Todd Itami on Building AI Inside a Law Firm

Thanks, Daniel. Great to Great to talk to you and Brandon today.

Yeah. So you know, kind of like, half my time is spent doing your normal looks like normal eDiscovery lawyer stuff. You know it whether that's on big giant Cids, big giant products cases or smaller cases. You know, hatch waxman act litigation teams of 5 or 6 lawyers. So it's kind of your normal what everybody else is doing at law firms in terms of eDiscovery. But the other half is really focusing on the integration of legal tech more broadly. And of course, you know, the AI gets all the attention today, and it is awesome, and I'm a big nerd for all of it. But you know it's kind of looking at solutions broadly. What things can we do to deliver new value to clients with the tools, whether that's automation?

Or integrating the new technologies in ways to save money, or, better yet, kind of deliver new value. Kind of get more insights on the data than we've had before do new and cool things. And so it's just trying to be the glue, and, you know, operate in this kind of intersectional space between. Maybe software engineers, developers, analysts and lawyers. And so it's really kind of operating in that space where that's that's where I have fun.

Yeah, it's interesting. You know, the question of like, how long will it take law firms to adopt these technologies and react to these is, you know, we could talk for a long time about that, and reasons why it might be faster or slower than past adoption of technologies, because in some ways, you know, the practice of law has not changed since, like 1870, in a lot of ways in the way lawyers do business. And yes, we, you know, have some refinements around document databases and the word processor. We've decided to integrate that one now. But this is going to be a big shift. And this industrial revolution of the AI era. You know, the question is like, Okay, how fast are firms going to react to that?

And I think everybody has to react. I think everybody is going to need to change the way we do business eventually change the way we leverage these tools. I think we'll staff things differently. And it's really, you know, people kind of look at it. And AI and these technologies as just efficiency amplifiers. But yes, yeah, it's going to do that. But you know, the other stuff is the stuff that's really interesting, the new value that it's bringing, the way it's going to let us interact with the law. And the role of the lawyer is going to be very different. And so I think, will big firms have to. Yeah, they're going to have to do something, you know, whether they move affirmatively and whether they need to be kind of 1st movers or in front of the wave.

or they can do the old, the classic like, well, let's just wait and see what our competitors are doing, and then kind of like follow suit. I don't know which one of those it's gonna look like. Exactly. I tend to think it's going to be the 1st one more than the latter. But yeah, I'm not sure people have a choice, you know. I'm not sure firms will have a choice eventually.

Yeah, I really, I think this time is different. I don't know in a lot of ways. And so I guess the question is okay. Why? Well, you know, why do you think this time is different? And in some ways, like all of a sudden, software is cool. You know, lawyers are paying attention to this. Lawyers know they need to learn something, and there seems to be an energy around it that's past, or at least kind of extending the hype cycle curve if you will, that it that it seems to have momentum. It seems like we're going to have to do this. And the truth is that like it is just such a valuable tool, even if you're a lawyer and you're using vanilla X model, you know to do your day to day.

I've yet to meet somebody that really gives it a chance really digs in with the tools and tries to use them, and then, after 10 or 20 h, says, Yeah, never mind. You know, I'm going to wait for it to progress a little bit more like, you know it's just so valuable. It does so much that I think it's going to be hard to resist this one.

Yeah, I mean, we're, are we talking in the eDiscovery context, or.

No, no, I think it's, I think, kind of there's the law writ large, and how the technologies will play into the practice of law as a whole is one thing, the discovery function, or this place that we're in here talking about that. So I think there are a couple of things here. Well, one. the end of the rainbow, I think, looks very different than now, you know. Right now we engage in very elaborate processes to capture data in a forensically sound fashion. filter it, then go give it to lawyers, and they read it through it, and, etc, etc. But if you, if you think about that, the point of collection system files and all that amount of data compared to what actually gets produced in the case, you would blow my mind if it's 1%, you know, because you do some initial NIST, you get away the system data all out of the way that takes out a big portion of it.

Then you run these search terms that usually cuts it down another 80%. And then the lawyers review it and decide what's relevant, what's not relevant. And then you produce these documents and then what the reality is. In any case, only a handful of those documents make it into trial right? Only a handful of those documents are in any way important enough for somebody to look at a trial. And then how many of those are actually outcome, determinative. It's nuts. If you told me 1% of what we collected at time, one is important enough to get the trial, I'd be shocked that what kind of case is this? Maybe in, you know, very low discovery. Very few documents or criminal cases, or something like that, you know.

But when we're talking kind of like large scale corporate, big law, firm stuff. it's wildly inefficient, the process. And so it's kind of like, yeah, that is gone, you know. I think it's we're moving to a place where we're going to see Enterprise platforms have on them. We're going to 1st get over this indexing problem that we have. Now, that's essentially the issue that we have and why we go collect all these documents right? Because the data is sitting in place. and we don't have good ways to search it. We don't have good ways to search it where it's going to search attachments or text that is not recognized as extracted text. We don't have good ways to go and look at it in place.

And so that's why we do these big forensic collections and run them through eDiscovery tools. So we produce things that are right. Well, what happens when indexing improves, and we're already seeing this? We're already seeing a lot of attention with AI. All of a sudden there's attention on indexing, because people realize, wow! If we could just get this data into the LLM, let's go do that. And we're seeing really cool stuff. We're seeing, you know, software's tokenizing and creating really elaborate indices that allow the LLM to interact with it. Well, for E. Discovery purposes. That's the whole game. And so I think at some point the Enterprise platforms themselves. We're going to be able to have a conversation with them. and we're going to say, here's what the case is about, what's relevant.

Look at everything. That's the end game, and you know, that will trickle down to the rules of procedure that will trickle down to a whole new way of doing eDiscovery, because we don't need to go through the elaborate song and dance of mass collections. And even if that happens on a shorter scale, even if that happens not quite that extreme, it's still going to really, radically change the way we look at Ediscovery and change the whole complex of the industry. Daniel, go ahead.

Do it.

Totally. Yes, for the people who haven't switched episodes yet. So you know, it's indexing. So we have files sitting on a computer. And you could imagine telling a computer to go search through them and tell us, Hey, go! I'm looking for a document about cats, and so then it will go and read through everyone. And physically, you can imagine doing that in a linear fashion would take a long time. So you know, we have these things indexes or reverse indexes, whatever you want to call them. And literally, you know, the oldest and simplest form of this technology like it, would exist in a document review. Database looks much like a book. You know every document that has the word cat. It lists it there. So instead of having to go search through 10 million documents for the word cat.

You go to the index, go to the C cat. Okay, that's in documents 1, 2,005 and 1 million, and then it returns those values. And so just imagine that. But getting even more complicated, you know, creating rather than it being a simple book, index in the back. Think of things that are more elaborate phrases or concepts. Call them tokens. You can call them vectors whatever you want, but the idea is you're creating that book index. But in a much more elaborate fashion. and in a way that's more aware of its context. And so the processes now are not only creating the book index, so the LLM. Can simply find it, but it's also creating, extracting these more complex forms of indexing. And that's what it's feeding to the LLM.

It's feeding into it. It's much more efficient because the LLM. Itself also doesn't have to go through a process of, you know, converting the text. And so the idea is, once we're able to convert or look at all the data on a computer, once we have this elaborate index built. Well, then, we just have to look at it. And we don't have to go through these processes, because you can imagine. Now, if there's data on somebody's computer. And it's in a Zip file that Zip file is attached to an email. And that email is attached to an email. And on the Zip file. There's a Pdf, and on the Pdf, there's a photo. And the photo has text on it. That's important to the case.

Yeah, the currently, the technology doesn't exist like, run a search for that text. And it's gonna show up. But that could be very important. So now we have to like. grab the files, bring them into an eDiscovery tool, expand everything, run Ocr, and go through these steps, but if we had better tools it would be possible to go in and index the container and do the Ocr. And do these things that oh, it just so happens that Llms are really good at some of these functions. Lms and deep neural networks that things like. Ocr, it's great. And so that's really the idea is building a better book index. But kind of an elaborate and very fast one.

Yeah. So you know, within the firm or within an organization. You know, we have a lot of focus on training finding the best training tool to match your culture requiring training, you know, setting up that expectation and making sure everybody knows the limitations of the current tools they're using while I think the alarm around Llms in the sense of information, security and confidentiality was like sounded way too loudly. I do really believe in understanding technology, just like we should for a search engine, you know, just like you're not going to go pound some confidential information into the search engine. Say, can you translate this for me? Like. Yeah, don't do that with your clients. Data, you know. Be very careful with it, just like we have to be careful with all these other tools.

So it's that educational curve, but kind of like on a personal level. If my friends asked me, what should I do? You know I'd like to encourage people that? you know it's this tech is accessible tech it's accessible. It doesn't care who you are. or where you came from, or any of these things, and it's part of why I love it. You know I grew up in a very small town in the middle of nowhere, Idaho and my parents. We didn't have a lot of money, but they had this idea that, like computers, are the future, you know. And so they bought a 486 Dx. 33, Megahertz computer for me when I was in, like, I think, kindergarten or 1st grade, and that was awesome.

You know. It was a way of kind of connecting me to the world, and I could do things and program things and talk to other people and operate kind of on this doing, covering new ground, and it kind of like put felt like, oh, I'm kind of on the map like I get to be part of the larger world and doing cool things. And that's how tech is. It's accessible. And sometimes I think you know, you can listen to enough like eDiscovery, or read enough eDiscovery articles. And you'd be like, wow! It's kind of like there's like this click. They're like people who are in the know. And I'm just like that's so, not how it goes. You know, this is the this is the most accessible thing for everybody, and I encourage people to dig in like, go learn it.

I think it's it's a pretty good idea right now, if you're a lawyer to like, okay, you know. Go spend some time. There's great crash courses out there by some very famous professors and data scientists that you can do with AI that are like 4 h or 5 h. Those are all worth it. It's worth it to go even deeper. And you know the conversations I have with my clients about their models. In cases, you know, everybody just assumes you can read Python. Everybody just assumes you understand the basic architectures of these Llms and prompt flows. And how that interacts with cloud computing. And I don't think every lawyer has to operate on that level. But I would encourage lawyers to like. Okay. take the technology seriously and approach it with some humility and really dig in.

It's it's worth it. It's a no regrets proposition to like, go take one of these courses, you know. Go, take the Andrew in crash course, or go take a python crash course and learn these things. That's like a no regrets. Time investment. Yeah, especially with the coming wave of technology. So that's kind of my macro micro advice. I give at least.

Yeah.

End up in law school of all places. Yeah, I asked myself, that a bunch of times I don't know. You know. I did. I did college. And I started a business, you know, just doing started with web design web design web applications. And then it kind of expanded to. You know, any sort of b 2 b software or custom kind of software solutions for small medium sized businesses. And I, I did that. And I just got to a point where I sold both of the ventures that were taking up all my time, and I was like. I like school, and I always wanted to be a lawyer. So I when I went to law school, I thought, Yeah, okay, I'm going to be wordsmithing briefs for the rest of my life like I'm going to be showing up in court and talking to people.

And I'm done with these computers, you know, like that's that's the old Todd. I'm going to go be a lawyer. And so that kind of that lasted for like I don't know. 2 years after I graduated, you know. And then they're Oh, hey, we have this discovery database over here, I'm like, Oh, what's what's the architecture, you know? Like, okay, yeah, like, actually, do know about that. And then I was like, Okay, never mind. Let's just embrace it. So kind of just right place at the right time looking for something interesting. Look at all my friends like we're going to law school. And I'm like, Okay, took the Lsat. Got a good score. It's like, all right. I guess I'll go to law school, you know.

Not the best, you know. Not the most well thought out plan there, Daniel, but.

Yeah, yeah, it's it's really interesting to see who's using it the most how it's being integrated. You know, my friends, at a couple of my friends who are plaint slurs, you know. Good old fashioned plaintiffs, lawyers who handle these big claims like, Wow, they are really deeply using the technology. They're very interested in it. Then they're kind of like operating on the right out on the edge of how much value it will deliver to them. And it's very impressive to see the ways they're doing those things. And it's kind of different than you know, a big law firm. And the things it. How hard it is to kind of steer that ship, and I don't know, you know, for me personally it's I can see both worlds, you know I could see people who want to go to a smaller place, and it's more nimble, or perhaps you know, they their clients, all their clients are very tech forward, and they want to do these things.

But it's also great that you know the great variety of clients that I get to see on a monthly basis. You know. major tech companies get to kind of go behind the veil, see what they're working on. Advise on various issues. And that kind of like buffet of being able to look at everybody's stuff. And you know, see the way they're thinking about it and spend time, hey? You know, with a Ml. Engineer. You have a few minutes just to chat, you know, or whatever it's I learn a lot, and I love it, and I kind of like have this whole universe of things to look at. And it's, you know, being in a law firm. Everybody wants to show you their products, or everybody wants to talk to you about technology.

And that's nice, too. Whereas if you're at, you know a vendor or a smaller place like that might not be the case. You know your competitors might not want to tell you what they're what they're using, and so I get the kind of be spoiled to look at a bunch of things, and I really like that.

Yeah. I mean, I think you know, that's like, that's, I think the end game of eDiscovery proper is just okay. The software does a lot of it. and it's like, I don't know. People don't want to. There's a lot of reasons why the you know, the predictions in the legal market are different. You know it's there's a lot of voices and people who the industry will change, the industry will be transformed, and it will look different. And there's a lot of people who've invested a lot of money in the current industry. Right? And so some of that is that pushback, I see is kind of like, okay, you know whether it's important that we get it right? I don't know. But I guess the question is kind of like.

really is it, are we? What do you think's gonna happen? And I'm kind of like, yeah, really, I do. What's stopping it. You know. I there were probably there are 3 or 4 products, you know, startups this year at with Web summit that I'm just like, Wow, this is totally focused on indexing. totally focused on, and you know they're coming at it from a place of hey? We'll let you know what's in your files business, and you don't have all this rot, you know, or dark data that you can't see or you didn't know is there sitting around, and now you can access it. But it's really like, Wow, that's 1 stone throw away from being an E discovery tool and solving the problem. But I don't know.

I think that the kind of the large context features of Gen. AI and Ediscovery are really going to demand that people use it sooner than later. You know we think about it just like A tar replacement like this is going to be like better tar or faster tar. And that's fine. But it's in some ways that's kind of the most predictable use case. Great like. You don't want to be the somebody wants like, go put that before the court be person number one be my guest. But you know, in the meantime there's a lot of other stuff that it's going to do, and maybe cooler stuff than just determine relevancy and make yes, no choices on kind of sending documents in and out the door.

And that kind of new value things it's doing that are different. That seems more interesting to me than like, we have a better tar here, like, yeah, the better tar is awesome. I'm all for it. There are some issues. It'll be interesting to see how they play out using that. But yeah, it's the other stuff, the new value, the things that it's going to do that. We would never imagined possible that I think you're not going to be able to resist it. You know you're not going to be able to get away from that. Some lawyers going to be doing that, and if you don't do it, then they're going to go to that lawyer, because it's going to be cheaper and better.

Yeah. So right now, some of those. And I try to be product agnostic as I can. But I will say that. yeah, there's a the pricing, the premium for a tool that is legal quote. Unquote is quite a bit right now. But you know, I think that it's it's difficult for me to see how secret sauce legal tech secret sauce can outpace vanilla models on some level. And you know, when people lawyers are very obsessed with tools of legal tools, which tool and see this and that, and it's and I like them as much as anybody else. I mean, give me a break. I'm like a computer, you know. I'll try anything. But it's on some level. Okay, the people matter more. But the question is like, okay, it's hard for me to see how your kind of AI secret sauce is going to outpace so much the Llms that are changing.

We see these different companies kind of trading blows on functionality, on performance, and especially if the last 2 or 3 months haven't showed you everything. It's like, Wow, okay, things are still improving. You know, we there's kind of a hype out there that we hit a wall, because this one way of training like wasn't paying off, and we scaled it more. But I think we're kind of okay guys that's like one dimension of it. There's still a whole bunch of different ways that we can make this better. And so it's hard for me to see how a particular application or a particular thing is going to be so much better when these other models are kind of everybody's leapfrogging each other. And so it's not bad.

I mean, if you have some review platform and you're feeding into it like vanilla Api with X good model. That's you can get some pretty valuable results out of that, you know. If you if you sit and spend a little time of it. Wow! You can do some pretty cool stuff that you can't couldn't do before. And so I think that. And we have seen this from a few companies just saying, Okay, yeah, we just have a pipeline, you know, you pick your model. It's a pipeline and you, you know, we have XY and Z kind of grounding or rag techniques behind it. And that's I think that's gonna lower the prices down quite a bit, and also, by the way. the prices are very different, you know, if you go and choose the exact state of the art bleeding edge compared to something that was bleeding edge last year.

It's like 98% difference in some case. You know what I mean. The prices are falling, and especially if you can kind of think about these tasks in ways that you could stack models because you don't always need the Cadillac for everything you're doing, maybe you can send one of the smaller models through, and those are like pennies on the dollar. So I think the market will come around eventually. Maybe not. Maybe folks will be able to have the. You know, the legal versions of these things. They'll be able to charge whatever it is now, 10 or 20 X, of how much it actually cost, but to run the data through the LLM. But I don't know. I think the market will come around eventually. It's just we're just at a place now where it's as much as everybody's talking about this stuff.

It's just moving very slowly, and that you discovery context, you know, actual people using it. Implementation is not moving as fast as you know, the pundits, like myself, with like light up computers in the background, would have you believe.

Yeah. So right now, I think it's It's somewhere in the middle. There are some things that you can do. You know the big. You know, hyperscalers have very powerful platforms that make it very easy to set up a prompt flow that actually make it really easy to kind of do a diy rag model, and that's something that like I don't know. The other day I like set one up in a coffee shop while I was waiting for somebody on my phone, you know. And so it's it's it's really simple now. And a lot of these the software packages and modules that the hyperscalers have available make it really easy to do your own development work. let alone, like, you know, bringing it down, even me, writing my own model in python, like, okay, Nvidia, like kind of figured out everything with cuda.

And I just call it. And it does all the hard programming, you know, operate talking to the Gpu. And so the you know, really it's easy to do, Diy, but sometimes the devil's in the details for what you're trying to do. And I think that there is a place and a role for Diy, especially when you want to keep the data very contained. And you know, using a model that you can keep in your own system, you can execute it yourself, and the data never leaves your firewall. There's a role for that. And but, generally speaking, I kind of think like, Okay, but are we gonna outpace the rapper? You know I'm I'm kind of people talk about the rapper of these Llms like in a derogatory fashion.

But I'm like, Oh, no! I want somebody who's doing the rapper work for me. that then I can go give it to 2,000 lawyers. That's different than Todd has a particular problem that he wants to solve on a particular case and save the client a bunch of money. Okay, well, that's different. You know. You can kind of operate in the micro there. But when you're talking about, let's give this to all the lawyers. Then I'm like, Okay, are we really gonna be able to develop a ui, that's better than you know. Whoever else is out there with their ui? Probably not. And so it just depends on who the user base is. And what you're looking to accomplish. But I avoid the most. Hey, this is our secret sauce.

This is our LLM like it. Really, you know. Okay, like, maybe there actually are better than these models that doing legal work. I don't know but it's kind of hard to see. I haven't seen very many examples of anything comes close, but that's kind of how I look at it!

Yeah. So I think that there's there's this whole school thought by a professor named von Hippel that users are going to do. And this is far precedes. You know, AI, that users are going to be able to develop better functions or better use cases for any technology than designers than architects than people who are sitting around like me saying, like, I know how this is going to work. And so I think that getting the tools in the hands of the lawyers and getting them immersed in the technology. That's the number. One thing you can do start using it, and every little, every practice is different. People's workflows are different. And so, both on the individual level, on a practice level, you know what sort of legal work you're doing.

I really think that somebody. any associate sitting using an LLM. All day long in a transactional practice that I know nothing about, or maybe even a litigation practice where I don't operate frequently is going to be better at coming up with what are the good use cases here than I will, and so for me the step one, and it's a huge step is, hey? Get some sort of tools in the hands of lawyers and develop a use case registry. So put it, put it in the hands of everybody you can, and have a registry where we all come together and talk about it together. What are you using this for, and I think you know, that will create synergies within the firm, saying like, How do you use it?

And we'll be able to. Then you'll really be able to see where it will deliver value for your particular firm or your particular organization. So I think. you know, step one is, get something in their hands, and then have everybody put in a central place. What they're doing. Have somebody in place to go look at those, use cases. curate them or summarize them or something. But that's that's, I think, the 1st step. And in some ways I'm like, Yeah, take that 1st step and you'll there'll be more. You'll see from there. You know what it's giving you, and you'll kind of have an idea. Wow! Do we want to kind of create our own, you know, in house prompt flow for this? Or do we want to take a body of work that perhaps we've done, that we'd want to have the AI just reference that you know you're gonna see all these different places where it's valuable to you, and the light bulb is going to click on with X body of regulatory work or X code that we always think about.

Or you know, whatever database or knowledge management systems you know. the light bulb will turn on for everybody. So that's kind of like my one step like, get your feet wet. Go, go, put it in the hands of users, and then start listening to them, and that will that will help at your company, you know, of course. training you know, of course, letting every having the infrastructure putting, realizing like, Hey, this is gonna cost money@firstst it will improve our efficiency almost guaranteed does seem like another. No regrets kind of decision. But you know, I think my roadmap that would be stop number one for any organization.

Yeah, we love NeLI is I. It's small enough to where everybody's at arm's length. You know, you're talking to people. The conference is great. Everybody's in a room. They talk to each other. It's very friendly. It's very accessible. It's a great place to go and hear some really amazing ideas. All the speakers at the Conference this year were awesome way more worth your time. Listen to than Toddy, Tommy, but it was very cool, and you know, we have Federal judges. leaders in this industry talking about these ideas, and you know it's in a room, and people are raising their hand and asking questions, and it's a conversation, and that's very helpful to me. You know. I'm at some of these other conferences or things like I am just like totally saturated with marketing.

I am totally saturated with all the you know, the deck that everybody uses for X. And so I really like it. When things get a little bit more in depth, or are more conversational. It seems a lot more valuable to somebody like me, but we also, you know, had a bunch of law students there this year who I was talking to a few of them afterwards like this is great. You know. I learned a lot. I'm really interested in this now. People would just like stop by to check it out, and then stayed for the whole conference. So that was pretty cool, too.

That's great!

Well, I had so much fun. Thank you so much.