AI Revolution in Legal Tech

Thank you, Daniel, and really, excited to be here as well. Thank you for inviting me.

Thank you. Yes. So I am a partner at Redgrave, the, country's preeminent eDiscovery boutique.

I am head of Redgrave's second request and antitrust practice, and my practice focuses on serving as discovery counsel for companies across a range of industries on not just antitrust issues, but other matters as well. And then I also help companies on information governance and other eDiscovery type counseling issues as well.

That's right. So I was formerly a partner at Sidley Austin and head of co head of Sidley's eDiscovery and data analytics group, where I spent, as you said, my whole, career at Sidley roughly about twenty three years before coming over to Redgrave.

The move was about Redgrave and what Redgrave could offer, specifically a focus on information law, on information governance.

That is all they do. Unlike other firms who, you know, naturally are much more diversified, but Redgrave's strength is in its focus.

So because all they do is eDiscovery and serving as eDiscovery council, that's where they invest and that is what they focus on. And because of that focus and because of that investment, that offers significant advantages to not just the companies that I was working for and with, but also really a large number of other companies that work with Redgrave as clients and specifically the focus on information law as a practice and the deployment of really cutting edge technologies, practices, workflows, and then a focus, honestly, on the non attorney specialist as well. What really attracted me to Redgray was not just their incredible list of top eDiscovery professionals and lawyers, but the large and really preeminent list and group of the non attorney analytics Relativity project managers specialists that really can make the difference in service and really in the success of a project.

And couldn't say enough about my experience in, just a short couple of weeks where kind of everything that I thought about Redgray that attracted me to the firm has really been proven out. And so it's really a great group of people and really excited about kind of the future here.

I would say it's two things.

So the first is that when Gen AI first came on the scene from the legal industry, it was much more about the promise of GenAI and what it could do.

More recently, what we're seeing now is the actual commercial products being rolled out by different providers.

And we're now seeing and we're using, for example, on our matters, we're deploying many of these products in live environments. So really in twenty twenty three, the beginning of twenty twenty four, it was much more about kind of sandboxing, testing different Gen AI applications on legacy matters, proof of concepts, and the like. And now the difference is we are starting to deploy it on LiveMatters and starting to kind of leverage the technology, you know, going forward to help benefit our clients. So that's one. I think the second big thing is that we're continuing to see the power of Gen AI increase significantly while the cost continues to decrease.

And the pace of that is where I have really been surprised. Like, you kind of think about that, like, this would be likely to have been given so much investment in the space.

But really, it's the pace of the technology increasing in sophistication while at the same time getting cheaper.

And then you're starting to see that then filter into not just performance of products and eDiscovery, but also you start to see the pricing be modulated as well. And so I think both of those two things are important for adoption going forward.

I think in the short term, I think we're going to see more diversification.

And I think that's a good thing. I think this space really needs competition and really needs options. And I think having more providers working on their own Gen AI solutions, whatever the use case, I think the better for the industry.

I think the other thing that we will start to see is further improvements in the UI, which I think is particularly going to be another important aspect to adoption.

Because if you think about from the TAR perspective, right, The technology of TAR in two thousand eleven wasn't that significantly different than the TAR technology in twenty nineteen or twenty. The difference was the UI of the applications that were being used and they were much more user friendly. They looked all that prettier, frankly, and they were more intuitive and thus attorneys themselves could in many instances, you know, be taking ownership of the TAR tool itself.

And so I think improvements in the UI is going to be, you know, another link in the chain to move GenAI forward.

We're still in the beginning stages. So at least what we're seeing as gen AI is much more in still in the field of the specialists.

But I do think that GenAI will have a better adoption her than TAR. And the reason is, is that TAR was always so mysterious kind of black boxy. That was the complaint from the very beginning, this kind of black box nature. It's not that an LLM is any easier to understand the technology, but the output, because it's so prevalent through other types of commercial applications, that I think that will help drive adoption. And the other thing, you know, we'll see how this develops, but, you know, at bottom, the prompt is words, right? As opposed to TAR, which was, you know, numbers and ones and zeros and vectors and the like.

In that sense, I think that will help adoption as well.

Right.

Yeah. So it is important.

And this is, I think, a really opportunity on the technology side and the providers to really help move things forward on the engineering aspect of this. So, for example, when you type in some type of prompt and then hit return. There's an enormous amount of engineering in the background that goes on that takes the actual words that you put in and translates it into a series of instructions for the LLM and such as formatting. It's not a accident. It will generate either, you know, bullets or structure. There'll be consistent style of format, and that's the engineering behind it.

And there's enormous opportunity for that in the eDiscovery space to kind of translate more natural language prompts that people will put into the tool and then actually then translating that into series instructions to make that tool kind of accomplish what the person is trying to do and then, you know, apply it to, let's say, a document set in a way that is most likely going to result in, you know, an output that the user wants.

I think it's going to be a continuing process, but we do see significant differences.

And so I think the biggest point right now for people to consider is that the use of GenAI is iterative on a dataset. So it's not just coming up with a prompt and expecting it to work perfectly right out of the box.

But coming up with a prompt, applying it to the dataset, looking for a sample, oftentimes looking at how the tool performs on the sample and then adjusting your prompt from there. And that's where we see that it oftentimes takes four or five, six times iterations of, like, looking and working with samples to get where you want. That said, there's a continuing, like, evolution, at least on our side, on the prompt where, like, for example, we have a e doc summary prompt that our partner Ray Mangum, who's a former computer engineer, has developed. That's twenty five thousand characters long, right? That's a lot.

That's a lot. But it took a long time to develop the first time around. It takes a lot less time to alter for future matters because you're adjusting to the specific facts of that case. And so we at least see on our side, you know, increasing performance over time as we're able to kind of essentially transfer learning, you know, from one matter to the other with respect to Gen AI.

That's true. But we're also partnering with different providers to help develop the Gen AI tools. We see that as, one, helpful, but also just moving the ball forward generally, but also useful for our own kind of learning and making sure we're we're still on kind of the cutting edge of the technology.

Yes. It's a great question. And this is where I think really the iteration is helpful and that the same techniques that we have been using for well more than a decade to validate TAR or if you wanted to validate your search term results, for example, those same techniques are the way to validate and make sure your GenAI process is defensible. So, you know, looking at samples, running illusion testing, doing validation sets, all of that is equally applicable to a GenAI review or really the use case of the GenAI. It doesn't have to be a review. There's lots of different use cases out there. But however you're using it to, like, you know, employ similar validation techniques to help make sure the process is defensible.

This is still a nascent technology and we have client considerations on all of this, right? And also there is friction with negotiating with the other side on a JNI protocol for the first time and that is going to result in a lot more friction than, you know, using a standard TAR protocol, even if substantively it shouldn't be that different. And so all those are barriers to adoption.

So we have a lot of challenges.

So I don't think you're gonna see that human replacement anytime soon.

I think there is instead, it is going to be a complement for reviews in different ways, and those use cases will gain traction over time as, you know, you demonstrate success. But I think there's going to be a central role for the human attorney in the years to come. And, you know, I don't think that is going to be in danger of a replacement.

You know, it may, again, not right now. In the years to come, it may affect the size of their review team, You know, where just like the use of TAR has affected the size of the review team, where, like, for example, I do second requests, which are merger related.

Back fifteen years ago, before TAR, it was routine to have contract attorney teams of four hundred to five hundred people, you know, on a second request.

Now with TAR, and we use TAR on essentially all of our matters, Our review team averages between sixty to one hundred and twenty people that's still a lot but it's way less than what it was. And so if GenAI kind of proves out then that team may decrease more, but you're still going to have a review team for different reasons. You're still going to have a place for outside counsel and otherwise to go forward. So, I think that the threats that are definitely talked about, I think they will eventually materialize in some way, but we're years away from kind of seeing that actually take effect at scale across our industry.

Yes. So I think there will be a role. You need to code sample documents, for example, and then run GenAI against the samples.

You definitely need some type of validation for that, and you need validation for all the other aspects, not just responsiveness, but for privilege, for if you're using it for issue codes and the like. And there's a certain benefit to validating on larger datasets that, you know, it may not make sense for outside counsel to do. And so I do think there's going to be continuing to be a role.

I think the power of the technology is particularly just beneficial in the face of continuing increases in volume. Because what I was saying is like search terms were revolutionary when we were using them in two thousand and four, right, as far as reducing the data set. But the problem is, is when your search terms are effective at culling seventy percent of the population, that is great when your population is one hundred thousand documents. It is less great when it is ten million. And so we just need, you know, better, more advanced technology to help us as the volume of documents continue to increase.

In the shorter term, I'm the most excited about GenAI's ability to create privilege logs.

Privilege logs is one of the least efficient, least useful aspects of our legal system right now.

And you could spend an enormous amount of money creating a log of privileged documents where the best case scenario is that the other side never looks at it. And it is very wasteful, very inefficient, the current system.

But it has perhaps the most more short term ability to be addressed by GenAI. Because GenAI.

Is particularly great at summarizing documents. Now it's a hard problem because it not just necessarily has to summarize the document, but has to summarize the privilege part of the document and has to do so in a way that under the rules allows the other side to be able to assess the nature of the privilege. So it's a very difficult problem to solve but it's where we've seen significant progress, over the last, you know, eighteen months and we are starting to see tools being rolled out. I personally see this as an area where we're going to see the requesting parties, the government, the plaintiffs attorneys, and the like, I think actually preferring Gen AI created logs as opposed to traditional logs because they can be created faster.

They're not being filtered. It is coming direct from the algorithm with, you know, QC from the party producing it. But I think it has a real advantage in speed and decreasing the burden of the review.

And so that's where I'm particularly excited about.

Thank you for having me. I really appreciate it.