Get inspired by Laura’s journey from courthouse to tech boardroom.
We sat down with Laura Safdie, co-founder and former Chief Operations Officer at Casetext (now Head of Innovation for Legal at Thomson Reuters) to learn about how she helped start Casetext, and the company’s eventual successful merger with Thomson Reuters. Casetext was one of the first companies to create an AI assistant for legal use cases, and was a BuildGroup portfolio company. Today, Casetext’s marquis product CoCounsel, a trusted generative AI assistant for legal professionals and more, is growing and expanding globally at Thomson Reuters. We hope you enjoy reading our Q&A with Laura as much as we enjoyed the conversation live.
- Team BuildGroup
I'm a child of immigrants, and all of my grandparents have stories of being refugees or having to flee a place and work up from literally nothing to start new lives. I was raised in the aftermath of those experiences, and having grown up in an environment of stability, with access to good education, I always had this sense of deep responsibility to do something meaningful with that opportunity. I really cared about the administration of justice and policy development, and gravitated towards reading and policy discussions. I thought law would be one of the most important tools to drive social change and legal change in this country. That's why I went to law school, and part of why I've always felt this urgent push to make use of that opportunity.
Jake and I have had parallel career paths. We went to high school together. We both clerked in federal court. We both worked in D.C. and actually lived together in the summer when he worked at the White House and I worked at the Senate. That summer, we both worked on the Sotomayor confirmation hearings, he from the White House and I from the Senate. There would be newspaper articles saying, “White House delivers Sotomayor documents to the Senate”. And it was Jake with some boxes in a taxi and me in the bowels of the Senate basement being like, “thanks, bud”.
We were both at law firms, but he had this incredible background as an engineer. He grew up in Silicon Valley, and his dad was an internet entrepreneur in the 90s. We were both experiencing practicing law and how challenging it could be to find answers to legal questions. When I clerked in New York in federal trial court, I saw firsthand how shockingly disparate the submissions for different litigants were based on the level ofof representation they had and the tools their lawyers had access to. I saw the difference between someone who was operating pro se (that didn't have representation); they had their handwritten submissions, which is very common, by the way. It struck me that having effective technology—and tools that provide access to legal information—is a crucial part of achieving justice for all.
Jake had an idea to automate common legal tasks by harnessing the knowledge that already exists within legal texts and deploying Silicon Valley level engineering, legal informatics, data science and machine learning. I said, “Wow, that's a great idea. You should definitely go do that thing. I will be here at my law firm in my secure job cheering you on from the sidelines”. But we talked about it all the time, and I was giving him advice – both solicited and unsolicited. After a while, we said, “why don't we just do this?”
Yes. My experience in government was gratifying in many ways, but it was clear that you could be smart and work incredibly hard, and yet you could be prevented from making progress because of things totally out of your control. It is very hard to make systemic change consistently. Whereas with software, you can reach millions of people. And maybe it doesn't solve all problems, but it does create a real, important, and powerful platform where you can make an impact on real people's lives. That was a new concept to me and super inspiring. I thought, okay, well, if we can make legal information more accessible, not just physically accessible, but also easier to access, easier to understand we might be able to have an impact. We didn't even really understand the full complexity of the problem and how this was actually going to end up being a life journey for both of us. But we thought it was worth a try.
Well, Laura the lawyer was a perfectionist, and that is not actually a great characteristic for a founder. I had to learn to get comfortable with launching before something is perfect, because actually, you don't get to perfect, ever. Perfect takes too much time when you're trying to build something, especially in an iterative way informed by user feedback.
That was my biggest learning; how do you maintain excellence at speed without insisting on perfection or even subscribing to the fallacy that perfection is a thing at all. But still excellence, and still maintaining enough structure in which to make intelligent decisions. Finding that balance has definitely been an evolution for me.
We made a bunch of mistakes along the way, like everyone does, and some of them were really painful. It's weird to say, but my favorite thing has been the growth that came from those experiences–it was exceptionally valuable.
Like times when we hired the wrong executive for the moment. Doesn't mean they weren't exceptionally talented, just the wrong fit for the moment. It's particularly painful because you've affected that person's career and you've lost time, which is the most valuable thing you have in a startup–and credibility. And so I feel like a lot of the learning happened through those times when we had to recognize an incorrect move and change course.
We also learned a lot from working in a resource-constrained environment. Figuring out how to do a lot with very little is a really good skill to have across your life. You need to be effective and efficient in a law firm, but it is quite different when it’s your job to make sure this place doesn't run out of money. But we also needed to make investments in growth. So how do you operate under those constraints? It makes you get much better at making tough decisions quickly.
I would also say the relationships I've developed with the senior leaders from Casetext. That's been really gratifying over the years; building a team of exceptional executives, all of whom moved with us over to Thomson Reuters, except for one who retired. And supporting their growth in both their careers and leadership as we've evolved together as a company and team over the years.
When you're trying to invent something new it's nearly impossible to get everything right. Learning how to be nimble is key. The first time we had to pivot took us much longer than it should have. Then the next time was a little faster, the next time a little faster. And then the world changes around you and you need to react.
Jake and Pablo (co-founder and former Chief Innovation Officer at Casetext) were watching closely as neural networks became a thing, staying close to the scholarship. As the first large language models were coming out, they were testing and working with the foundation model developers, even though the tech was not anywhere near ready for prime time for professional use cases. We didn't know where it was going to go, but their foresight and work in this space put us in a position to adapt when the world turned upside down with GenAI. We developed those skills and built those building blocks to keep updating our mental models as new information came out. I think we proved at the end that we got really good at that.
The other thing is the leadership team we had. It was so good! It took us years to really learn what we needed at what moments and how to build a team that worked well together; that had enough flexibility but enough structure, the right personalities and the right level of skill, that wasn't so senior that they wouldn't get their hands dirty, but wasn't so green that they wouldn't know how to deal with issues as they came up.
When we discovered GPT4 and realized the enormity of the technological breakthrough, we pivoted the company in a week completely. That would not have been possible if we didn't have such a strong leadership team.
It was quite a story. Jake and Pablo and others on the team had been working with OpenAI and other platforms, testing different models as they were coming out. It was always like, wow, this is interesting, but not ready for professional work at all. And then I think in August 2022 or so, we got access to GPT4 under NDA.
Suddenly it became clear to us that you could delegate substantive legal tasks to an AI model and expect it to be completed at a high degree of quality and accuracy. And there were things that they tested that they learned over the years by working with these models that were the key tests, for example, can you point it at a document and tell it to answer only from information from within that document – a critical limitation for legal practice. That is important because even if a fact exists in the world or in a model training set doesn’t necessarily mean it exists in a litigation record. And you can't bring that outside knowledge into the litigation record.
They tested a few things and said, this is different and we can do something big here. And the question was, what and how and how quickly can we do it? And so we crystallized a plan for an AI legal assistant that you can delegate tasks to like you would a person, which now sounds maybe obvious because there's a market for it, but this wasn't a thing that existed at the time.
The problem was that we were under NDA. OpenAI wasn't letting us share with anyone because they were keeping a really tight control over this information at the time because it was so early. They hadn't even launched ChatGPT at this point, so no one knew this technology existed yet. We actually had a quarterly executive offsite planned that week, where we were supposed to talk about our plans for the business – but the whole agenda was from a pre-GPT4 world.
We begged OpenAI to let us extend this NDA to the whole company. We really needed to be able to tell them what was going on. We came in from around the country, we sat at the table and we said, “trash the agenda. Guys, we've got some news for you. Let us show you the future right now”. And so we showed our executive team GPT4, before anyone in the world had seen ChatGPT.
They were like, “holy crap, what do we, what does this mean for us”? We needed to do this very different thing because the world was about to change in a really big way, and we had this opportunity to build an application. You couldn't just take GPT4 out of the box and expect it to have the accuracy controls needed for law. That application that didn't exist yet. The thing that we needed to build would allow us to take the engine and the power of GPT4 and deploy it for very specific legal use-cases like legal drafting, legal research, analyzing legal documents. A very different set of things than the general LLM engine work.
We were lucky that we had all the leaders at the table. Literally at a table for the week. And we planned our future. We planned the product that we were going to build at a high level, the timeline that we needed to build it on, the inputs that we needed, the customers we could trust to work with us on this, etc. By laying that foundation, we were able to execute at pace.
Over the years, our vision never really changed. It was more about the ways that we were going about trying to achieve that vision. And a lot of it came down to, how can we leverage technology to automate pieces of a lawyer's workflow that we believed should be automated – that weren't the highest value use of a lawyer's intellect or training, but are a really important part of getting law done? In retrospect, these ambitions exceeded the state of science at the time.
At first, we took an approach where we thought, well, there's so much information out there in the world that we could leverage to annotate the law. The best legal research tools relied on many people creating that content. And we thought, there's so much content out in the world already, both in the law itself and within cases that we could extract by using really clever data science. legal informatics. and machine learning work. But also, lawyers write a lot and how can we get them to write in ways that we can then leverage to scalably annotate? We had different versions of that approach. Just to launch our basic product, it took us three years. We needed the law itself, we needed the annotations, we needed the secondary content that historical companies built over decades with hundreds and with thousands of people.
Our first product that we launched was CARA, a brief analyzer tool. Basically you could drag and drop a document and have research done automatically for you, which was a totally new way of interacting with legal documents. In the past, you would read the document and do complex searches and then build your case. Then we iterated on essentially an automated drafting tool which we ended up sunsetting because, in retrospect, it really needed the power of a tool like a large language model – which of course did not exist yet. But we learned a ton in that process, and in that process we developed what we called Parallel Search, which was a new kind concept-based way of searching. You could search a whole sentence as opposed to very rigid search terms. Parallel Search leverages a concept map to find the information you're looking for, even if it doesn't match traditional search.
In retrospect, we needed all of those pieces when we got to the point of building CoCounsel with GPT4. The reason we were so well positioned to create the application that was able to take the engine of GPT4 and apply it to law with the accuracy limitations that we needed, was because we had all those building blocks that we built over the years. It wasn't like we scrapped everything and did this totally new thing.
We took everything that we built, and the knowledge that we built, and were able to deploy it in a way that when they launched GPT4, we were one of the only fully developed professional AI tools because we had built all those pieces that we could then leverage in creating this application. That's CoCounsel.
I've been really focused on driving the adoption of AI across the legal profession, and a project especially close to my heart is AI for legal aid and pro bono work. AI allows you to save enormous amounts of time. Nowhere is that time needed more than for the people who represent communities that don't have access to adequate legal representation.
One of the organizations we're working with does exoneration work. It's an innocence project; people who are in jail, who have evidence that they are innocent and improperly imprisoned are writing to these organizations saying, “I've been sitting in jail for three years, and I didn’t do it. There is evidence in the case showing I didn't, and I didn't get a fair shake for X, Y or Z reason”. This happens a stunning amount. It is one of the biggest blots on our justice system in America.
So how do we leverage CoCounsel to help support exoneration work? The actual petition is called a habeas petition. The organization sat down with one of our customer success managers recently because we really want to invest in getting in deep with these customers. We wanted to find out, what are the needs? What are the bottlenecks? What prevents them from serving more people? What would help them do it faster? And they came out of this session with a dozen use cases across a habeas petition. From identifying all the witness statements in the record, to identifying all statements made to the police, to creating a timeline from that, creating a table of contents, a table of authorities, drafting the email to the court asking for a conference.
All these pieces that in and of themselves may not seem transformative, but when you add them together, we saved hundreds of hours per habeas petition, and they are sitting on a pile of applications for representation. Like, hundreds of people sitting in jail, waiting for them to read their applications because they don't have the people or the funding. If we can magnify just that use case across all innocence projects in America, which is what we aspire to do, we’re saving hundreds of thousands of hours that can then be spent serving more people.
Those kinds of use cases stop me in my tracks. It's obviously really compelling because those individuals are sitting in jail. Every hour, let alone every month that they're imprisoned as an innocent person is a tragedy. But even still, taking a step back, imagine the impact that could have even on just commercial litigation or real estate. All the things we do that take way too long to get access to justice or resolution. So much so that many people don't even avail themselves of the justice system because they fear it'll be too expensive. It's going to take too long. That's why we do this work, right? The impact is so clear. And now that we're at Thomson Reuters, we can see that impact at scale globally, at a scale that we never could have imagined at Casetext.
We’re trying to partner really closely with legal aid organizations that do work like innocence work, immigration, federal benefits, landlord-tenant... all the common access to justice legal practices, and learn from them about how we can best deploy AI. We want to build these use cases together and create basically a blueprint AI for asylum applications, AI for habeas petitions. I really want to get this in front of the innocence community.
Early on, it was really common for me to be the only woman in a room. Not at Casetext, but anytime I was interfacing outside the company. When we were raising money, I was usually the only woman in the room. And initially I noticed. There weren’t moments where I thought, wow, if I was a man, this would be so different. It was more like, you really can just sense your otherness when you're in spaces where no one looks like you. But over the years, I think I got better at feeling like I belonged in those rooms and had something valuable to add. I was ultimately one of the founders and thus responsible for this business and this team and the outcomes.
So, my advice would be, if you find yourself as the only woman or one of the few in the room, don’t let that shake your confidence. Instead, see it as a reminder that your voice matters all the more. I was really lucky, but also chose to be in a company with people who value diversity and who weren't going to sideline anyone based on their background or who they are.
Going back to my law firm days, I tried to dress in a way that would make me disappear and talk in a way that would make me disappear and not highlight the fact that I was a woman at all. And now as a leader more mature in my career with the benefit of experience, I actually lean into my full identity. I know it's a cliche, but I bring my whole self into work. I want people to know that I have to go to a parent-teacher conference, and I want them to know that you can still be a high-impact executive and have personal responsibilities. It's important for me that people know that about me and that they know that it's safe for them to bring themselves to work too.
The information in this blog post is provided in good faith without any warranty. It does not constitute investment advice, recommendation, or an offer of any services or products of BuildGroup Management and it is not intended to provide a sufficient basis on which to make an investment decision. This document is provided for educational purposes only. Discussions of current or former BuildGroup portfolio companies are intended for educational and discussion purposes only. Any portfolio company so discussed has been selected based on objective, non-performance based criteria.
This content does not constitute or form part of an offer of any investment advisory services of BuildGroup Management, LLC, nor does it constitute or form part of an offer to issue or sell, or of a solicitation of an offer to subscribe or buy, any securities or other financial instruments, nor does it constitute a financial promotion, investment advice or an inducement or incitement to participate in any product, offering or investment.
©BuildGroup 2024
©BuildGroup 2024
