Speaker 1: Welcome everyone to the LTH product briefing. Very excited today to be introducing a brand new product. I have here with me today, Jake Heller, CEO and co-founder of Kacetext. Hi Jake.

Speaker 2: Hey, good to see you.

Speaker 1: Good to see you. Thanks so much for coming onto the LTH product briefing. Most people know you, most people know Kacetext. It's been around for about 10 years. And because people are familiar with Kacetext, what I'm most interested in, in terms of your founder background is what was the mission upon which Kacetext was founded originally?

Speaker 2: The mission we're trying to solve for is to empower lawyers to provide such high quality, efficient, and affordable services that their clients benefit. And we look at this providing a lot of, I think, medical technology companies look at as providing and empowering their doctors to do better. If you give an MRI machine to a doctor, they can provide better health to their patients. We believe that by providing the best technology to lawyers, the end outcome is that they're able to provide better justice.

Speaker 1: I love that because it allows Kacetext as a company to evolve as technology evolves. So we originally knew Kacetext as a legal research company and then evolved through Compose and Parallel Search. Jake, what is the product that you're here to talk to us about today?

Speaker 2: This product is called CoCounsel. CoCounsel is an AI legal assistant that lawyers can delegate like substantive legal tasks like doctoring review, legal research, reviewing contracts, assuring contract compliance with a policy or set of playbooks within a firm, delegate those kinds of tasks to the AI for it to do a lot of the heavy lifting on for the lawyer.

Speaker 1: CoCounsel is exciting for numerous reasons, but one of which is the underlying technology, something that people have been talking about a lot is generative AI large language model. That's correct?

Speaker 2: Yeah. And large language models have come a very long way. We've been working with large language models since their inception about five or six years ago. And that's actually the same exact technology that empowers some of our search engines like Parallel Search. So it's been a very exciting and powerful technology. One of the things that I think everybody's seeing right now, if you use products like ChatGBT is that it's getting more and more intelligent. What we're seeing now in the latest and most modern applications of this technology is getting to a place where these are AIs that can read, write, and understand very complex texts at a postgraduate level. To us, that's an inflection point. That's a point at which something that was kind of neat and interesting, how it would do okay at that in previous kind of previous models, the next generation of models do it at a level where a lawyer can feel very good about delegating those kinds of materials, those kinds of tasks to this AI and for it to get extremely good answers at superhuman speeds.

Speaker 1: I know that we have had a lot of debate in the industry recently about large language models and their applicability in law, whether this is really an inflection point in the industry and beyond, and whether indeed there are use cases for large language models of the kind you're speaking about in law. So I think probably the best way to demonstrate that is to jump right in and take a look at CoCounsel. So let's jump right in.

Speaker 2: This is CoCounsel. I'll start here, like with other chat applications that people may have been aware of, if you want, kind of ask CoCounsel to do anything for you via chat. So I can say, write up a quick summary why legal tech enthusiasts should be excited about large language models. We've programmed CoCounsel to have a number of legal specific skills. The six that we're launching with are things like asking it to review documents, to do legal research, to extract data from a contract, to summarize large and might be difficult documents, other similar skills. So to give you a sense of how this works in practice, I might ask my AI to do legal research for me and say, what is the leading standard allegations under the False Claims Act in the Ninth Circuit? I'm just talking to this like I would a colleague. What's happening behind the scenes is it's deciding what search queries to run based on this query. It is running those search queries against our database of daily updating cases, statutes, rules, and regulations. It is reading hundreds of results that come back from these searches. And then it's going to use that to put together a legal research memo in response.

Speaker 1: How can lawyers who are using this know that the result that comes back is reliable?

Speaker 2: It's a great question. The short answer is two pieces. The first is I'll kick off a Docker review project that's kind of similar. I'm uploading 33 speeches that Barack Obama has given and you can ask any questions in it. Does Obama make a joke? Does Obama mention his daughters?

Speaker 1: Does he mention his dog?

Speaker 2: Does the speech include veiled criticism of Republicans? What this is going to do is bring back answers relatively quickly. It's also going to empower you to find the exact page where it claims it got the answer from for you to verify for yourself. For all of the skills, it's going to answer questions based on the cases that it's self-reading. It's also going to answer questions based on the documents that it's reading. So you can verify that entirely yourself. There's also another set of things that we do on the product side to make sure that it's answering based solely on the information that it's, that it's reading, that it's not pulling from extraneous information from kind of from its memory. Already it's pulled dozens of cases. We'll in a moment kind of start writing its reason why it believes this case is relevant for every single of these cases.

Speaker 1: In addition to actually generating the memo, it's also showing you the cases from which it has drawn the expertise in the memo and the sections of those cases that are relevant.

Speaker 2: That's exactly right. You can see right now is currently typing out its memorandum. This is the kind of work that as an associate would take me hours. It's giving me this deep and nuanced on space in the crew.

Speaker 1: I think some people worry that this would replace a lawyer, but really what it's doing is giving you an, an extremely good place to start with as well as the cases so that you can go and look at them and tweak the memo and save many hours of time that you would have spent, it still requires the human to look over it in the end and ensure that it's expressed the way that it would like to be expressed.

Speaker 2: You go a step further than that. First of all, I think that lawyers empowered with this are going to be able to do so much more, so much better and so much faster for their clients that they're going to be in much more demand and going to be put in a position to do really great work for their clients. I think that's going to be really empowering. They're not going to see less work and see a lot more demand for their work at a cheaper, lower price. I also think it's actually knowing what questions to ask and knowing what to look for in the documents, what's looked for in the cases, how to conduct that legal research. I think the new skill is going to be about deciding how do you delegate to this incredibly powerful asset that is going through hundreds or thousands of pages of text incredibly quickly. So for example, for every one of these, it's saying, did Obama make a joke? He jokes about not having remarks in front of him and also use a layer tone to talk about his presidency.

Speaker 1: That additional information is so critical for lawyers and so useful to get that on the page. Exactly.

Speaker 2: It's kind of explainability, including pointing to the exact right page in the case and even open up the page for this speech to confirm or deny is a really critical aspect of working with this kind of AI.

Speaker 1: I can see that in some cases it said insufficient. So if you ask it a question and it does not know the answer, will it actually say, I don't know the answer or I have insufficient information in order to respond to this question?

Speaker 2: That's exactly right. That's exactly what it's doing.

Speaker 1: I had heard that there was a limitation to the number of pages that this type of technology is able to read, but you've just added many pages worth of speeches. Have you architected something on the backend that allows co-counsel to read many pages of documents? Is there a limit?

Speaker 2: That's one of the major engineering breakthroughs that we're working on to make this kind of technology work, which is putting it in a position to read through thousands of pages, a hundred thousand pages for the large ones that work with us, they can read up to 500,000 pages of documents overnight using this technology, putting them in a position to do an intense amount of e-discovery or categorizing documents, their cam database. That's a major breakthrough that has been worked on.

Speaker 1: Jake, this is fascinating. I noticed that one of the skills is summarization, which is another use case we are seeing in the industry. I'm very interested in the extracting data from a contract.

Speaker 2: Here, for example, I'm uploading, you know, a few different commitment letters and private equity deals for leveraged buyouts for loans. You can ask any question like who are the parties in the contract? What is the termination fee? I can specify that I want that in the form of a number. By the way, if it's, it's an interpretation wrong, introducing kind of everywhere, give you the opportunity to re-edit your question and that way you'll have a higher likelihood of it really finding things you're looking for.

Speaker 1: It's helping you engineer prompts that are effective. Exactly.

Speaker 2: Yeah, this is ultimately extremely flexible. So you may have as a firm or as an attorney, your own set of questions you want to ask against contracts. And this will, like document reviews, pull out deep answers to those questions.

Speaker 1: I'm curious about a couple of other things. One is security. It's one of the things that law firms have been concerned about with ChatGPT, which is an open source. Firms are often very concerned about security around their own work product and contracts within a transaction. How can they feel safe using Co-Counsel on their own internal work product?

Speaker 2: Being in legal tech for the last 10 years and working with 60 of the largest law firms in the trade means that we totally get the security concerns that firms may have around these kinds of things. And part of what's important to distinguish is when you're working with a public tool like ChatGPT, every time you put information into it, that model is learning the information for your discussions with it. Large companies like Amazon and even Microsoft, which is a big investor in open AI, the folks who make ChatGPT, tell their employees, don't put any confidential information in here. It can leak into the big model. For us, we have our own private servers that host these models. It never learns by design, any of the information you're putting into it, never retains any information that you're putting into it. The only way that the data comes in contact with our AI is for the processing, for reading and understanding and putting it out. And then it's gone forever. That's a really, really important part of our architecture. Frankly, firms should ask these kinds of questions. There'll be many more in the future that attempt to do these things using this brand of technology. There is a major difference between the companies that can provide that kind of level of security and capabilities and those who cannot. Firms have to look into that when they're trying to decide whether they leverage this really impressive and exciting technology, which they should. They have to look into whether or not the model is being trained on the documents that are being put into it.

Speaker 1: Are you able to save a model effectively? So a list of questions so that you could reuse that for later. And with data that is extracted, are you able to tell it to populate a summary table so that it extracts data and then also populates a table or a database with the data that has been extracted?

Speaker 2: Absolutely. So what we do, for example, here is I've uploaded these documents earlier. I'm going to select them and I can run the same analysis of those documents. It also saves previously used questions. Those are the questions that we asked earlier. You can heart them. So they'll be at the top always. I can kick off a doc review with exactly those kinds of questions again. And we're going to make it even easier in the near future to save templates. In terms of getting the answers out, we always allow you to download things, for example, to Excel spreadsheet, or in the case of a lead research memo, to download it as a Word document. And you will get, you know, all of this context, all the cases, the answer, links to the cases, the quotes in the cases that are relevant, et cetera, all in a version that is easy for you to use.

Speaker 1: Oh, this is fantastic. Jake, you're already working in beta with a number of large firms and other entities as well.

Speaker 2: That's right. We found it very important over the last nearly six months now to engage with folks who are at large law firms, small law firms, nonprofits, in-house counsel, to really put it through its paces and make sure that it works really well. In the early days of the beta, they caught issues with the way you interact with it, what kind of skills we should really prioritize and use, help them make it faster. All those kinds of things that the beta customers did really great work and really great feedback for us to get to the place we are now. It puts us in a position that when we rolled out to the broader market, a lot of the kinks have been ironed out. Still new technology still needs to be worked on and always will be, like we're constantly improving. But a lot of the kinks have been ironed out and we get to come out with a pretty fully baked product.

Speaker 1: Well, this is fantastic. It looks like CoCounsel is the lawyer's new best friend. CoCo, their best friend. Yeah. CoCo, I love it. Like their mascot, their mascot. As of the day this goes live, March 1st, is it available? Can firms access this, call you, demo, pilot?

Speaker 2: We're going to be working with a smaller handful of firms at the very beginning because we want to make sure that as we deploy this technology and send it in responsibly in a well-trained kind of manner, that we have the server capacity, which is very expensive and difficult to obtain, to make sure that everybody who's using it gets a great experience, but we'll be rolling it out to as many people as we can, as fast as we can, making sure that the first customers who are onboarded have a great experience. So if you're interested, come March 1st, we'll have very easy ways to get in touch and know that we're working around the clock to make it so that after that first batch or two of clients who come in early and get rolled out, that we'll get to the wider market as soon as we can.

Speaker 1: Jake, thank you so much for showing us CoCounsel or CoCo. It looks like a groundbreaking new product. So congratulations.

Speaker 2: Thank you so much.