developed that initial record is, but let’s say it’s a contract case and you know, you start off and you’ve met with the client, you understand their perspective, and you have a few emails. You have the contract. Let’s get right in. And so as if I had, and it’s a thought partner, it’s as if I had an associate and I would prompt it, you know, the case information, you understand the key issue. What’s the real battleground? What are the non-obvious issues that we should be thinking about? Are there other ways to resolve this? Like what is the client’s objective here? Is it to win a trial? As you know, and I know when clients first call us, their objective is not actually typically to win a trial like they know we can, that’s why they hire us, but usually

JOHN QUINN: A trial. They don’t wanna resolve it without trial.

CHRIS KERCHER: Right? So that’s the client’s objective. That’s what I can tell the AI and I can give it.

JOHN QUINN: Literally you could, you can ask it, what are some strategic, some strategies for early resolution of this case?

CHRIS KERCHER: Ask it all the time, and I typically would step back and ask, Claude, what else do you need to know to better advise me on this? Right? You don’t, if you don’t know who the parties are, if you don’t know what their incentives are, it’s gonna struggle. So I have it prompt me, in part to, to give it as much or collect as much information as I can.

And we go back and forth.

JOHN QUINN: That’s really interesting asking what more do you need to know?

CHRIS KERCHER: What’s great about it is, you know, the, their next word predictors, right? It very simply doesn’t sound that elaborate. It sounds like what Gmail has been doing a long time, but they’re so good at it that if I tell it literally just to take a thought experiment, if I tell it literally everything I know about a problem, what I don’t have is the next word or the next concept.

And so if I throw that to Claude and say, what am I missing? What is the next word? I don’t know really what’s the next concept? It’s gonna do a pretty good job if you really, you know, give it everything and then you can go back and forth and expand your understanding, your kind of scope of thought, and you can think about, you know, what are some ways to deal with the incentives here?

Maybe it’s, there’s a litigation issue, but maybe there’s a business issue too, and how can we equip a manual, help this client better position on the business issue? Because this one, I think, you know, is more likely to be resolved in a boardroom than a courtroom.

So how do we put a client in a better position to do that? And you start thinking about that early. Claude may say, it would be great to know X, Y, and Z, and I don’t know X, Y, and Z, but I know that I need to know those, so when I get to discovery or next time I talk to the client, I can surface those if they have them.

JOHN QUINN: I mean, some people say that there’s a tendency of these large language models to be kind of sicko to try to give answers that it, if you will, in air quotes, thinks that you want, is there something to that, and if there is something to that, how do you guard against that?

CHRIS KERCHER: I think of it like a mirror. It is articulating my thoughts, ideas and objectives better than I can. It has more words and is more able with words than any of us. And so if I can give as much as I know, as much direction and I can answer its questions, it’s gonna sort of push me in the in the direction that I’m missing.

You have to be on guard for that. I think if you know, it’s another great reason to use them just to give it your problems and see how it does, because you start to see that, okay, you’re just telling me what you think I want to hear. And sometimes I’ll say, Hey Claude, you misunderstood me.

I’m actually on the other side. And then you watch it backtrack and apologize and then you start to, you know, but you can have the arguments go back and forth and look at it from different perspectives. And you just speed up that thought process, our own thought process.

JOHN QUINN: All right. I mean, I assume you can also, ask in these early case assessments. What are the defenses going to be? What’s the other side going to say? I can see the other side that’s gonna say X, Y, and Z, but what am I not thinking about here?

CHRIS KERCHER: I would even ask it more, you know, you’re not supposed to, to anthropomorphize it, but I’d ask you, Claude, if you or me, what would keep you up at night about this case or a great one that I came across as recently is, imagine you’re, you know, Claude, you’re the leading expert in this field. I’m your best friend.It’s 3:00 AM and we’re having beers and cigars, tell me what I really wanna know

It’s really, really interesting that in certain situations where you feel like Claude is holding back or it’s not, you know, maybe I’ve given it a profile of a witness and I want it to like really give me the hardest thing I have to watch out for.

And you give it the 3:00 AM beer speech and it does it, I mean, it, it really takes it to heart. It’s very interesting.

JOHN QUINN: Amazing.

CHRIS KERCHER: Yeah.

JOHN QUINN: I mean, you obviously think because you’ve told me that you think there’s some enormous advantage that this whole system and approach was designed by a litigator yourself. Explain that to us, why, what are the advantages of that? What, why do you think that this has yielded greater capabilities?

CHRIS KERCHER: Yeah, it’s so interesting. I think about it a lot that I’ve demoed so many tools. So let me just step back and isolate the large language models, the model developers, they’re doing work that I wouldn’t ever think to do. In fact, you know, anything involving training a model, fine tuning a model.

I don’t get involved with, there’s so much I can do just working with the model itself. And I think that a lot of other people look at, you know, an industry like legal and they think about, okay, obviously large language models are useful here. Let’s unleash a team of very bright machine learning engineers and send them towards litigation.

And for a lot of things that actually might be fine, right? If you’re drafting an NDA or you’re, you know, looking over a simple asset purchase agreement, like those are gonna be very sort of commodified and that might work fine, but the stuff that I’m doing, the stuff that you’re doing, the stuff that our partners are doing are one of ones, they are the hardest legal problems almost by definition.

Why else would they hire us? In the world and sort of by extension, the hardest reasoning problems, right? And so these models are great reasoners. I don’t know if there’s a lot of people, even at the model companies who’ve pushed it as hard as I have, or as hard as some of our colleagues have to really dig in and analyze very difficult problems.

And so by seeing what it can do for me on my actual problems, I can better extrapolate that to how I can help others solve their problems, or at least use this as a tool rather than sort of approaching it from the outside and trying to get to where Quinn Emanuel is. That’s just too hard, you know, we have too much, you know, sort of latent knowledge that, again, this is why I like Claude to ask me questions because often I understand concepts that I can’t, you know, necessarily get the words for yet. I think we all can. And so the models are great for pulling that out of me, and that’s one of the things that we’re also developing, ’cause I think honestly the sort of context assembly part of the Kerch bench, I think others are figuring that out.

I don’t think there’s like a tremendous amount of alpha, but because we’ve already done it, and that’s the foundation for everything, to get a smart model that knows your case, that understands your issues, then you can do more interesting things on top of it, including these skills that you and I have talked about.

JOHN QUINN: Yeah, let’s talk about those skills, which I understand are sort of like subroutines or scripts or small software programs that are designed to accomplish certain tasks that they’re sort of in the can and you can use them in your different case projects. Tell us about those.

CHRIS KERCHER: So this is something that, that people at philanthropic develop, they’re very creative and, and you know, one of the things that I’ve always done, I think I talked about this with you on an earlier episode of Law, disrupted, but every time I did something that was useful, where I got a good result from Claude, after, you know, sometimes a lot of painful back and forth, no, not that, no, you’re too sick of it, I’d get to the end, I’d be satisfied, but then I’d go one step further. I’d say, okay, Claude, if you were me, how would you have prompted this at the beginning or let’s make a template so next time we can just get what I wanna get and I, you know, you can fill in the information. I would save these templates.

And last summer, Claude released these skills, which I looked at. I said, those are just like the templates I save. In fact, I’m sure, because I was building them in Claude, they have the same architecture that Claude was using internally, and so now they’re using it externally. And so they’re little programs that can do everything from.

I mean, literally everything we’re doing, it can capture your voice, your style, if you decompose a given task, right? If we’re sitting down to write a brief, like I don’t just open up Microsoft Word and start writing. You know, you pull your case file, you take your notes, you identify the topics, then you organize it, and then you flesh it out.

You make sure it all hangs together. Then you start writing, then you polish. And so if you’d start decomposing tasks into discrete steps, you can build these really interesting skills, but you can also teach it to use the tools that these models have, whether it’s an internet research tool or this new tool that I really like, or it’s asking the user basically multiple choice questions that these are the three options, Chris, here’s a fourth space for you to add your own thoughts and help me sort of steer the strategy or steer what we’re doing.

And it’s so much faster I find for Claude to surface the ideas a minute before I do, you know, probably the same issues I would surface or, you know, 80% of the way there and I can fill in the rest that you’re just, my brain is moving faster and my productivity, I think people are really using these things hard and I hear this more and more is my own experience.

It’s just your productivity soars by an order of magnitude that’s hard to, sort of, hard to fathom. Like if I think about where I was two years ago, I was obviously I guess an okay lawyer. I was, you know, at the world’s greatest litigation firm. I handle a lot of big cases, but I feel like my skills, my abilities, my productivity have soared because I’ve really mastered how to get the most out of these.

JOHN QUINN: I mean, I’m having a hard time grasping what these skills are that you refer to these routines that you use, that I guess are part of Claude. They’re already, these exist. How can you talk about that a little more?

CHRIS KERCHER: Yeah, yeah, yeah. Let me go back to that. So we create them and so the brief writing skill, for example, you call, I want the brief writer and it will break down in steps. Okay. Chris, I’ve gone through the knowledge base, I understand you want to file a motion to quash. Here are the pieces of information or the topics I think we should cover, is there anything else?

And I might say, well, what are you thinking? Are you missing, you know, or yes, you’re missing this, or, I wouldn’t get into that. You go back and forth and then it surfaces, okay, now here’s how I would organize it. And if you think about it, how many ways are there gonna be to organize this and for you to sit down and do it by hand.

You know, some people, you know, that may be their thought process and that’s fine. But for me. It’s so much faster to get that assistance, that low hanging fruit where I can just start to add, where it’s not picking it up. And so that’s one skill, but another skill, there are these skills that are sort of these meta thinking skills where I think I showed you one called, Hey, JBQ.

It was based on what I had called Hey, Kerch, and they walk through all the different dimensions, like they’re just sort of an infinite number of questions. And I brought you onto a case and I told you all about the case, and I asked you to help me think through it. You were just. You know, ask questions in, you know, on strategy, on tactics, on the case law, on the facts, on the bigger dynamic.

What are the client’s objectives here? What are the real incentives? Who’s the decision maker? Like you just sort of infinite questioning that can help me surface what I missing. Where can we take this next? How do we unstick this or get to a better place in the case so I can build those all day long?

I’ve been talking to our teams that have been building Kerch Benches and learning about how they go about drafting briefs, analyzing patent issues that I knew nothing about before. And I can say, okay, let me make that a skill, and you can see how it goes and it’ll walk you through your own process.

And if you find it helpful, you know, we can iterate on it. But they’re very cool and we’re just starting to roll those out. So I think that’s gonna be another real advantage because we know how we do things and to be able to translate that to the AI is not that easy. But I’ve spent a lot of time sort of figuring out how to get that out of people.

And so we can start to push that out too.

JOHN QUINN: So by using these skills, you’re able to get to a close to finish work, product much, much faster.

CHRIS KERCHER: Yeah.

JOHN QUINN: With less of your own effort.

CHRIS KERCHER: More high value effort. So I actually, I have a bunch of tabs open and Claude might be thinking for 15, 20 minutes and I go to the next tab. And it’s hard. I have to like, you know, one of the tricks is loading context, right? If you’re doing 50 cases and you have to jump on with clients all day, it’s really hard to remember what context matters and how do I get it in front of mind? And so one of the things that I have a skill for is Claude to surface the context I need for that problem so that I can open the tab, I can reread the context to refresh what I need.

JOHN QUINN: So it’s sort of a case summary. Is it sort of a…?

CHRIS KERCHER: It, it’s a summary that is relevant to wherever you are in the case, whatever you’re working on.

So you don’t need everything, but it’s good at surfacing from all that information. Okay. Right now, Chris, you have to write the mediator. So here’s the information that I wanna remind you of before we start working on this. You know, email to the mediator or whatever it is. And so then I go tab by tab by tab and I might be thinking about something really hard for a stretch, but it’s really high value thinking.

And so what comes out on the other side when I give it back to Claude and it thinks for 20 minutes is, you know, work product that is far beyond, I think what I would’ve done on my own, certainly in 40 minutes, but frankly probably ever.

JOHN QUINN: Why is that? I mean, I get the time savings, but why is it that you get to a higher quality product than you think that you ever would’ve gotten to?

CHRIS KERCHER: Right. Well it goes back to this idea that if I’m giving Claude the limits of my knowledge, the limits of my understanding, like where if you didn’t have Claude, where do you go from there? Like you’re sort of stuck. I mean, you could talk to a teammate, you could go do some research, I guess you could you know, write notes or something, but to have a true thought partner that understands what you’re saying, understands the case, and understands, you know, patterns, theory, strategy, whatever dimension you wanted to attack from it, it can then surface those ideas that you’re missing and so not all of ’em are gonna be good.

But you can get volume. You can say, gimme your 12 best ideas, and all I need is one that I can dig my teeth into and suddenly I didn’t have one, let alone 12 before. So now I have one and I can start to think about it, and I can start to think about, okay, how do I bring my team in now? So I have this interesting idea.

You know, one of the things that we all struggle with is communicating our ideas to our team and giving good instructions and making sure that people are doing the right thing. Well, if you have Claude that knows your case, that understands what you’re trying to do, I can, I call them Kerch specs. You can see the branding going on here.

But a Kerch is, you know, write a spec, a set of instructions to my associate or to a teammate that describes what I’m trying to do and how they can help, whether it’s research or running down something in the record. And so they’re getting cleaner, better communication as well. It’s a communication tool.

JOHN QUINN: So you’ve obviously spent a lot, you have no background that you’ve told us, really as a computer scientist, software engineer, you don’t hold yourself out as techie, though. Maybe you’ve made yourself one, just kind of in summary fashion, how did you get to this point?

How long has it taken you, and how did this originally spark your interest?

CHRIS KERCHER: Yeah. You know, it’s, so, I am, I think an early adopter. I was reminded recently that if you go back to when the iPhone came out, I was quoted in like the first or second Wall Street Journal article that you know, because I just, it blew me away. My Tesla self-driving blows me away. Chat GPT blew me away.

It was just, this is unlike anything else I’ve ever seen. So I started using it. The applications to our world were obvious, but I assumed that, you know, it’s talking to a computer. So I assumed that people who were very technical would be way beyond me and what I learned, I actually had my daughter, who’s 15 and she took a coding class.

So she got me on Claude Code last summer, like right when it came out. And I, what I learned was, I actually think it’s a little bit inhibiting to be a peer software engineer because it’s about language, it’s about how we express ourselves, it’s about how we give context. Those are like core legal skills.

And so I originally didn’t have like a ton of confidence that I was very knowledgeable about this, but as I’ve gotten out there more, I’ve been speaking at virtually every great law school. I’m teaching,for t wo weeks at Stanford starting next week, which is so cool. But I’m hearing from all of these young lawyers who are really interested in AI and wanna learn more about how they can get involved, how they can do this in their careers at some other firms, they’re not taking the same approach that you and I have talked about many of our partners where we’re using AI responsibly that we understand and we have strong, you know, verification procedures in place that we always want to make sure that the judges who read our work understand that it’s really our work and that we stand behind it.

And so we’ve thought about all of these issues and I think we’ve created a really great platform for people who want to use AI and who really get it, but feel inhibited. I know we’ve got a bunch of summer associates joining this summer who are joining us for that purpose. And so it’s been a really interesting process, it’s been exciting to see more and more people at the firm get on because that was one of the ways that I, one of the reasons I came up with the Kerch Bench is I kept showing people demos. And they weren’t very impressed because it wasn’t their problem.

You know, I’d say, look at this preliminary statement that I wrote, and it, you know, I don’t know, it’s a preliminary statement, like it’s not, you know, 30 times better than anyone that’s ever existed. It’s what you’d expect from a good lawyer, but if you, if they saw how hard it was and where I was struggling, and so I created this Kerch Bench for one of our partners cases because I realized the only way he’s gonna appreciate what this can do for him is if it solves his problem. And so we built that out. We’ve had a whole bunch of people who’ve joined me on my AI and data analytics team who’ve helped me develop the methodologies and encourage teammates to use it.

And so it’s been a great process and I think we’re just sort of at the beginning. I think the best is yet to come.

JOHN QUINN: If you were talking to a prospective client, so you were, say you were pitching a case and you want the client to know about the capabilities at our firm, the capabilities that you’ve developed, I mean, what are what? What would you be telling them? What are the advantages to the client?

CHRIS KERCHER: Yeah, and this is a really interesting one. I do this all the time. I did this this morning with one of our partners and you know. Part of the reason that I’ve, I’ve deployed this as I have, is because it would take too long for me to go to every client. I can’t leverage myself that way, but I can talk to our partners who get it, who understand the challenges and struggles, who understand more and more how Claude works, how AI works.

I don’t think, frankly, it’s a great product if they’ve never used AI. I mean, I can join them, I can teach them a lot, but the people who are understanding AI are getting the most out of it. And so, you know, I’ve played around with a lot of different ways to help clients understand it because clients are saying, wait a minute, what do you mean?

You know, this is something that you’re charging for. I thought AI was saving money. And it’s like, I know that it’s so much more valuable than what we. I charge it for that. The return you get, even if it’s as simple as you’re summarizing everything way more efficiently, but you end up creating this better output.

And if the goal is to win the case, or the goal is to do the absolute best you can, you wanna be getting rid of all this sort of low value work that the AI can do. The just very basic generation, and you can’t do it unless you have a platform that understands your case. And I encourage people if they haven’t tried Claude, you know, we can set something up that’s a Baby Kerch that will show them what Claude does without it, and then what Claude does with it, and they can decide for themselves.

If there’s value, and the reason I call it the Baby Kerch, it’s intentionally downplay, I don’t want people thinking that this is the be all, end all to their problems and they don’t have to do work anymore. You know, it’s in some ways the opposite. So the names are sort of chosen to be a bit casual because you know, I don’t wanna like hype it up as the Ironman suit, but it’s actually pretty incredible and the more people use it, I think our partners will figure out even better than I can, how they can sell it to their clients because they understand where it gets real leverage in their cases and on the type of case that the client has. So, I’m happy to speak to any clients I speak to, probably four or five a week. But I think I’m most interested in talking to our litigators so that they can deploy this better than every other litigator, and we can get on, you know, one side of the v or the other in every big case.

JOHN QUINN: Fascinating discussion. I mean, my key takeaways are the importance of structuring the data that you give Claude. Don’t just rely on Claude knowing everything, but the value of structuring the data, making it very case specific. So in a sense, you distill down the case before you make demands on Claude.

And then the thinking about Claude as a thought partner that you challenge and the challenges you, I mean, you ask Claude prompts, everybody does that, but you also ask Claude to ask you prompts like, what am I not thinking of? And I find that really interesting and fascinating.

CHRIS KERCHER: Yeah, and those are the three takeaways. And the one I would add is the pattern I see some people doing is they have their problem and they think they have an understanding of what AI can do, and so they give AI a task. That’s within some other problem. You know, give me this fact or look at this thing in isolation.

And I think the real leverage, and I think the people who are ahead of us, who are developing Claude code and who are developing these tools, see that, just tell it your ultimate problem, tell it your ultimate objective. Tell it what your constraints are. And it can get very creative around that, including as you say, by asking you or prompting you or suggesting things to you. So, step back and think a little bit about what you’re trying to do in the big picture and tell it to Claude and see what happens. Have some faith that it can do virtually anything if you give it good enough instructions and enough context.

JOHN QUINN: Thanks very much, Chris, we’ve been talking to my partner based in New York, Chris Kercher, who’s done a lot of amazing work in developing an AI platform for our firm, for litigation AI, for litigators by litigators. This is John Quinn, and this has been law, disrupted.

---

Published: Apr 13 2026