How We Built 50 AI Agents to Create a Virtual Law Firm

Sean Weisbrot: Arpan Nanavati is the founder and CEO of Cimphony, an AI that's helping startups to get access to lawyers with a much lower cost than going with a traditional law firm. And a human lawyer remains in the loop to ensure that the work being done is valuable and legally binding. In this conversation, we talk a little bit about the future of AI in the legal field, the current situation with AI in the legal field, as well as how he built the technology, how he trained the models by getting access to the data. And a lot more about the business and just working with AI in general. So if you like this kind of technical focused episode, then you're gonna love this interview. Let's get to it. Now, before we jump into the episode, I do need to mention that this episode is sponsored by Symphony. They are the modern law firm for startups. They started because they realized that trying to get. Legal matters handled for startups is quite difficult and very expensive, especially when starting out, and sometimes it's prohibitively expensive. And so they thought that if they set out to use AI to make it possible for startups to handle things much more quickly and cheaply, then they would be able to capture a large segment of the ING startup ecosystem. If you are looking to have help with your legal matters. Then go check it out. symphony.ai. Let's get to the interview. Why is AI so important to small and medium sized businesses and how can it help them from a technological point of view?

Arpan Nanavati: Services as an industry today is about a $20 trillion market. Um, but most of that is consumed by larger corporations or with small to medium sized businesses, not really having an ability to reach higher quality services, in our case, legal services. It's not like you can just pick up the phone, go to your top five law firm and pick a partner and call them and be like, Hey, be my client. The whole network is, is is a closed ended network where you need to have a fat wallet to get an introduction. And based on that introduction. You get access to that high quality console for which you need a legal service for. So, um, I think where AI comes in as AI becomes this, uh, deflationary democratic access to a very high end quality service, which traditionally was powered by humans, uh, and AI really has the ability and has shown the potential, uh, to disrupt these high-end quality services without having. A centralized check and balance in place, uh, but just more wider access and more importantly, cheaper access. Uh, because small and medium sized business don't really have the wallet to hire those high quality legal console, uh, in our case. But with ai, they have the ability to use a high quality. AI legal counsel, uh, which gives them similar to sometimes even better, faster results, uh, than they would get from a traditional lawyer. Not because lawyers are bad, it's just the, the system, uh, is broken. Uh, my parents are lawyers. Uh, as an example, but just the, the business model of, of legal services and law firms is based on an hourly model. Uh, and it's, it's all about how can you increase the billable hours, uh, which means you need a very, very fat wallet. Uh, you even pay for things like reading an email. Uh, Hey, can I get on a call? Type of email you'll get charged for it. Uh, so just things like that. I think that business model is, is, uh, ripe for disruption. And that's where AI has the power to cut through a lot of that slack, uh, and put in the right human supervision to give the desired high quality results for small, medium sized businesses.

Sean Weisbrot: I have to ask, how do your parents feel about this?

Arpan Nanavati: It's an interesting conversation. You know, I think it's a generational gap, uh, uh, in many ways. Uh, it starts with being in denial. Then it starts with being. You know, the traditional, uh, way, way, way of disruption is it's denial. Then it's the making fun phase. Then it's the slowness phase, then it's the adoption phase. So I think we're still in the denial phase, not just with us, but just the services industry and overall, uh, AI wanting to replace the services industry overall.

Sean Weisbrot: I love the idea that AI has the potential to enable a lot more innovation faster because the cost is so low because you're basically paying for API calls, which. If you're a business and you're working with another business and you're, you're using their APIs. The APIs are generally pretty cheap, thankfully, because their focus is on scale, which is also how a lot of AI startups have been able to grow from zero to 10 or a hundred million ARR in a year, so often and so fast in the last year or so. If your goal is to help startups have access to AI for the purpose of making their legal work easier and cheaper. How do you justify that with your investors and your team? Like, our goal is to make this cheaper. Our goal is to make it easier, but we have to make money too. So how do you, how do you balance those two things? Trying to democratize and, and make something cheaper while trying to grow revenue?

Arpan Nanavati: What does this mean for the traditional SaaS model? In general, which has traditionally relied on high ACVs and longer sales cycles, et cetera, et cetera, type of stuff. I mean, again, this is my personal opinion. Uh, the, the answer is still outstanding. Um. But what, what we, what we are seeing personally, we've moved away from the SaaS model, uh, and it's a pure transaction based model. Like you pay for the outcome. You pay for alignment of the incentive with what the customer needs rather than just paying for something when the customer doesn't need. Uh, and number two is managing the investor expectations. I think this is, again, going back to the. Overall, we see industry for the last 10 years has been, has been heavily focused on SaaS for a good reason. SaaS was a very lucrative model, high ACVs, you can tell in the larger corporations, et cetera. Um, uh, in our case, we still have high margin. Uh, the top line is more of a game, and that's why we picked an industry which applies to everyone. That means the. The, the area, the blast radius is much larger, which still allows us to grow, uh, and create a market which is a very large market, at a very healthy market, if that makes sense. Like everyone has a legal need, whether it's a personal need, whether it's a business need, um, um, whether it's a compliance need, and, and that means today everyone in one way, shape, or form is, uh, is consuming that legal service so purposefully. Applies to a very large population, uh, which in turn means it's a larger market size. It's a very large time. Uh, and the fact that the cost of our infrastructure is low compared to the human powered cost, uh, of the traditional model, uh, it allows us to generate really strong margins of 75 to 80% margins.

Sean Weisbrot: How can you trust AI to work in a field like legal? When we've had clear examples recently of AI hallucinating, legal precedents,

Arpan Nanavati: you can't really mess up, uh, even 0.1%. So the way we handle it is, um, on the platform, on the infrastructure side. Um, our platform is extremely capable to not have any hallucinations. And the way we do it is we've digested, uh, uh. Data, uh, uh, and case law from the last 50 years, uh, dating back to 1975, uh, for example, across Supreme Court all the way down to the county court. What that gives us and our models is an extreme ability to give specialized feedback and avoid hallucination. So the model hallucinate when they don't know what to do, but they're still forced on the response. In our case, it, it doesn't happen. And number two is, um, this is still a regulated industry, which means any legal advice still needs to be given by a certified bar certified lawyer. Um, and the way we do it is human in the loop, which is a very commonly used term within the AI industries or regulated AI industries where you'll have a human. Just to make sure there's no hallucinations, there's cross checking, there's double checking, et cetera. It makes sense in our industry. And, and we adopt the same model where we have a human lawyer who will oversee the output of ai, uh, to make sure it's, you know, accurate. And it comes with the official legal advice, which in turn, you know, this is all about establishing trust with the customer. Not that the customer doesn't have trust in ai. I, that trust is be earned. 1, 2, 3, 4, 5 years. Once they start seeing many, many transactions with us, the trust for AI becomes higher and higher and the need for the human becomes lower and lower at that point in time. Um, but we've taken care, like I said, on the infrastructure side, we have a massive training data set. Legal data set that it's given to our models to avoid hallucination. We have case law from last 50 years. We pair that with a license attorney, uh, who sits on top of this. Now, I think this is a very important point. Um, generally when AI products, uh, sell human in the loop, what they're selling is the customer becomes the human manipulator, which is the customer still self servicing the ai. Which may make sense in a lot of, in, in legal indu, customer generally doesn't know what they need to do. Uh, like as a founder, you don't really know what the legal needs are, leave alone their legal ramification. So what we have learned is our human in the loop is better suited to provide the service to the user. Uh. The customer being the human in the loop. Uh uh, so I think that's a, that's a key learning that we've had in the last 18, 20 months that we've started, is um, customers really want us to be the human in the loop, uh, versus them being the human in the loop, which makes sense.

Sean Weisbrot: So I wanna go back to your training data set. I was reaching recently, approached by. Uh, agency, I don't know what you would call them, that was trying to get me to license my podcast content to be trained on an LLM for like business focused content. Did you license this? Did someone have this data set available and you licensed it, or was this publicly available and you're able to access it? Like how, how did you find this data? How do you know that it's good?

Arpan Nanavati: Because your legal data is open source data. Uh, legal court systems, um, your regulations. Constitution, all of that is open source. Uh, and by open source, you can find it, you can look at it. These are government websites, uh, that host this information paid by, uh, the taxpayer money. So within the legal domain, uh, data by itself is open source, and data in itself doesn't become the model. It's the application of that dataset. Uh, it's the application of how you tune the models and the fact that you're building a tech layer. Between the customer and take abstraction layer between the customer and this dataset, which are fragmented, frankly speaking. Uh, sometimes they're even handwritten, uh, in many cases. So, uh, it's the ability to digest that dataset in a digital fashion and apply that set towards the outcomes. That's where the mode comes in. So execution is the real mo is what I tell people at least. AI legal domain data is definitely not the one

Sean Weisbrot: I've been building my own software that I'm trying to launch soon. Not looking to get it fundraised, 'cause I did that before and I don't wanna do that again. But I realized that part of what I was doing could be enhanced if I had an AI that could ingest the data and. Allow the user to communicate with the AI about the data, and I'm using AI to develop the product and the AI is really good, but in this case it's really struggled to help me to build what's essentially an agent that has this separation layer you are referring to where. You have the database and then you have the chat bot, and then you need an agent in the middle that separates them so that it can take in the data, assess it, analyze it, and then be able to answer questions about it. Something like this. Was that the most difficult thing for you in building this? Was this, this agentic layer?

Arpan Nanavati: I wouldn't call it difficult, uh, but it's definitely a much needed layer. Um, and, and to my previous point that. Execution is a real mode, which is that abstraction. There is, is our real mode. Uh, anyone can build a wrapper, an AI wrapper, and you know, from off the shelf models and kind of build a POC, but for production use cases, uh, and, and in legal use cases, which are highly, highly regulated, there's just no room for error, high risk use cases, uh, the abstraction layer becomes really important. So what we've done is we've built. Uh, uh, about 40 to 50 agents that are highly sized for their particular use case within their particular work call and what our infrastructure looks like is. Is a team of agents similar to what a law firm would have, where you would have X, Y, Z person who specializes in A, b, C law, and then they have a team of five to seven people underneath them who do different things. Like one might be doing research, one might be doing drafting, one might be doing cross checking. One might just, you know. Doing X, y, Z use cases. So that's how we've structured our agents similar to functioning, like what a human function would do. And what that allows us to do is give very fine tuned, very specific, uh, unlocks back to the customer. Because like I said, legal is a very high, uh, risk domain. And, uh, and yeah, that's our, that's our mouth is that software layer that we've built on top of these. Fine tuned models, uh, and our fine tuned infrastructure, which consists of these, uh, highly specialized legal data sets.

Sean Weisbrot: Hey, just gimme 10 seconds of your time. I really appreciate you listening to the episode so far, and I hope you're loving it. And if you are, I would love to ask you to subscribe to the channel because. What we do is a lot of work and every week we bring you a new guest and a new story. And what we do requires so much love so that we can bring you something amazing. And every week we're trying really hard to get better guests that have better stories and improve our ability to tell their stories. So your subscription lets the algorithm know that what we're doing is fantastic and no commitment. It's free to do. And if you don't like what we're doing later on, you can always unsubscribe. And either way, we would love a, like if you don't feel like subscribing at this time, so you're referring to abstracting this away from the users so they don't have to see it, which I, I think is extremely important. Um, and something that I'm trying to do, even though it seems like what I'm doing just needs a single agent, I could be wrong. Uh, maybe I'm, I'm not looking too deeply enough yet at what I'm doing, but. To have so many agents I Is it common to, to need so many agents or were you just trying to like digitize the illegal office basically and trying to just get everyone involved in replacing them?

Arpan Nanavati: Most of the startups, we ourself come from tech and we've been part of other pro, you know, startups. We talked to other founders, all our friends are founders, et cetera, and they getting crushed by 20 k invoice on. The question came up is what if this legal work that they're getting charged for can be done by ai? Uh, and and that's where it started this journey of us trying to answer that question. And the first attempt of that was to just take off the shelf models, try to attempt to do some of this legal work. And obviously we failed to what you're, uh, noticing right now. And, and that's where. We spend more time on the architecture and design, which is how do these elements really work? Agents by itself is a cool concept, but can we, can we get the agents to do what we want in a repeatable, predictable fashion? And, and that's where spend a lot of time, uh, proving this tier and building this infrastructure out. Um, I don't know what others are doing. Um, I've seen cases where. Where it's single I, it just depends on the use case. Use case, uh, within the legal domain. If anyone's building within the legal domain, I would highly recommend that. Training and training their models to then building this specialized abstraction layer, uh, to get the quality of the service, to get a very high quality of the service for, for the, for the customers.

Sean Weisbrot: What has been the most exciting thing for you about this whole process?

Arpan Nanavati: It's the adoption. Uh, frankly speaking, you know, I think we underestimated. The ability to gain trust with the customer. Um, personally, I didn't expect our growth rate. Uh, and, and the growth rate just speaks to that the customers are willing to give trust, uh, on a highly regulated, highly risk risky environment, product, or pro service, uh, is, is one. Number two is the, the, again, the appetite for. Services to be disrupted by, by software. AI is just software at the end of the day. Um, uh, and third I think, is again, a pleasant enterprise of folks. Really, uh, aligning or accepting, uh, the fact that they wanna pay for their outcome. They don't want to pay for renting fees, which is being the traditional SA one, right? Like you pay a monthly fee, you pay an annual fee regardless of how many times or whatever you're using the product per seat, pricing, all of that. I think the consumers, our consumers, at least what we're seeing right now, um, is they're fatigued by that. That model and they just wanna not like, Hey, I'm doing X, Y, Z, I only wanna pay for X, Y, z, I don't wanna pay for your servers that are just running, um, on the site. Makes sense, right? Like that's how you, that is the service model, right? Like you consume, uh, a service, you pay for the service when it's done. You don't pay for it for perpetuate.

Sean Weisbrot: That's something that I thought about back in 2020 when I was doing my last tech company. We were building something to compete with Slack and we hated Slack's business model, which was we're gonna charge you to enable being able to access all of your chat history, which I thought was a disgusting model. It's like, wait, you're telling me if I don't pay, I can't have the entire history of my team? Like I thought there was a better way. That was one of the reasons why we wanted to go after Slack, but we went back and forth internally about whether to have a usage based model. Or a flat fee, and our conversations with potential customers told us that they didn't really care. They would pay either because the way Slack works pisses them off. If we could do something better, they would pay basically whatever we wanted them to. I think that's a very specific example, but I loved the idea of usage. I still think it's a great idea, and I think I may implement a usage based model for the new software that I'm trying to launch as well.

Arpan Nanavati: It's actually not usage based model. It's outcome based model, which usage is a precursor to the outcome, right? We don't charge until the desired outcome is done, so we don't charge for just. Partial outcome as an example, and that's what the customers like. Is, is, I think that's the differentiation that we want to drive, is usage based is, is a common metric, which is you are calling X number of APOs or you're consuming wide number of tokens, but what, what outcome did that give to the customer? Customers don't know that. What customers wanna know is like, I want X, Y, Z outcome. I wanna know it's gonna cost A, B, C and I'll pay it to you once it's done. So that's how we, we price our, our, our work is anytime there's work coming in, we're able to estimate, um, based on our pricing models, et cetera, what the cost will be. We tell the cost of the customer upfront. They agreed to that cost. When it's done, they'll get an invoice, which they'll pay.

Sean Weisbrot: So you, you automate that pricing strategy and, and offering to them. How do you come up with that?

Arpan Nanavati: A lot of the legal work, uh, in that sense is commoditized, at least the transactional work. We, we don't support litigation work. Like there's a lot of things that we don't do. Today, litigation is, is an example. Uh, we don't deal with like fund formation for hedge funds and et cetera type of stuff. So there's a lot of things we don't do, but there's a lot of things we do. Uh, and a lot of things that we do is, is a classic. Uh, day-to-day need of a startup founder, which is, I have an idea, I wanna set up my Delaware C Corp. Uh, I wanna do all my post incorporation docs, uh, then I wanna start fundraise, I wanna start hiring people, I wanna manage equity, et cetera, et cetera. So we do that really well and, uh, we are able to guesstimate. Safely, uh, with the room for buffer as to what our pricing should be. And obviously it's never perfect, um, but we try to eat the cost. Anytime when it's not perfect, the customer just gets an upfront cost. They never see any surprises on the invoice.

Sean Weisbrot: I think that's kind of the best you can do, right? Like I, I see with the ais that I use, that they will charge me a, a credit for a message and if. I don't get the response that I want, and I, and even though I know I'm prompting it correctly, and it takes me 20 more credits to get the desired result, I'm still being charged all 21 credits, but I feel like I shouldn't be paying for the AI's inability to do the job that I've asked, especially when sometimes the AI hallucinates or the AI tells me it's done something and hasn't done it. I can prove that it hasn't been done. This was a big reason why I, I was using lovable before and I got away from lovable because it was hallucinating. The, the deeper into the project, the more it hallucinated to the point that it was rewriting ui, it was removing pages from the route configuration. It was all sorts of. Ridiculous stuff. And so I was like, I can't, I can't justify paying for this, knowing that it's, it stopped giving me what it was giving me before it used to be good. And the more I use it, the worse it gets.

Arpan Nanavati: Yeah. Personally aligned with the same model, which is, it's a, it should be an outcome-based model. We're in the service business and service businesses, uh, are, are all about you get the service and you only pay for the service if you're satisfied with the service. Basically an outcome if the customer, client. Uh, got their desired outcome, then they pay for it. So we truly believe in that model. We're not a software company, we're a service company, is what I.

Sean Weisbrot: This is what I'm doing as well. I've built a software to enable me to provide a service rather than just outright having a software for people to self-serve. I think you can make a lot more money a lot faster because you could charge more by having that human in the loop to provide the service. And I, as of now, I'm the human in the loop, but eventually I may ha, I may develop an agent that's much more thoughtful, that's able to actually run the entire service on my behalf.

Arpan Nanavati: Again, it goes back to industry by industry. Within the legal industry, we, we need the human in the loop and the human needs to be a certified lawyer. I'm not a lawyer. Um, I can't give legal advice. Uh, for example, t can't be the human in for the client, but my partners and my team members or lawyers become that human in the loop. That way they're able to give and crosscheck and make sure the, the work done by AI is backed by the lawyer is backed by the professional liability rules, et. So again, it goes industry by industry is what I would say. Within regulated industries like ours, we definitely need the human in the loop and customers frankly deserve that, right? If you think about it, law firms will charge anywhere from partners at law. Firms will charge anywhere from eight 50 to $1,600 an hour. Um, and with us they still get the similar type of quality at $10 an hour, um, and type of cost. Which is incredible savings, uh, for a similar type of quality, what they could expect from another EM one.

Sean Weisbrot: Is there something I haven't asked you that you feel we would be missing out on if, if you didn't share?

Arpan Nanavati: I think where does this go? Um, is is the great question. Um, uh, I think this was the, the preamble where, where we first started talking is, is where does the services ministry go in general? Uh, and to my point that we're going through a massive change. Customers are more than willing to adopt AI and trust ai. We see that day in and day out, and corporations and businesses that are willing to adopt ai, uh, not only willing, but willing to be native to AI are gonna be the ones that personally, in my opinion, are gonna come out as winners over the next decade. Uh, versus trying to establish friction, uh, within the services domain to try to not use ai, uh, for various reasons that they don't have. So we're, we're going through a massive change. I think, uh, services in general will become highly deflationary, net positive for the end user. Uh, services will come really fast. Business models will be more outcome-based business models. Uh, all in all, it's gonna be a massive win for, for the customer.

Sean Weisbrot: This episode was sponsored by Symphony. You should go check out their website, symphony.ai, the modern law firm for startups. Go check them out now and they will help you square away your legal matters when it comes to documents for your startup. Thanks again for watching this episode. I look forward to the next one.