SFG 44: Everlaw CEO AJ Shankar on how AI is accelerating legal tech
‍

‍Everlaw is a legal tech company started in 2011 that helps law firms and corporations run what’s called “e-discovery”— essentially a process of investigating often massive piles of documents to find the key pieces of evidence, the smoking gun, and then to construct a narrative from it. Last valued at over $2B, Everlaw has over 1,000 customers over the world.

‍

In this episode of the Startup Field Guide podcast, Sandhya Hegde chats with the founder and CEO of Everlaw, AJ Shankar.

‍

‍Be sure to check out more Startup Field Guide Podcast episodes on Spotify, Apple, and Youtube. Hosted by Unusual Ventures General Partner Sandhya Hegde (former EVP at Amplitude), the SFG podcast uncovers how the top unicorn founders of today really found product-market fit.

‍

‍

Episode transcript

​Sandhya Hegde

Welcome to the Startup Field Guide, where we learn from successful founders of Unicorn startups how their companies truly found product market fit. I'm your host Sandhya Hegde and today we'll be diving into the story of Everlaw. So Everlaw is a legal tech company, started in 2011, that helps law firms and corporations run what's called e-discovery; essentially a process of investigating often massive piles of documents to find the key pieces of evidence, the smoking gun, if you will, and then to construct a narrative from it. Last valued at over $2 billion, Everlaw has over 1, 000 customers all over the world. And joining us today is AJ Shankar, the CEO and co-founder of Everlaw.

Welcome to the Field Guide, AJ.

AJ Shankar

Thank you so much, Sandhya, and thanks for having me.

Sandhya Hegde

I am so excited for our chat. So AJ, you studied computer science at Harvard and then at UC Berkeley, finishing up with a PhD in 2009. How did you end up starting a legal AI company two years later? How did this happen?

AJ Shankar

Yes, it's not an obvious conclusion. I don't have a legal, law degree. I don't have any attorneys in my family. It's true. It was really happenstance.  At the time when I was in grad school getting my PhD, a law firm came to the computer science department looking for a technical expert. They had a relatively technical lawsuit and they needed someone to help advise on it.

‍

I happened to know the things they needed and so my advisor put me forward and I actually thought on a lark, hey, I'll do it. It'll be fun, I'll make a little extra money on the side. My fellowship is only so much right? And it was fun. But I also got a very eye-opening experience into the law, an area I had never thought about before in any detail. And I got to find out that the people in the space were very smart; the stakes were really high; I could appreciate what they're doing and the impact of the work; and then simultaneously I got to understand some of the technical challenges in the space, which were very deep.

‍

I was surprised at how high the technical ceiling was in this legal space, lots of hard problems, and that the technical bar was very, relatively low in terms of the technology they were using at the time. So high ceiling, low bar, relatively high stakes; I didn't know a lot about the law, of course, but knowing what I knew about computer science, I knew you could build something better. And that was very compelling to me, and it was an impetus to come back to it. I stayed in touch with the lawyers at that law firm. They kept asking me, was I going to come back and do this thing and eventually I did so that was it. Yeah.

Sandhya Hegde

Tell us more about the problem you worked on while you were at Berkeley. What was the challenge? How did they end up reaching out to a computer science department and your advisor? Was that something that law firms used to do often at the time?

AJ Shankar 

I think everywhere you want to find an expert in a particular area. This lawsuit happened to be about a particular way the internet worked. That is actually not that relevant to whatever law does it, the particulars of that lawsuit, but the idea which they also had to wrestle with which is hey, we have, in that case, it's 600 million records. They had to find some particular records that had a particular characteristic. How are you going to do that economically? The tools they had just weren't cutting it. And so it's really, it wasn't a legal problem, right? I didn't have to know anything about state or federal law. I had to understand the technical challenge, which I did, and recognize that you could solve it, right? And I think that was sufficient for me to walk away thinking, yes, there's a lot of legal problems that these experts can solve, but there's a lot of technical problems they can't solve. And actually what you really want is someone who has the technical competencies to solve them.

And I might be bringing that competency to the table so that was the opportunity.

Sandhya Hegde 

And what was your very kind of original vision when you first started the company? Did you have co-founders? How did it happen? And I'm curious whether it was very similar to what you eventually ended up building, or you started in a different place.

AJ Shankar

Yeah, I had the support of some of these partners at that law firm that I knew, but I was the only operational founder. So I did do it by myself from the beginning. And the vision is surprisingly similar to what we're doing today. And so the law has basically two halves. If you look at where work is done, there's the transactional side of the law.

‍

So contracts and M&A and that stuff. And then there's the litigation side, which is how do you resolve these disputes in court? And that's where we operate, right? Our aim is to sift through these mountains of evidence to find the kind of the needles in the haystack, right? The smoking gun as you put it, right?

‍

That vision is a kind of a big one. It turns out that to do that well requires handling incredible scale, millions and millions of documents over a long period of time, months or years. Many people, it's a big ecosystem involved in getting from haystacks to needles and then from needles to a narrative.

‍

And so the kind of goal was to build a piece of software that could be the platform, right? The platform that people would use to solve these problems. And at the time, and even today, in a lot of ways, there isn't a unified platform other than Everlaw. Like you might be using five or six or seven tools if you're a practitioner in the space, when you really should only be using one tool that's really well integrated, right?

‍

The same way that, Salesforce is a platform for sales, right? And NetSuite is a platform for finance and Workday is a platform for HR. There wasn't such a platform for litigation and that was the idea and it remains the idea.

Sandhya Hegde

Awesome. Now today, legal tech has become, fairly hot within applied AI. But I suspect it wasn't quite so 13 years ago, even though you were using fairly advanced technology in terms of like your vision, the kind of market for legal software wasn't as attractive as a lot of other categories that were getting investments.

‍

I'm curious what the early years of building Everlaw were like. Were you, you know, able to raise the kind of funding you wanted and build an early team? What was the first couple years of building Everlaw like?

AJ Shankar

Yeah it definitely was not the kind of space which had a ton of interest coming into it. And it's a hard area to get into if you're a technologist. The more you know about it, the less exciting it is. I joke if I know everything, if I knew everything back then that I know now, I never would have done it. But I'm so glad that I was completely ignorant of all the obstacles because once you start tackling the problems, you keep going. And before you start a problem, it's easy to find all the reasons to say no. It's just difficult for this or that reason. And you might go to an area that seems easier, but you know what? There's a lot of people trying to solve that easy problem. And your likelihood of success might actually be lower counterintuitively. So I'm glad I was ignorant. I say ignorance is bliss in this case because I was excited by the technical problem; I was excited by the product challenges; and then, the market needed a solution.

‍

And I would just argue forgetting the technical side, what we do in the law, the rule of law as an ideal is a really important ideal, right? The ideal that outcomes in the United States and these other countries are determined by a system of laws, not by some random person's diktat, right? So that's a powerful ideal. It needs great technology. And we are a truth finding machine in this process. It's going to lead to just that. That's just worth doing. So I would say appreciating that took me some time. I was first drawn to the products area. In retrospect I think it's very powerful and I wish, now I think more people are aware of the value of the rule of law. Let's just say in the last 7 or 8 years I think it's been a little clearer. But when I started, it wasn't as obvious. 

‍

Yes, it was not an area where a lot of VCs had any kind of thesis. We had some early two seed rounds of funding, relatively small, one of which was led by a mentor I knew from the prior startup I had done, but it took a long time to build a minimum viable product with a small team. And you can imagine why, right? The stakes are high in our space and if you're a law firm and you're going to a client and saying, look, we've got this big lawsuit, we're going to bet the farm on it. I found these two people in a garage that have really great technology; we should use them. No one's going to say yes so you have to actually build something substantial and build a reputation. And then also the surface area you have to cover is also not small, it's a relatively sophisticated workload.

‍

So it took five years of building, but we had some small amount of revenue early on, enough to keep us going. And then, around 2016 is when we really hit that first level I would say, as I know now, of product market fit and we went and raised that series A. And when we did, it was a lot of effort and a lot of rejection – I'm happy to talk about that – but it was definitely not an easy first round mostly because VCs didn't really have a thesis about the law.

Sandhya Hegde

Maybe starting with your approach to product development, how did you think about getting to understand your customers? If you look at law, there's so many different types of laws, so many different types of lawyers, and different types of organizations. How did you start building empathy for this end customer that you were building for? And I'm curious whether you ended up hiring anyone with a background in law into the company or was it, no, we are just a very tech company, we're going to like work on understanding how lawyers operate. Would love to understand your approach to kind of, you know, developing authenticity as a founder.

‍AJ Shankar

Yeah, it's a bit of all of the above. Like right now we probably have 40 or 50 people at the company with JDs. Most of them don't practice law, they just work at Everlaw and appreciate the law, right. Back then we had very few, it's true. We had no one with a JD for the first many years.

‍

One of those law firm partners was on our board so we had JD input, but I honestly found I don't think you need a lot of that stuff. You just talk to people and they tell you what their problems are. And I would argue that most good product sense is just good common sense, like 99% just good common sense – hearing someone's pain; understanding enough about the workflows to make sure you can meet the parameters; but then actually just coming up with a solution that solves a problem. Everlaw does not have tons of minutiae about how the law works. It's really around finding needles in haystacks so we didn't have to get really deep into the law to understand the kinds of problems people were facing and we could get that by talking to customers and then understanding and internalizing the problems ourselves. I just find that you can learn a huge amount by that kind of process, I don't think that you have to come in as a deep domain expert. 

‍

I know a lot of people want this. I know a lot of VCs want this. I would argue that your deep domain expertise should be in the domain of solving these kinds of problems, which is really hard and requires a very specialized set of product and engineering skills, much more than it should be around the domain you're trying to solve for because you can actually learn the essentials of that domain a lot easier than you can learn the essentials of building sophisticated software. I know this is a somewhat controversial point of view, but I definitely feel that pretty strongly.

Sandhya Hegde

Do you remember, like, how many, like potential customers and design partners did you do your user research with in the early days? And what questions did you ask them?

AJ Shankar

Yeah, that's a good point. And I forgot to mention that we didn't start by saying we're going to solve all of the problems in all the litigation for the law. We started on the plaintiff side, which is that firm I'd worked with and other similar firms that had a similar set of needs relatively, not homogeneous entirely, but similar overlapping sets of needs, chipping away at what they needed.
‍
And we had, early on, maybe a couple dozen customers; I don't actually remember the customer count. I know we were doing about $2.5 million in run rate when we raised our series A. That's as far back as I can remember, but we inched our way up there, right? With this ecosystem of law firms that are all kind of frenemies, they work together sometimes, they compete against each other sometimes, but it was a good way to chip into the market and get an understanding of a subset of the user base that we could tackle.

Sandhya Hegde

Got it. And what kind of questions did you ask them as you were trying to figure out, okay, what do I prioritize for product development?

AJ Shankar

Yeah. First of all you figure out a lot of our industry is just about dealing with the data. How do you get the bytes in? How do you get the bytes out? At the end of the day is the question. And so you understand the formats that they're dealing with. It turns out that it's the law so every exchange is custom and negotiable, and you have to be very flexible in how you accept inputs. You understand the kind of problems they're trying to solve for; what kind of documents they're looking for –  you understand their workflows. How do they organize their team to sift through these documents? We would even do ethnographic kind of viewpoints, which is to sit with someone for a morning and they'd let us do it and just observe what they did. Everlaw or not, what tools were they using and how, and that would inform where we would go next. So it's just a pretty deep curiosity about what people are doing with their hours in the day. And we want to make sure we're solving the thorniest, most painful problems and then the problems of highest time magnitude in that day.

Sandhya Hegde

And what were the technical skills you were looking for or building in your team? I’m oversimplifying, yes, it's keyword search, but I feel like that you can have really bad keyword search that ends up in a lot of wasted time, or you can do something that really works for the legal workflow. What was your approach to ‘what is it that you need to build’ in terms of the underlying technology that was specifically like what the legal customers you're working with needed?

AJ Shankar

Yeah, it's a very reasonable assumption that this is just a keyword search engine, and it turns out that is just wholly insufficient. When you have millions of documents that are not friendly – they may not be fully adversarial, but they're not friendly, they're just communications and powerpoints and whatever – you can't rely on keyword search to find the smoking gun. Elizabeth Holmes is not writing, ‘Hey, I'm committing fraud now.’ That's just not how it works. So you actually have to be very creative.

It turns out you have to build a lot of technology. Ingesting this data is very complex: extracting the maximum amount of information from it; imaging; audio transcription; OCR; certainly really sophisticated searches, so complex Boolean queries, then data visualization on any number of dimensions; graph visualization; supervised and unsupervised machine learning, so clustering tools, predictive coding tools we call them; and now generative AI – you have to build a ton of stuff and you have to do it at scale. So significant scale, we're talking individual data sets can be in the order of terabytes. So you need to now think about databases and distributed systems and networking challenges, right? And then complex enterprise software operating at scale. There is no simple solution here. You have to approach this with every tool you have in your arsenal.

‍

So what we look for, what I look for, is people who are very good at systems development. You can't really write a bad algorithm in Everlaw. You will blow up the system, right? If you have an algorithm that's, you know, quadratic or worse in a not confined way, you will pay the price because someone will run it on a giant data set and things will explode.

‍

So, you need people that actually have a relatively solid computer science education at a theoretical level and then the chops to build that kind of into an implementation, which is some amount of systems level programming or application programming that requires an understanding of things like as I said, database access, but also concurrency and scale and networking. And so this, it's a, not an easy order, but for the people that know this stuff, it's a very enjoyable kind of job because you get to apply everything you've learned.

Sandhya Hegde

Makes sense. I'm curious when you think about kind of the status quo your customers had at the time, right? Like a few years in now you have a mature product, you have built out kind of most of the functionality they need, what were you replacing? What did the current status quo look like? And I'm curious if you had any surprises or like big aha moments around, what the impact of this could be for your customers?

AJ Shankar

Yeah. Yeah. So we were the first cloud native solution in our space, right? So moving from a model where you actually had to buy servers and install software, which is painful and hard to maintain, but also not very scalable to the size of data we were then seeing to our distribution model where you log into Everlaw, you have new upgrades every four weeks, and it can scale to whatever you want, because we built it cloud natively, was a very different experience.

The other thing was, we really focused on a consumer grade UX, which is very different from the kind of very clunky tools that were around. And just scale, speed, ease of use were just, I think, were key hallmarks. It's just completely different. We moved the tenor of the experience people had from ‘this is a necessary evil, I have to use this awful tool to get my job done’ to ‘this is actually great and it helps me do my job better and I'm excited to use it’.

And that I think was a very powerful kind of emotional kind of target we had. It's also really satisfying. We can make people go home at the end of the day happier than they were before. And that's very meaningful. And yeah, some of the surprises were recognizing that impact. We've definitely had people come, our customers would come to our own internal kind of kickoffs and, when I remember one of them, just this last year saying, “I would not be in the law anymore if it weren't for Everlaw. This has basically saved my career.” That's very meaningful, and many stories like that.

And then at a societal level, we have definitely hosted on Everlaw multiple cases that have society level impacts, front page of the New York times stuff. And it's, we're behind the scenes company and we're happy with that, but we know those cases are on Everlaw.

We know they make a difference. And, I think we are in our own way proud of that kind of impact that we're doing. It's not just optimizing someone's return on some investment, but actually leading to the kind of impact that can move society forward, I think is really meaningful and definitely not something I appreciated early on, or even anticipated we would have.

‍

Sandhya Hegde

When you think about what is it early customers really appreciated about Everlaw, was it more of okay, we were able to find things that we would have missed, and this helped us get there? Or was it just, okay, the hours in my profession are brutal, this helps me actually get the e-discovery done in a reasonable amount of time, as opposed to taking months and months? Wonder was it more okay, this makes something possible that I just couldn't do? Or was it about just the sheer productivity unlock?

AJ Shankar
It is actually both. We definitely have people say, “We used Everlaw and we found evidence that we'd been looking for for years.” They moved the case on. Yeah, oh absolutely. There's long running litigation so they moved data on and they immediately found new stuff that they didn't find before.

‍

And of course, the day to day thing is not just a series of these amazing aha moments. It is just an improvement in quality of life and improvement in efficiency. Being able to find better evidence is ultimately, faster is really important. Your client doesn't have a lot of sympathy for you. If you don't find the important pieces of evidence and the opposing side does, so just the, there is a lot of that just incremental quality of life that really makes a difference when you think about the thousands of hours people spend on these cases.

Sandhya Hegde

And what was it, did you get any kind of memorable customer feedback that for you really helped shape the product vision?

‍

AJ Shankar

I mentioned some of the kind of life changing experiences that were really satisfying. We have customers send me personal Christmas cards and stuff. And I think I remember one customer sent us a check for extra money, just on top of the invoicing. We just loved it so much. It was extra money. And it was a very, yeah, I don't even remember what it was. It was like $500 or something, but it was really meaningful for us at the time that we could have that kind of impact.

As far as product influence, countless times. We routinely talk to customers and have thousands of conversations a year that impact the direction of the product. It all gets consulted in there. And we have a pretty broad customer base. We certainly sell to law firms, but we also sell to in-house counsel; we sell to public sector entities; state and federal entities like the Department of Justice; and so there's just a lot of different kinds of points of view that are all worthwhile to incorporate.

Sandhya Hegde

Makes sense. You mentioned it took five years to get to the Series A milestone. What was fundraising like? Please share some stories.

‍

AJ Shankar

Yeah. Every round has been very different from the A through the D. And the first one was the hardest by far. When we, when I went out to raise, we had a relatively healthy business and we had a really good product that actually demoed well and people who weren't lawyers could understand it. So I found I would get a lot of first meetings and then second meetings. But ultimately you go to these VCs, they want to have a thesis around your space. They want to understand why they're investing. And most of the time they had never, ever, they'd never invested in a legal company for sure, but they had no thesis around the law and no domain expertise.

‍

And so I spent a lot of time accumulating notes. It wasn't just an immediate reduction. It was like let's come back. Oh, this is really neat. Let's talk about this. And eventually, no. I think I had 29 no's before I got a yes, 29. And it ended up being, I think the best possible yes we could have done. Andreessen Horowitz is a fantastic firm to work with. They, I think saw the opportunity and then Steven Sinofsky, who is a board partner at Andreessen, who was a former president of Windows at Microsoft, former personal assistant to Bill Gates, had been involved in the DOJ antitrust trial against Microsoft and actually understood enough about the law. And that's really what it took was someone to appreciate what we were trying to solve and immediately say, this is worth doing. And so all credit to him for that. And of course I pitched to Mark and Ben and Andreessen and the whole team. And we got a yes. And it's a reminder, at the time it was really tough. We weren't going to run out of money, but we would eventually have. But I felt really bad. I felt like I was letting the team down. It took several months of no's. But I forget, I think it was Henry Ward of Carta who says, this is not a, popularity contest. You just need to find a match. And then once we did they saw in us what we saw in ourselves in a sense. And I think they, I'm very happy to say, I think they've done very well by that investment. Then the subsequent rounds, we just built, everyone kind of had an idea. They had a legal tech thesis. It became a thing that people recognize that legal was a vertical worth playing in.

‍

The series B was, I think 12 days from first pitch to term sheet. Series C, we got preempted. So like it very much changed as we, of course, developed as a business. And then, of course, as people recognize that the law was a really interesting space. But yeah, that first go around was tough and I remember making a lot of trips down to South Bay. I live in Berkeley and I learned a lot from that including what not to do, which is maybe a topic for another podcast, but I definitely learned some lessons about how to pitch well that I have taken with me for sure.

Sandhya Hegde

Makes sense. And maybe coming more up to present day or near present day, right? Enter around 2022 and now there's a lot of chatter about generative AI and how it will impact search and RAG systems for Q&A and there's also a lot of apprehension about what it means for the future of knowledge work in general, right? What does it mean for the paralegals who do a lot of this work that might get automated? I'm curious, if you go back a couple years or so, what was it like in the Everlaw executive team, the board, like, how were you responding? Especially given that you have always been an ML, NLP, computer vision company, like you've always been in this space and suddenly now it feels the conversation has changed and there's kind of a new player in town. I'm curious what the early reactions for you were like, and like, how your view of the opportunity here has evolved.

AJ Shankar

As just someone who's built technology and studied computer science for a long time was just of sheer excitement. I will say I don't think I could possibly have looked at the technology and not thought immediately about the positive implications it’d have, which I think are sundry.

Then of course you think about the business and it is true. It's a fundamentally different kind of technology. As a person, as a team building software it's interesting because it's so powerful and it doesn't require a lot of technical expertise to incorporate.

So that is actually a big deal, like all the other tech we built ourselves, our supervised unsupervised algorithms, we have a patent, like all that other stuff requires deep technical expertise. Here it's making an API call. And so that is a profound change, and you have to understand the implications, which you spend a lot of time thinking about.

‍

I do think our thesis has evolved a bit, but it remains relatively constant that you can't just create a wrapper over these tools to solve problems. I'll just give you an example. I think ChatGPT is phenomenal. But it's really a phenomenal tech demo. I don't use it all the time every day because it's just a standalone app I have to go to. It doesn't have the context I need in my workflow to do whatever I'm doing. When Gmail has their generative AI in Gmail, I'll use it, right? Excel, like that's how things are going to be done. So in our point of view is that Gen AI is going to be everywhere, but it's going to be embedded in the workloads you have in your day to day because you want to remove friction and you want to have the right context and especially you want to have the right guardrails for safe usage, right?

‍

And that those guardrails are context and use case dependent. So luckily, in a sense, you still need Everlaw to do all the other things so that you can use Gen AI at the right moment. And that's how we've approached it. We have multiple touch points in the system where you use Gen AI. It's a very simple, easy to use thing.

‍

You, as an attorney, do not need to become a prompt engineer. And you don't have to worry to the extent you normally would about things like hallucinations, and you don't have to step out of your workflow. You're just there. And for that reason, I would say, as it's always been a truism that AI is a tool, you actually have to solve a problem first.

‍

And luckily, there's lots of problems to solve in our space and AI will be one of the ways to solve them. But I don't think you just come in with a blank box and say, we're going to solve discovery with AI. It seems unlikely now. It might have been possible. I don't think, maybe in five years when the technology is transformed substantially, I don't see it going that way.

‍

So then it came down to how do we use it well? And so we've spent a lot of time understanding how the tech works. We have a relatively strong point of view on the competencies of these gen AI systems, the ones that we want to be able to leverage in our tools and the ones that we don't. And most prominently, especially in the law, the embedded knowledge distributed in a gen AI model is a risky place to go when you're asking very high stakes and high precision questions. It can hallucinate, that's risky. So we want to make sure we don't put our users in a position to exploit it for its knowledge of the law. Luckily, there are many other competencies that it can really assist with and that's what we stay with.

‍

And that building it into the product allows us to craft an experience with those guardrails so they have a lot of confidence in using the system safely. That's worked out well for us.

Sandhya Hegde

And how did your customers react? Because I think there were definitely stories around people using, not Everlaw, but chatbots to generate, do research, end up with hallucinated citations and case facts that never happened.

‍

And so I'm curious, how did your customers react? Has it changed over the last 18 months or so as general curiosity, understanding adoption of what our generative AI does, has improved?

AJ Shankar

Yeah, it's changed quite a bit. Certainly there's a very high profile, embarrassing gaps in the law with kind of solo people plugging stuff into GPT, and then the court's responding, I think, probably too strongly as they're trying to learn about what's going on. I think the pendulum will swing back to somewhere in the middle, and it's evolving in the law.

But certainly a lot of what we do is education. And I actually have done this many times, where we explain to customers and prospects. It's not that these hallucinations are uncontrollable and happen at random. They actually happen for a very specific reason, and I'll explain to you why they happen. And it is possible to mitigate some of the risk here, and reduce the likelihood of hallucinations in the following manner, and here's how we do that. So we actually work through those things, and what's amazing about this technology is that at a high level, it is somewhat explainable to people. You can explain how these models are built, and how they're trained, and how inference works. In a way that it's actually much harder to even explain something like logistic regression, which is a lot of math. And so you can actually explain some of these things, and people come away with an understanding and an appreciation. And I think that education is actually probably, at the moment, a critical part of adoption is just getting comfortable. And so a lot of our work is just that education, which we think is incumbent on us to do as the technology provider to explain, here's how these systems work and here's what we're doing to keep you safe in a sense, but there are still some risks and here's how to deal with those risks and here's why you should check your work and here's how we make validation easy and all those steps are part of the delivery.

It's not just like the API call, right?

Sandhya Hegde

And how do you think it will impact the legal industry over time as these tools get more widely adopted? What do you see changing and maybe give me your most aggressive or optimistic scenario and also maybe you know a realistic one, like, if you had to look five years out like how would what would be the bounding box that they would put around this?

AJ Shankar

Yeah. A realistic scenario is that these tools remain in what various people call the system one or what I call the intuition box of being able to do a particular set of tasks that require language fluency and a certain amount of reasoning quite well. And what that means is we will be able to have better coverage of these million document haystacks.

Right now, we, you just, it's never feasible financially to look through every document. You're going to have these systems actually sift through these documents or assist you in sifting through these documents, helping you construct narratives. Allowing you to do a more comprehensive, better job and basically allowing more truth to come out, which is both parties will benefit in a sense – the justice system will benefit from having more access to the truth.

It will change how people work. It'll allow people to have what we call a smart intern that can work night and day and it will, but you're still going to be wanting to check its work and apply your reasoning and your knowledge, but it will get better and better at its ability to do this context windows, sizes will increase. The reliability of reasoning will increase. So all these things will continue to improve incrementally and I think it'll be just that amazing assistant you'll have.

Law firms could conceivably take on more work. There's an access to justice problem today. It's hard to solve all the problems out there. So it's not like there's a shortage of work in a sense. I would say the other outcomes are much more speculative, involving improving, and this is scientific discovery level improvements to the ability to reason in a way that's currently limited in how these systems work. They have a kind of a very linear thinking process. They have to think out loud. Many people are studying how to improve this with chain of thought and then tree of thought and all the different ways, but the ability to backtrack to explore hypotheses and then ultimately to contain sufficient context to conceivably understand what's at stake in a matter of wholesale, rather than at a smaller kind of document quantum level would be a significant sea change and would really change how the law works. But it would also, I would say, change how most of society works. That ability is approaching AGI in a material way. So I wouldn't even want to spend time on the implications of the law there. I think you'd be wanting to spend time on how this is going to impact society, but that's going to require, I would say, significant possible breakthroughs in foundationally how these systems work already.

There's a lot of work being done to make certain things that are hard tractable. So function calling and using outside oracles for things that – these systems can't really do five digit multiplication. There are very good reasons for this, like how they work and how word embeddings work. They now can, because they'll just write a Python script and execute it, or the call tool from alpha.

As you continue to build these competencies out, it'll be interesting to see how that evolves with the law. But fundamentally, when you look at one of these tasks, in discovery, you have to look at an email between two parties that's an email that has a huge amount of context and ambiguity and understand why person A was saying this to person B.

And that's just not something these systems are capable of yet. But we're moving in that direction.

Sandhya Hegde

Yeah I think I had, remember, maybe this was almost a year ago now watching this interview between the NVIDIA CEO and Ilya from OpenAI, where Ilya said the goal is if there's a mystery novel and the last sentence is, okay, who's the murderer, can the model predict who the murderer is? Which I feel is, in retrospect, actually a very close analogy to e-discovery. But you need all the context and the information, but also a lot of things that are not explicitly written or said anywhere, right? They are implied, but not explicit, which is today something very much a human process, right? Of what's implicit, but not explicit. And reasoning requires that skill set. And it's a big question whether the transformer technology being used today actually gets us there.

AJ Shankar

Yeah. I actually haven’t heard it put that way and I like it. I would say discovery is like that, except maybe the novel is 1 million pages long instead of 200 pages long and there wasn't an author intentionally planting clues so that you could derive the answer. Sometimes it's unknowable. I will say that reasoning by itself is not something magical. It's actually quite mechanical. People have made theorem provers for a very long time. In my area of programming systems, we use theorem provers a lot. But it turns out, I think one of the most surprising things about large language models, is that in order to do next word prediction, you not only need fluency in actual natural language and syntax and grammar and et cetera, you need to have a certain amount of reasoning.

A sentence has to be coherent, a paragraph has to be coherent, a chat has to be coherent. You have to go from one topic to the next. These systems already have a surprising ability to reason to an impressive degree, approaching human level degree in some areas. I would argue that if you look at human evolutionary history, people actually will posit that our cognitive abilities took a big leap with the advent of language, maybe for this reason.

There's a certain amount of reasoning that you can do, but to your point, and what people will call this kind of system two, but deeper ability to reason, explore a state space, not out loud, but in a very sophisticated way. That's the gap. I don't know that you need that to solve a murder mystery, believe it or not. I don't know. I don't want to... these people know this stuff a lot better than I do, to be clear. But I'm willing to bet that you could put a murder mystery into an LLM now, and just based on the cues in the book, it might be able to get pretty far. But I would say that the actual problems in the real world, they're much more ambiguous and a much larger scale, and you may not be able to get that far.

A murder mystery fits in the context window of state of the art large language models today. We have corpuses with over a billion characters of text easily. Those don't, right? So what do you do?

Sandhya Hegde

Makes sense. I'm so excited for what the future of Everlaw is going to be. We'll be tracking your progress, AJ. And maybe to wrap up, I'd love to understand how you have thought about your own evolution, especially as a solo founder and CEO. I think, I've worked with so many founding teams now, and I think being a solo founder is making what's already a very, challenging and lonely journey even more. So I'm curious what that was like for you and, how did you invest in growing as a founder and CEO over time? What would be your advice to founders in the audience?

AJ Shankar

Yeah, it's a great question. I don't want to short sell solo founding. I don't think it's as hard as people make it out to be. Don't cry me a river or anything, right? There are worse jobs. There are benefits. You don't have to spend a lot of time getting alignment between multiple people.

You can align with yourself at any point, anywhere you want, day or night. Taking a walk, in the shower, that's actually really helpful. But yeah, there are times when, ultimately, the buck stops with you for everything. And that's where it's important. And I have recurring chats with other founders, some further along than I am. And I really appreciate their time. Some less far, but that's a very healthy way to balance. But yes the role has evolved substantially in ways I didn't anticipate. In retrospect, it all seems obvious, but at the time I didn't know. I started Everlaw cause I love the product and technical challenge, and I love building.

I love writing code more than anything and I was a programming languages PhD. I could geek out about all kinds of stuff, but that was what drew me to the company. And as we grow, as we grew I should say, the role changes. Right? So one, I stopped making stuff and I started becoming an editor. Right?

People bring me things and I vet them and I give them feedback and I would argue that's a really important role. It should be done well. But as someone who likes making things, it was challenging, it felt a little inauthentic at times and I had to come to terms with the reality that my job primarily now is editing and not making. And then, another thing is moving from making decisions about ideas, which I really enjoy, to making decisions about people, which is really important, but I would say not as enjoyable.

People's careers are at stake.

You have to make a fuzzier call in a fuzzier space about someone who's going to do the work rather than making the call yourself. And so those were meaningful changes. And so there are many more I could get into. You're constantly trying to understand what's next. I've learned many lessons on leadership and certainly on delegation, but on clarity of thought – ‘alignment’ is a buzzword I would have laughed at 10 years ago. Now I think it's probably the most important word in the business. But many more lessons to learn as well. It's just a continuous path of learning from day one, for sure.

Sandhya Hegde

And what's been your approach to learning? Is it, like, you mentioned, like chatting with other founders? Is it books? What helps you actually try to almost be a different human being as your company grows?

AJ Shankar

Yeah, you have to be. I think there's a lot of wisdom within the people at the company. So I talk to folks at the company a lot. I read absurd amounts. I don't really read business books. I know I should. I read other books about everything. Just, I'm a curious person; I'm constantly reading and so that's informed my point of view on a lot of things. And you just stir that in the pot and synthesize it in and that's really helpful, but then I also just am a first principles kind of person. I'm always trying to understand fundamentally why are the reasons I'm doing something, and if I can't articulate them, I don't know that I'm making a good decision.

So I spent a lot of time internally in my own brain getting to walk around with myself articulating a point of view and stress testing it and checking the logic. And that's actually really helpful, just to really understand why I might do something or suggest a course of action. I owe it to everyone at the company to be able to explain why I'm coming. I have that point of view and that exercise is just so valuable and I do that all the time.

Sandhya Hegde

Awesome. Thank you so much for joining us today, AJ. It was really great learning more about you and the journey of Everlaw. And, just cannot wait to see how this evolves. I feel like knowledge work is definitely the most impacted, whether it's code generation, whether it's professional services like law, like consulting.

So really excited to see where you take Everlaw.

AJ Shankar

Thank you so much, Sandhya, and thank you so much for having me.