Legal Tech StartUp Focus Podcast
The Legal Tech StartUp Focus Podcast covers the startups that develop and sell legal tech products and services. Through interviews with legal tech startup founders, investors, customers and others with an interest in this startup sector, the podcast's host, Charlie Uniman, and his guests will discuss such topics as startup management and startup life, startup investing, marketing and sales, pricing and revenue models and the factors that affect how customers purchase legal tech. In short, the Legal Tech Startup Focus Podcast will focus on just what it takes for legal tech startups to succeed.
Legal Tech StartUp Focus Podcast
Purpose-Built AI Tools That Genuinely Help Litigators
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Fifteen million pages of litigation documents are not a “busy season” problem. It is a systems problem. We sit down with Tiago Luccini, CTO and co-founder of Theo AI (https://theoai.ai/), and Sarah Johansson, Head of Legal Product, to unpack what it takes to make AI genuinely useful for litigators and in-house legal teams without turning every task into a prompt-writing contest.
We walk through Theo AI’s two product lines: a claims and litigation management platform for corporations that helps with document intake, analysis, and at a glance reporting for legal leaders, plus a focused solution for MDL and mass tort matters where scale and repetition can overwhelm even the best teams. Along the way we talk about why niche legal tech can outperform general purpose chat tools, how “AI for the service of the human” changes product design, and why verification against source documents is non-negotiable when lawyers stay accountable for the work product.
Then we go deep on agentic AI. We define what agents are, why swarms and configurations can create AI fatigue, and how Theo AI tries to hide that complexity while still delivering automation that triggers off real workflow events. Finally, we tackle AI security and privacy: read-only preferences, blocked internet access, auditability, and even a kill switch for single-tenant deployments. If you are weighing buy versus build for legal AI, we also cover the ongoing cost of maintenance, testing, and keeping guardrails current.
Subscribe for more legal tech startup conversations, share this with a litigator or GC who is buried in documents, and leave a review with your biggest question about using AI safely in litigation.
Welcome And Guest Introductions
SPEAKER_02Welcome everyone, loyal listeners to the Legal Tech Startup Focus Podcast. As always, your host, Charlie Uniman, and I'm very pleased to introduce two guests today. Each is from a legal tech company called Theo, T-H-E-O-A. First, uh the uh uh Theo CTO, I think I have that correct, and and uh a co-founder, Tiago Luccini, and also the head of legal product at uh Theo AI, uh Sarah Johansson. Welcome to uh each of you. Thank you. Thank you. Thank you for having us. Uh pleasure is all mine. Um as we often do in the Legal Tech Startup uh focused podcast recordings. We uh we begin by asking for a little bit of background about our guests, how they got into Legal Tech, and uh from there we're gonna segue into Theo AI, uh, particularly finding out what exactly it does. I've had the benefit of speaking beforehand to uh Tiago and Sarah, so I know a little bit, but they're gonna educate me further and educate all of you, and then we're gonna see what uh its secret sauce is, how it distinguishes itself from its esteemed competition, and then get into more general topics. So let's start uh ladies first with uh Sarah. How did you get to uh the OAR?
SPEAKER_00Yeah. Thanks for having us, Charlie. Uh and my background is is highly legal. I had my I got my law degree at Queen Mary University of London in the UK. And then I worked in commercial litigation in London for a boutique firm and um was mesmerized by sort of the intensity and ebbs and flows of litigation in general. Um and as my next step, I took a little turn. I did an LLM here in the US in national security law, which I also found really, really interesting and and um sort of engaged my policy brain and and my international relations brain, which was a lot of fun. But um ultimately I wanted back
How A Litigator Enters Legal Tech
SPEAKER_00into the hustle of bustle of litigation. And I was introduced to Tiago and Patrick in the early days um of VEO in sort of a research capacity, and was thrilled to be looking at cases again, and um got more and more involved and ultimately uh started leading product for us with with my litigation background as we're building. Uh I'll get into it a little bit more, but tools for litigators. So that's my joke.
SPEAKER_02I may have I I may have joked with uh both of you when we last spoke, and I'm sure my podcast listeners have heard this, but uh when I was a corporate young corporate lawyer, every time I thought about becoming a litigator, I would lie down to let the feeling go away. Um but um uh some of my best friends uh uh were litigators and I uh I I regard them very highly. Tiago, how did you get to TheOAI and legal tech?
SPEAKER_01Of course, my my and Charlie, once again, like to thank you for for having us. Uh we are we're we're glad to be here. Um I have uh of course a slightly different background and and uh and uh story um uh because I started in tech, uh, have been in uh uh uh in the startup space for a while. Uh TheOai is uh my startup number five. So I've done it before uh and and explored uh a bunch of different uh areas, different technologies. Um and then even though I have had uh AI uh as a feature or as a as a tool in some of my previous um uh engagements, uh never at the core. And then uh this was one of the things I was uh uh looking for uh after my last uh uh uh exit. Uh I was really looking for a place where um uh AI would be core and uh and and front of like everything. Uh but I I wanted to actually find a space where um uh the AI made sense instead of actually uh doing AI just for the sake of AI or um uh uh so many other um uh uh technical founders uh who end up uh scratching their own each and uh building tools for themselves. Um I was really into uh finding a a market and finding an industry that uh that could benefit from from AI and that uh uh was not necessarily well served. Um and then when uh you know talking to Patrick and then uh uh uh hearing the whole pitch of uh the legal space, like you know, back then we were not even uh you know uh uh uh uh focused on litigation that much. Uh but we're like, you know, what about legal? When like what is gonna happen in the with the legal industry as a whole uh throughout this AI revolution? And that got me very excited. Um uh and uh I think that we started the UAI almost right away.
SPEAKER_02Very good, very good. Well, I I love that lawyers such as uh former lawyers, recovering lawyers, yeah, um uh like uh Sarah are are in the uh on the pitch and and uh helping to develop uh legal technologies. Uh but I uh appreciate especially engineers like you uh who bring an outsider's uh perspective uh to the somewhat often insular world of of legal. And um I I welcome uh I welcome you to the space. You've you've been at it for a while, and it's a wonderful time to be involved with legal tech.
SPEAKER_01It is it is definitely the space that I'm actually uh enjoying the most. Of course, like you always
Why A Startup CTO Chose Legal
SPEAKER_01have to enjoy what you're doing uh uh as you were doing it, but uh, but it's been uh it's been a fantastic ride.
SPEAKER_02Yeah, yeah. A lot of wind behind your sales here. Um so tell us now, uh, with with that description of your backgrounds, um what is it that Theo AI does? We know it uh by now is involved with uh litigation, but more particularly what what does it do in the litigation field uh more generally? Um you can decide how each of you handle that. I'll turn it over to you too.
SPEAKER_00Beautiful. Yeah, so currently we have two products serving um litigators or attorneys managing litigation. One of our products is um a platform that supports corporations in their daily management of claims and um anything even from pre-claim up into uh the litigation stages. And so we support sort of document intake, analyzing the docs, um condensing information in ways that are um make it easy at a glance for a GC or an in-house attorney to have an idea of what's going on. And then because we're looking at all of these cases, compiling that into interesting um data cuts strategy and um reporting that make it a little bit easier, hopefully, to be an in-house attorney managing that volume.
SPEAKER_02And I'm I'm glad you're you're in the in-house space because it's hardly neglected, uh, but it it's a good place to be for a number of reasons, particularly that uh you you'll find, as I'm sure you have uh you don't have that uh dreaded committee of partners uh uh to get through uh when you're when you're selling to the enterprise. And often the enterprise as opposed to the law firm, although it's changing and changing rapidly and favorably, is is more open to tech technology solutions. So that that's one part of what you do. What what's the second product offering that uh um you you're involved with uh at Theo AI, Sarah?
SPEAKER_00Yeah, the second one targets uh litigators at law firms specifically and is a little bit more niche, but helps manage um particularly massator claims, MDL, sort of those larger um repeat uh claims where the that the data vastness is really a blocker to quickly assessing the litigation, understanding where it's going to go and and working through complex problems. So trying to support attorneys um to make quick strategic and um helpful legal decisions there by understanding that that document set.
SPEAKER_02So does the the uh law firm lawyer involved with uh you know uh multidistrict litigation or mass tort liability are they're feeding into Theo a uh a large corpus of documents and you're trying to find the uh the nuggets? Uh uh do the lawyers point uh Theo's AI in the right direction? Is Theo just trained to uh take all that in and and output uh things, a combination of the two? How does how does the latter product work in that regard?
Two Products For Litigation Teams
SPEAKER_00Yeah, so it's a little bit of both. We engage on a project basis for the specific MDL. So sometimes we're working with one law firm, sometimes it's multiple of them in sort of a defense corpus. And we, of course, have um signed all the requisite orders from the court uh and work often with the corporate clients in some capacity too. And um to your point about how we direct it and training, we have done these types of projects several times. And so there's certainly a lot of robust infrastructure in how our agents are built and sort of the pipeline for uh understanding all of this information and analyzing it. But we because it's also on a project basis and it can be relatively self-contained or completely self-contained, there's no training that sort of lends itself between projects, but we can work with the attorneys to address the questions that come up in that particular litigation and really point it very specifically at solving only the problems relevant to the MDL at hand or Mastord at hand.
SPEAKER_02And I would think being, as you put it at the outset of your description of the the second product category, I would think being a bit niche uh could serve as um uh you know, the term moat is uh overused, I think, but I'll I'll use it for want of a better word coming to mind. Could serve, even especially in the face of some of the more uh uh uh well, some of the foundation labs attempts to offer uh something to legal. I I would think that uh aiming what you're doing in the SAC product in particular at a niche area such as uh multidistrict litigation and mass tort and liability type litigation would serve you well in trying to uh you know distinguish uh yourself from from other more general tools and including in the uh purview of general tools, the tools that you know Claude uh from Anthropic and OpenAI and and others of that elk may offer. Is that a correct way to look at it?
SPEAKER_00I think so.
SPEAKER_01And maybe like a no, I can add a little bit of call there as well, because like no, no, no, no, you're you're totally right. Like you're on the on on the money there, uh Charlie. Like it is definitely a moat. Um uh and I would go even uh uh uh two steps deeper because uh uh uh part of the moat here as well is like not only the volume itself, but uh it is uh privileged access, you know, to some extent. And that applies to even both uh product lines, uh, where uh we get access to uh settlement data that uh uh very often is not uh uh broadly available, right? So having having access to that data as well creates a a yeah, moat nowadays is uh is a tricky word, but like some level of data mode as well. Uh but we also have an inexperience mode uh to some extent um uh uh that really differentiates us from uh from the traditional uh uh models and and other providers because uh we uh we build our product in such a way that uh uh we internally we say that it's like AI for the service of the human, not the other way around.
SPEAKER_03Right.
SPEAKER_01Um because uh we we feel that the industry has been you know pretty much like uh us having to sit on a prompt and have to uh it to explain to the AI what it is that we want. And uh we've taken a more proactive approach where uh Theo uh does a lot, uh Theo understands a lot, Theo goes uh very deep on the documentation, um uh uh and exposes all those things back to the user in uh in a more mastigated fashion, if if that's a good word to explain this. Um but but definitely like in in the insights realm or the in the intelligence realm, more than uh uh uh you know waiting for a human to ask a smart prompt.
SPEAKER_02Yeah, and I and and and your ability to do that in in a in a um you know and and demonstrate that expect uh uh expertise that is the models post-trained acquired, put it awkwardly that way, exactly is is is an indeed important part of the MOOC. Yeah. Uh and one that a more generalist approach to litigation or to legal, I I would expect, just can't offer. And would confront the lawyer using the tool with that uh uh somewhat tiresome and often frustrating uh back and forth and back and forth that uh brings the model up to speed. Uh here the you know the post-trained version of what the AI has to offer, I imagine, is already up to speed the minute you open the app.
SPEAKER_01Yeah. And and this is like where uh the niche and the focus really help because like a we can we can go very deep on the specific questions that that kind of MDL uh uh uh uh raise, right? Because they are right there. And they are repeatable. They are repeatable at scale. We were talking about uh hundreds of thousands of documents and uh and millions of pages that need to be uh processed as well.
SPEAKER_02So well, you know, it's interesting. I was I was at a conference, uh Legal Tech conference recently. You mentioned hundreds of thousands of documents. I remember as a young lawyer uh being asked to do due diligence work on uh I don't know how many documents there were, but uh whether it ranged in the 100,000 category or not, uh, you know, we were just with no tools of this kind at all available to us, uh, said go ahead and do it, analyze it, and do it completely and do it accurately. And one of the speakers said, you know, AIs are not human, but the tasks often that were asked of lawyers were inhuman tasks. If I had objected back then when I was a young lawyer, I would have been told to stop complaining and come back to work. Um but it it's so true. Uh you need these tools because what's being asked of any of any group of people, uh law firm in-house, uh to analyze hundreds of thousands of documents, that's an inhuman task. And to expect it to be done without the aid of these tools, especially nowadays, is is is uh uh uh just unfortunate to say the least. And they very well border on malpractice as the malpractice bar develops its tools.
SPEAKER_01And no, you you you are you you are on the money again, uh Charlie, that because like uh it's like even you you almost described one of our uh uh sales slides where we uh we show uh and we do this on a case-by-case basis, but like uh we we try to uh uh uh predict and infer uh the total amount of uh documents and pages that a single case is gonna generate uh throughout the next couple of years. And we say, like, here's like the number of pages, like another 15 million pages that are gonna need to be read and validated. And do you want to go through that? Do you want to do this? Uh and of course, our system, uh uh uh,
The Niche Advantage And Trust
SPEAKER_01which is also part of um uh uh some of our ethos and and and mindset, uh, because some people are scared of like also delegating 100% to AI, right? So we we have been taking a lot of uh care and attention to make sure that uh the user experience that we provide, the the intelligence we provide is always uh referred back to the documents, that you can put them side by side, that you understand what's going on, that you can correct the system if the system got something wrong as well. Uh because trust is a big component of that uh that thing. Like to the extent that like again, talking talking about motes, as you as you mentioned before, uh creating that kind of trust is also part of the moat uh based on the experience that we have had with our some of our customers who did not like AI with their first couple of experiences with AI products.
SPEAKER_02Well, yeah, I mean the lawyer uh he or she is uh ultimately accountable for the output and the work product. Um the uh the the word that every lawyer should have uh burned into her and his mind is verification. And to the extent that the AI tool is designed to assist in that verification is is uh you know vitally, vitally important. And and we're gonna uh I should ask before we get into other topics, how long has the UAI been commercializing its offering? How long has it been out in the market?
SPEAKER_01We started in February 2024. Um, but as uh you know usual in the in the space, uh, we pivoted a couple of times uh because we did start with um uh prediction models um that should then led us into insights that then led us into uh a repeatable volume. Uh so uh with this current product, the product line that we've been uh talking about here, like it's been in the last uh 12 months.
SPEAKER_02Yeah, it's hard to believe, but in uh AI, generative AI uh uh time uh metrics, uh 2024 is perhaps a century ago. Yes. Given all that's happened since then. Uh and and you you operate throughout the United States, I assume. That's right. Uh and uh any plans to go abroad to make the offering available internationally? Is that in especially in the English speaking uh parts of the world?
SPEAKER_01Sorry, you you you you you are our one of our one of the one of the many foreigners in the team. Uh yeah.
SPEAKER_00Yes, but we don't have a timeline for it quite yet.
SPEAKER_02So because I I would think that there are an analogs, well, certainly on the in-house side, there are analogs to the product need that you satisfy in the U.S. And I would imagine too in in in uh abroad outside the confines of the U.S. there are you know similar um uh litigation projects that uh uh uh act like and behave like uh multi-district litigation and mass litigation. So you would you would be uh and London's calling me.
SPEAKER_00Yes, as well as always.
SPEAKER_01The thing that I can say is that like I know uh uh in in the uh in in the in-house uh side of the product, like uh some some of our customers are global uh uh uh corporations, right? Like that do have litigation at a global scale as well. And then once they see uh what we can do, uh they they are also very interested in uh enrolling uh that kind of support globally. So like that that is uh also another uh uh entry point for us of the global marketplace.
SPEAKER_02Well, let's uh no, I think that's an excellent uh uh overview of Theo AI. And and I I do believe from what you've said, the the listeners uh can come away from this with a pretty good idea of what you're doing, and we'll give them at the end of the uh podcast some contact information in case they want to uh figure out how to uh explore uh Theo AI earlier. Uh hint, be careful of the email address. I misused it myself the first time I was reaching out to these people. It's a little tricky, but we'll cover that. Um nonetheless, you, I think it was may have been Sarah, may have been Tiago, I don't remember, had used the word agent. And I want to turn to that next. Now, I'm sure most of our listeners know that that has become the agent, agenting aspect of generative AI and chatbots, or what have been used mostly as chatbots AI tools, is has come into its own prominence. Um but uh tell us what it means uh generally, and and particularly, of course, with TheOAI. And I'd appreciate hearing how uh you know a user of the AI, Theo AI who who uh may not have dipped his or her toe into agent use learns how to um uh get familiar, how familiarizes himself or herself with agent use and and how to use it. So so however you want to split the discussion, let's get into agents and and we're going to uh you know, we're going to talk about um some security and and and privacy issues, but let let's get into agents more generally.
SPEAKER_01And and uh yeah, and thank you for the question 'cause like uh Agents and and you mentioned uh in this AI space, uh, sometimes 12 months is an eternity.
SPEAKER_02Like uh, we we have an internal joke that sometimes even 18 minutes is uh is uh is is already well if you count the models that generate yeah yeah uh with their new version numbers and capabilities, it probably is around it, I'd say 48 hours is a is a century. Um, but go ahead.
SPEAKER_01I interrupt and and then agents particularly like they they they became a huge thing for this uh last year for the whole industry, for the whole tech industry, but like uh basically this this is affecting uh uh every single AI product out there, right? Right. Um uh because it's basically taking the the power of the model uh but but but but taking it to the next level, right? Like the power of the LLM who could uh uh generate uh content and thoughts, uh, but also now calling functions and tools and making decisions along the way, right? So the whole agenc flow, agentic conversation or agent conversations basically by you know giving as much agency as possible uh to the AI so that the AI can make decisions along the way.
SPEAKER_02Uh so let me let me interrupt there. So what I'm hearing is, you know, as an AI user myself, but not not as a lawyer anymore, I retired 12 years ago from practice in New York. But here we are. We started with a generative AI, large language models, say back in in uh uh 2023, 2024. And that that was chatbotting, that was chatting with a bot. You dictated or typed in a prompt, as it's called, and you got some sort of word-driven output back. Yes, and then all of a sudden there came about this power on the part of the LLM, this agency power, where you could not only expect the AI to reply to your prompt with text and often well-formed text. Sometimes uh hallucinations uh appeared, other times uh it was just uh right on on the money, but always very fluent, always very coherent. But now you could ask the the bot, that single chat bot, to to act on your behalf and as you put it very well, call in other pieces of software and access not only tools built into the LLM, but also even other applications. So that's that's in my view at least, step one in agency. Is that a way to look at it?
SPEAKER_01That is a perfect way. Like, no, it it is it is the the the natural evolution from just this conversational bot that uh it's all about questions and answers. But now you can say, go ahead and do this, like go ahead and send an email to my business partner
What Agentic AI Really Means
SPEAKER_01or to this customer, right? Go ahead and read all the documents on this folder uh and help me draft out a Slack message that I want to send. And go ahead and already send a Slack message as well. Um and this this this path like happened in the last uh 12 months, accelerated in the last 12 months, and it became a little bit more accessible because uh it was rolled uh rolled into the the the chat applications. Um uh the the one example that I that I that I give to most users, like you know, if you install Cloud Desktop, um uh you will have on the top left corner as of today, and then they might move it around, but um uh on the top left corner you have like the chat interface, which is more conversational, and then you have the cowork, which is already uh connected with uh your Google Drive and with your local computer, and like it has access to some of your tools and uh uh uh the cloud code, which is uh then designed to even generate code and applications and size and whatnot, right? Uh so but those two uh second bot uh second tabs are a lot more agenc than than the first uh the first tab.
SPEAKER_02The first turn, yeah. And then then, of course, they took the word agent and and they added an S to it. And now you can have more than one agent, yes, uh, on on call to work with uh the the lawyer uh who's using theoai.
SPEAKER_01So tell us it it became like age sub-agents, agents of agents, agents calling agents, network or swarm of agents. I know, I know, dizzy, very dizzy, yeah. So but uh and it it it became kind of a crazy world, and and I think like this is also a good time to to to bring back to the i and like where our customers uh are because our customers are navigating this, and it's an important world to navigate and learn anyway, because it you know it's it's a powerful uh uh uh development for everyone, but also it adds uh uh a uh cognitive load, right? Because like now we need to understand all these things, we need to configure all these things, right? Uh, and then there is also like a lot of uh difference because like you know, the way that I configure my my agent, my agent, my swarm, if that's a little bit difficult than yours, and then we're trying to solve the same problem. Uh the model is gonna react completely different, and it's gonna get to different results. So uh uh uh it is a little bit tricky, uh uh you know, to standardize and to get to uh workable scenarios. So uh a lot of uh people got uh AI burned, AI tired, uh AI exhaustion, I think. Like I don't have a good good word for it, but uh but we hear this in in so many conversations with normal people, right?
SPEAKER_02I think normal people saying, I was supposed to do less work, now I'm even doing more work.
SPEAKER_01Exactly. Like, why am I asking this? And then I have to review all that AI's locked on the other side as well, right? Right. Um, uh so one of our approaches is that um uh uh you know, as I as I mentioned before, uh we try to be the AI tool that does not expose this complexity to users. Uh, it doesn't mean that we don't have that complexity internally, uh which is then uh not only placed to our advantage in the sense that uh um uh uh it's easier for users, like it's simpler for users, but also it allows us to express the problem space and the solutions that we're trying to achieve uh in a more uh surgical approach. Like and surgical is a weird uh adjective here. But what I mean in practice is uh we can have agents that um not only behave like uh normal humans would, but we can then also have agents that are very focused on a specific kind of disease uh and to find that kind of disease on a specific kind of medical record. And then that agent is only triggered when that kind of situation gets uh gets uh uh uh matched by other agents, right? So you can imagine a swarm of agents that are uh agreeing and deciding between themselves what is the best course of uh action and the best way to understand the case uh based on their specialties. And then you can also have like an internal judge that that that that that uh that uh uh decide one way or the other and expose that kind of judgment back to the user, providing a much richer data sets to the user on the other side without you having to ever uh do any agentic work yourself. Like the agent is the agent of the agents are doing that for you, so right.
SPEAKER_02And and so you've you've abstracted that agent work from the user.
SPEAKER_03Exactly.
SPEAKER_02And and they let me see if I understand this. Is it the typical case that you know the the agent tooling that you bring to Theo AI, does the user ever have to invoke it, or is it sort of invoked at the discretion of the LLM and Theo AI, I should say, behind the scenes? Or both?
SPEAKER_01There's a little bit of both, um, but it's not a direct kind of invocation, like it's not the traditional way of like having a prompt and then you have to uh write instructions. Uh it's a lot more circumstantial based on your workflow, right? Uh what do I mean about your workflow? Like if an event happens, uh uh if if a new claim has been filed against you, you don't need to prompt anything. Like, you know that that that is an event. Like, no, they should be processed immediately, right? Um uh it's so our system reacts to those events that happen externally one way or the other.
SPEAKER_02Gotcha. Gotcha. And and of course, um it's a two-sided coin and coin, the agency piece. It's powerful. It can uh lend a great deal of assistance that absent the agency, the LLM or the OAI, wouldn't be able to provide. But on the other side of the coin uh uh are the security and privacy aspects. I mean, if you know, I I uh sometimes wake up uh in the middle of the night and realize that I have something running overnight and uh that something is uh uh an LLM that has access to my email. And I've done my best, uh, but you know, it it it worries me that will it send an email uh uh on my behalf that could could be catastrophic. So so how how have have you and in working with your customers, how have you approached uh trying to deal with concerns the customers might have that these the you know this is not the matrix with swarms of uh octopus-like machines running around outside your control? In instead, we we've built in safeguards and guardrails. How do you get that uh across?
SPEAKER_01So to some extent, like no the the our our our approach of actually uh uh uh retaining ownership uh of the the the way that the agent works uh and kind of like black boxing it a little bit uh is actually a benefit, is it's actually a plus uh for those conversations because uh uh uh we are categorically not allowing uh some connectors that that would be risky, right? Uh uh so we we basically uh uh uh uh uh prefer read-only uh models, read-only agents, right? And then uh everything that is like uh write and mutational, we try to make it into any uh or we aspire to make them immutable so that we keep track of what's going on and we can revert and like go back and uh and do uh an audit as well, right? Like as if you would debug like how a human is thinking, right? Um uh and those are all the tools that uh again, supposedly, uh if you have you're sitting on Cloud Code or sitting on any one of the big uh uh uh uh uh legal products as well that are prompt-based, you would be able to do as well. Like if you you are an engineer and you start tapping into the into the into the brain of the of the of the agent and how the agent is working, uh, very few users are gonna do this or even care. Uh we do this uh not only for every single change and every single uh uh new feature that we have, but also on a recurring basis, uh, since we are like always testing, always evaluating, making sure that our system is secure. So actually, many of our customers, exactly because of uh uh at this approach, they're they're feeling more comfortable in in uh in engaging with us and sometimes even engaging with uh their internal teams for AI, which is which which has been a very pleasant uh uh realization.
SPEAKER_02Yeah, so the uh the fat-fingered lawyer uh uh of which I could uh be accused of of being one, uh it was awkwardly, but uh can't can't tap the button as
Security Guardrails And The Kill Switch
SPEAKER_02I might have done personally here with my uh with codex and and and revealed my entire email to the world by by some uh erroneous click of a button because you you have put that behind uh behind uh uh and abstracted away from it, but put it behind the user's uh uh uh reach.
SPEAKER_01Um yeah, and imagine that like no, just just because like this this is like uh uh uh one of the things that keeps me up an eye to like just make sure that we we provide this uh uh uh uh at the highest security level possible, because our agents are sitting on a lot of uh data that is collected data, PII data, medical records. Like there's a lot of stuff that uh those agents could just go berserk and start sharing with the world and whatnot, right? So even access to the to the internet is something that we that we that we block from our our agents because we don't want our agents like you know searching stuff online or or eventually submitting stuff online because they just decided, right? Um uh and then on top of that, uh for for our uh uh uh more security conscious uh clients that that go for our uh single tenant uh solution, uh, we also provide a kill switch. So again, because uh uh if you are not feeling comfortable at some point, or like if you feel something is being is being leaked or it's going to be leaked, uh we do offer this option that uh uh customers can press a button and then every single file, every single transaction simply becomes unreadable by uh humans or agents likewise. But uh basically it it destroys the system with one button within a couple of megoseconds.
SPEAKER_02And you had mentioned earlier, and I I I know that if you're a you know uh an AMLO 50 firm with an enormous budget, this doesn't pose for that firm as as much of a uh impediment to to buying and building their own AI. But you know, one one of the uh most important aspects of running software such as COAI is maintenance, the constant evaluation, the constant testing. Uh not only product improvement, but also the guardrail and and uh uh type of valves and and uh and and maintenance that has to go on. So, you know, if a firm is going to be building its own, one of the costs that has to be considered after it's built, tested, and deployed is you gotta maintain it. And you guys, and I I suppose it's table sticks for most SaaS uh legal tech companies, is is that maintenance feature that you regard as central to your own uh way manner of operation and and your own business, and you do it every day and every night.
SPEAKER_01Yeah. And I think uh the uh AI has uh we we have always had like the whole buy versus build conversation, right? Like I've been on the customer side as well. So like no, it's it's always a strategic decision, and you have to analyze like no case by case. Um uh and of course, in the in the last uh generation of uh SAS products, uh there was a a big lean towards uh buy instead of build because it was right there. Uh with AI, my feeling is that it's getting even even smarter and even better to buy than build uh because of these things, right? Like, no, uh acquiring talent uh is really hard uh in this space. Uh token cost for RD uh is no joke. Like uh there is a there's a whole conversation in the space and token maxing and whatnot that's been insane. Uh and then you're gonna have to maintain those things and make sure that uh the latest and greatest security um uh uh practices are in place, that you are red teaming properly, that uh your system is not going berserk like by accident, or uh there's so much so many uh meta uh features like once we start employing agents as well, because just as an example, and like some of your uh listeners might actually find it is interesting, but as soon as you start giving agency to agents, like uh the you can also have uh self-healing um uh agents that basically uh make a mistake and then they are corrected by another component in the system and they are like, oh yeah, yeah, I got it wrong, I'm gonna fix it, right? So there is a self-healing behavior, but very often on that self-healing behavior, you're also hiding uh some um uh accidentally hiding uh some concerns that could actually escalate uh out of control down the line. And those are failing completely silent because uh, you know, that self-healing is happening uh uh without you seeing if you're if you don't know where to look, right? Um uh so those are all tiny little concerns that uh that might uh completely get out of control if you're not careful.
SPEAKER_02Yeah, I don't want to go too deeply down this rabbit hole, but uh you know, speaking of hiring talent, um here you are, early stage company hiring great talent, giving them the potential of the upside that comes with being part of a uh uh a startup such as yours. I don't know how law firms, uh uh, since they can't offer that equity upside, uh, are going to compete for that talent. But uh we'll see. We'll see.
SPEAKER_00We'll be watching closely too to see how.
SPEAKER_02I'm sure. Um so we delved into the buy verse build thing, which uh a dilemma that that uh some customers are facing.
Buy Versus Build And Hiring Lessons
SPEAKER_02And and uh we we let's go further a little bit down the business side of things. We we often conclude the the podcast episodes with some business advice you've been the business in the business of offering a SaaS legal tech product for for a bit of time now, uh eons in AI uh uh uh epochs as measured. Uh what other business advice might might you give to our listeners, some of whom are uh many of whom I would say are legal tech startup leaders themselves.
SPEAKER_01Sorry, you had some interesting ones. You're gonna go first?
SPEAKER_00Sure thing. Um Charlie, one of the things that's been on my mind particularly has sort of after having been in a law firm, thinking through how are we meeting our customers where they're at. Um one of my interview questions super early on was about um who how to prioritize. And I was given a bunch of options, and the scariest person seemed to be the managing partner of the litigation team, and the the truly scariest person is the court. And I think that's a reality that sometimes we forget about when we're thinking about our customers as litigators. They might want to fix their workflows, they might want to do things in a new way, and they have clients asking for it and everything, but ultimately court deadlines and and their service to the court is at the forefront of their minds, and everything else is is tertiary. And so um as I think a lot of your listeners will identify with the legal sales cycle is is tough, but uh showing sort of building relationship and um showing interest in that reality for litigators has proven to be very useful for us and allowed us too to get an insight into what struggles they're facing and what why um the the acquisition of AI may be a much slower process than we might anticipate, seeing you know the how every person and their grandmother is using sort of Claude or or the newest uh thing to prompt for themselves in the legal realities, it's really different. But we've sort of to my point about our relationships, that's worked really well, and that's made hiring a challenge too. And and if I could give any advice is being conscientious about hiring the right type of person, not just the right skill set, um, because the people that will then touch your customers matter a lot to how successful you can move your how successfully you can move your customers along and and uh rise to the challenges of of sometimes not seeing contracts close as quickly as you want them to, but still having to build something phenomenal and trusting that you're on the route, uh on the path there.
SPEAKER_02Well, you know, you make an excellent point, especially as we think about people like anthropic and and uh open AI, Gemini, uh perplexity uh trying to come into the legal space. Look, God bless them, um competition is good. Um and uh you know, they may be the appropriate uh uh uh tool for the uh particular kind of law firm or NAS department. But let's let's face it, they those those people, with perhaps a few exceptions, I think it was OpenAI that had just hired a very well-respected senior person outside of a uh a legal tech company. But generally, they're not gonna have the appreciation that uses expressed for one of the concerns that litigators have to deal with, or if you're a corporate lawyer and you're doing public markets work, you know, you have you're always thinking about the SEC. Um and and well, they're not a court, they're a very serious regulator. Uh they're not gonna be able to, you know, the these lab models are just not gonna be able to bring, at least as of today, that kind of sensitivity to product build, to post-sale customer success that someone like you as a former litigator um uh and others, you know, in the legal tech space per se, uh uh can bring. Um, you know, I as a corporate lawyer wouldn't, a former corporate lawyer wouldn't have had the sensitivity to the court and what that means for litigators that you, Sarah, would have. And I might have had a better appreciation for what the SEC might mean for a transactional lawyer and how that concern about what the SEC would think of my work and react to my product, uh, I'd have, but I wouldn't be able to bring to the work that you guys are doing. So I think that that's a very good point, especially when it comes when it comes to hiring. Well, uh, you know, I I uh I want to um uh jealous of our listeners' times uh uh time, I'm gonna bring the podcast to a close, but I want to thank you both very much for an excellent discussion of not only what the AI does, but uh, you know, more generally, some of the things that we talked about, build versus buy and hiring uh the right kind of people and and agentic AI. Uh uh thank you, both of you, Sarah and Tiago, for for being guests. If people want to find Theo AI or reach out to you, uh what's the best way for them to do so?
SPEAKER_00We're both on LinkedIn, uh Sarah Johansson and Tiago Luccini, so feel free to look us up there. Um Tiago, do you want to give our further details?
SPEAKER_01Yes, further details, like an uh finding Theo AI, and back to Charlie's point, like our our domain is uh is uh is a little bit tricky because it's uh theo t-h e uh o a i.ai. So we're
How To Reach Theo AI
SPEAKER_01gonna be.
SPEAKER_02That's what I uh that's where I blew it on the email. So I was reaching out and didn't get it.
SPEAKER_01And there is a theo.ai as well. That's not it. Yeah.
SPEAKER_02So I found out. Yeah. So everybody listen, the-ai.ai when you're gonna visit the website.
SPEAKER_01Uh of course you can learn a lot more uh there. Uh you can contact us there as well. Uh and um thank you, you Charlie, and uh, for inviting us and your listeners for the for uh spending a little bit of time with us.
SPEAKER_02Yeah, no, I I I I welcome having you and welcome the chance to talk with you uh on air. And as I say to everyone, and I I uh very much enjoyably say to you, when you're out in my neck of the woods in New York, come visit. Yeah. Uh and uh you're where out in uh is it San Francisco?
SPEAKER_01So our our office is in San Francisco. We we are uh that's where most of uh our folks are located. Uh we did start uh uh in a distributed fashion at some point, so we have people all all over. Sarah is in Seattle herself.
SPEAKER_02Uh-huh. Okay, wonderful place. So again, thank you. Uh let me know when you're out east. I'm a very short way away from the city. Love to uh to get to you there. And uh again, a most uh enjoyable uh podcast to record with you. Thank you.
SPEAKER_01Thank you.
SPEAKER_02Likewise. Thank you for listening to the Legal Tech Startup Focus Podcast. If you're interested in legal tech startups and enjoyed this podcast, please consider joining the free Legal Tech Startup Focus community by going to www.legaltech startup focus.com and signing up. Again, thanks.