--:-- Sarah Guo: Hi, listeners. Welcome back to No Priors. Today, Elad and I are here with Jake Stauch, co-founder and CEO of Serval. Serval, with all the speed of the African wildcat, is trying to attack ServiceNow and more broadly, bringing the power of code generation to IT and enterprise teams overall. Welcome, Jake.
--:-- Jake Stauch: Thank you. Great to be here.
--:-- Elad Gil: Yeah, maybe you could tell us a little bit more about your history and how you started Serval. What is the problem that you started with, or why did you start working on this to begin with?
--:-- Jake Stauch: Yeah, so background: Serval is enterprise automation, enterprise service management. The idea is you can ask questions and get help at work and automate a lot of the parts of the job you don't enjoy. We do that through code-gen agents. That's what the company does. We compete with companies like ServiceNow, which is historically the big legacy provider in the space.
--:-- I discovered this category actually working at a previous company called Verkada, where we sold enterprise physical security products to IT teams. I spent many, many years talking to IT teams, figuring out what to build for them. One of the things I love to do with these teams is ask them, "What do you spend the most money on? And what do you hate the most?" ServiceNow would top a lot of those lists. I was always interested in this category, thinking maybe there's a better solution here. But it wasn't clear for most of my time there how you'd ever beat ServiceNow. They're a great company. They've done so much so well.
--:-- Elad Gil: What was frustrating for users about it?
--:-- Jake Stauch: It's so hard to build in the platform. It's a very robust, customizable, powerful platform that can map to all these enterprise workflows. Fundamentally, it's a workflow automation platform. It can do so much, but every idea you have for doing something cool, you have to go and implement it. A lot of times you're bringing consultants to implement it, so it's always running six to 12 months behind where you want it to be. It's fine if you're building something that's going to last five or ten years, like a big onboarding flow where nothing about your business is going to change. You can invest all that upfront effort, it gets deployed, and it works for a while. But for just trying to keep up with all the automation you want to build, ServiceNow is not able to catch up. And then it's very expensive and keeps getting more expensive over time. That frustration—it's not keeping up with where I want to go as a business and it keeps getting more expensive—that's what was driving a lot of the frustration.
--:-- Sarah Guo: I think one thing that you mentioned that's really interesting about what you do is if you look at the world of AI adoption, a lot of the really early adoption is coding agents. That's engineers using it to build applications for their business or products. What you folks have been doing is really pioneering expansion of the category in terms of saying we should let anybody internally—an IT person, et cetera—build internal products or tools that can be used broadly across the enterprise. You're basically providing coding agents for all the internal stuff for people who often aren't that adept at coding. Can you talk more about that?
--:-- Jake Stauch: Yeah, we love this idea of coding agents for the rest of us. So much of the early enterprise adoption in AI has been engineering teams. If you think about why that is, there are a lot of reasons—early adopters, obviously—but also they're able to build and experiment with these tools because they have this process of code review. It's very safe for even a 19-year-old intern to build something very powerful and then go through this process before it goes into production. We want to bring that same capability to the rest of the enterprise.
--:-- We build a platform that brings coding agents to IT departments who then can bring it to HR, finance, people teams, and legal teams. They can do it because under the hood, it's coding agents. But then you have all these abstractions, guardrails, and controls. Things go through approval, a staging and approval cycle, audit trail, and all of that. It's about how you build a system that makes it easy for everyone to be a builder. The core insight here was that IT actually wants to build stuff. For most of the world, they don't have an engineering department; they have an IT department. Most companies have an IT department, not a team of software engineers. IT is what's building and configuring technology for the company. We're building essentially the AI platform for them.
--:-- Elad Gil: How does that run into things like all the different workforce agents that are being developed, like DOT from OpenAI or some of these things where you almost have a virtual assistant that will go and do things for you?
--:-- Jake Stauch: I think those really work well for consumers and SMB, and even a lot of the mid-market organizations. In a large enterprise, there's so much process around approvals, around what you can and can't do. I didn't even really appreciate this going in. The permission sets of any individual are so limited that even if you are allowed to bring your agent to work—which in a lot of cases you're not—what it can do is incredibly limited. It can barely even read your emails. When you think about all the stuff you need to do at work, all the systems you want to touch, and all the interesting automations you could build, your personal permissions end up being so limited that these personal agents are also incredibly limited and not all that useful.
--:-- What we can do, because we work with these IT teams that can connect with service account credentials and admin level permissions, is create a structure where people can access elevated permissions and do all these crazy things in the org, but with very strict deterministic approval cycles, workflows that are deterministic, and an audit trail. Under the hood, we are also a ticketing system. So you say, "Okay, I want my agent to go do X, Y, and Z." Most organizations are going to be very uncomfortable with that at the large enterprise scale. But with Serval, that's going to create access requests that go through an approval process, all of which will be documented and logged. Then you can get just-in-time access to be able to run that deterministic workflow one time, and all of it will be tracked. Yes, it's a little bit more annoying than the magical personal agent experience we love, but it's what has to be true for a large enterprise—especially a regulated enterprise—to ever adopt these kinds of technologies.
--:-- Sarah Guo: This also feeds into the traditional diffusion of this technology, which has been consulting firms—the Accentures and Deloittes of the world—kind of owning a piece of this, doing the implementation for people. You're taking that and putting it in the hands directly of the IT department. Are you also exploring this more sort of forward-deployed motion, or how do you actually get this built out in the enterprise or used effectively?
--:-- Jake Stauch: Absolutely. Diffusion is obviously the gap here. The technology can do so much, and enterprises have barely scratched the surface of what's possible. A lot of that is just understanding the people, the process, what workflows actually exist, and how you want them to exist in the future. We think you still need people for that. So we deploy both a deployment strategist—we call them that—and a deployed software engineer that go on site with customers to figure out how do we implement this, how do we solve all your problems as a business, and then walk you through the implementation.
--:-- What we've found, though, is that a lot of times after we've got this implemented and deployed, customers then have a long tail of more stuff they want. As we've discovered with software engineering, there's actually unlimited demand for work. We've started to build out a team—we call it ProServal, or automation engineering is the field—and we bring in these automation engineers to do ongoing continued support and services. There's so much that they want to automate, and they need the help to be able to figure out what can be automated, what the processes look like, how to navigate all of our systems, and then put it into production with a product like Serval.
--:-- Elad Gil: How old is the company today?
--:-- Jake Stauch: Two and a half years.
--:-- Elad Gil: If you just look at ServiceNow as a business, I feel like Serval is one of the only companies that is like, "We are a direct attack replacing enterprise systems of record, including for large enterprises." At the beginning, did you think you were going to do that? How do you eat the elephant in terms of how big that product suite is?
--:-- Jake Stauch: I like to say everything has gone exactly according to our very delusional plans. We set out from the beginning to be the largest enterprise software company of all time. We actually made a list of all the companies, we circled ServiceNow, and we said we're going to beat them and we're going to build something much bigger. We said that on day one when it was two people in my co-founder Alex's living room. Even though that was crazy, that was always the intention. We didn't grow into that ambition; we set out from the beginning.
--:-- Elad Gil: A lot of people would have started with automation on top of their ticketing system, for example.
--:-- Jake Stauch: Exactly. The way you typically go after these large enterprise startups—at least the common approaches—is they build like the layer, or they go down market. "Let's build a small layer, play nice with the big incumbents, AI agent on top of Salesforce." That has never worked. It really caps out. I guess maybe the only one that worked is Viva or something, because they started off on Salesforce and then migrated to become their own platform. They're the only one I can think of that backwards integrated. It's really tough. It caps you unless you find some niche to expand into later on.
--:-- The other approach is just go down market—service the SMB and mid-market. You see that with a lot of competitors to Workday, competitors to Salesforce. They just say, "We're going to be that but for the lower end." We looked at that and said, "Why not just go after the big one?" It's much harder, but at least the prize is quite big if you win. I think that was so important for us because you end up making so many different decisions if you're building for the large enterprise from day one. All the product decisions, how you build your team, how you think about growth—they're fundamentally different than if you go serve one market and then hope to expand later.
--:-- Sarah Guo: How'd you choose this versus Salesforce? There are a few other sort of canonical giant companies in different categories.
--:-- Jake Stauch: I think we didn't really understand how we'd beat the other companies and why we had an advantage and what AI unlocked that they couldn't catch up to. What we understood about ServiceNow because of our experience with IT is that fundamentally, it is a workflow automation platform. Our belief—me and my co-founder had this idea early on—was that code was the right way to express automation in the enterprise. You should just build all automations in code. These drag-and-drop workflow builders are a painful level of abstraction because it's too complex for non-technical users and annoyingly abstracted for technical users. Code is a better way of building automation. The problem, obviously, is not many people can write code and do it quickly enough for it to make sense for automation. As AI made this possible, that was what clicked for both of us. If code is the right way to build automation, and now anyone can write code, we can build a better automation platform. And what is the best market for an automation platform? It's actually taking on these massive legacy software companies like ServiceNow.
--:-- Elad Gil: Two and a half years ago, when you got started, most of the coding stuff was sort of tab complete. It did not work.
--:-- Jake Stauch: To be very clear, this is a brilliant idea, but in practice, when we started building this...
--:-- Elad Gil: What was the assumption? Was it that the coding would get much better? Was it that you'd build tooling and that you didn't need to because things got so good? I'm curious about that transition.
--:-- Jake Stauch: The bet was that coding agents would get very good. The theory there was it's a verifiable domain, and it should actually hill climb fairly well. We believed fundamentally that coding agents would get good—at least for these deterministic procedural scripts of reset password, Git approval. All of these things are just so defined and verifiable that we felt like it would get good at those things. From a workflow automation perspective, we thought it would get there.
--:-- In the early days, it did not. What we ended up doing in all our early demos was we would prompt the LLM and say, "Here's a suggested way you might want to build a password reset workflow." We'd seed it with a lot of examples, and over time we'd make that library bigger and bigger. But then the models just got so good that eventually it became hard to break. It became hard for you to give it a workflow that it couldn't automate if the APIs existed.
--:-- Sarah Guo: When you think about where coding agents are today and what your customers want, are you and Alex excited about them getting better from here? Are you nervous? What does it look like in terms of impact on your own product?
--:-- Jake Stauch: We're so excited for them to get better because I think it unlocks a lot of the things we'd love to do around better browser use, computer use, and more of an approach to background agents and personal agents that can be really useful in the enterprise. We're excited for the evolution of the category. We've always believed fundamentally that we are not building the LLMs ourselves. We are building the guardrails. We are building the boundaries. The boundaries are the product for us because we are the way that a large enterprise with a lot of rules, approvals, process, and politics can adopt these tools safely and comfortably at scale. Serval is the boundary. Serval is the guardrails.
--:-- The intelligence getting better just makes our product more powerful and also makes our customers want to clamp down even further on those boundaries and guardrails. I actually think access is going to get more and more controlled for individuals in the enterprise. As access becomes stricter and you let people do less and less, your permissions get more and more narrow.
--:-- Elad Gil: Why is it going to happen?
--:-- Jake Stauch: Because if you think about it, I might safely give you more permissions than you need because I don't actually think you're going to do all that much damage. But if you have access to all these agents, then the damage you can do with any kind of slightly elevated beyond normal permissions expands. We're seeing this in cybersecurity risk right now. What I'm actually going to want to do if I'm a large enterprise is shrink everyone's permissions down to truly least privilege. You cannot access any document and you cannot take any action unless you absolutely need to do it for your job. But then you're going to have exceptions all the time. We are the layer that grants those exceptions and makes it possible for you to set up approval procedures around that, have an audit trail and a ticket associated with that. By owning how privileges get escalated in the enterprise, we actually become more powerful and more essential the more you clamp down on end user permissions.
--:-- Sarah Guo: Is there a story or a win you can share as to why people would be taking a big risk going with you to replace a very embedded incumbent? It's an investment of their own time. I assume it's personal risk. Why do that? Or is there an experience where people have already gone through to the other side?
--:-- Jake Stauch: We have to convince ourselves and our customers that Serval is the way you get promoted at work. It can't be that Serval is a risk you take to hopefully get something on the other side. We have to convince customers that you do this and you're going to make your career. You're not just going to not get in trouble, but you're actually going to get a promotion. This is top-down. We sell into the C-level, we sell into IT leadership in many cases, and then distribute into other departments. The message is: people built their career on ServiceNow; you're going to build your career on this next generation.
--:-- We're seeing this play out. One of our largest customers is Fox Corporation. Fox adopted Serval company-wide. The person who led that migration on the ground level actually got nominated to be the MVP of the entire company and got an award for his work there. That is really what we're about. We're convincing our customers that this is the way to not just prove that you're useful to your organization, but to prove that you might be the most essential person in your company because you're leading this change. We're so pivotal to the AI transformation story that I think it's something we can sell really nicely to our customers.
--:-- Elad Gil: When you started Serval, I think you were very adept at thinking ahead. You knew coding agents or coding would get better. You knew there's this way of transformation. You identified the good market segment to go after. What do you predict will happen in the next two or three years, not only for your company but for AI more broadly? Is there something else you're seeing on the horizon that you're very excited about or a big shift that you see coming?
--:-- Jake Stauch: I think consumer often gives us a window into what's going to happen next in enterprise. It doesn't always look the same, but I think it's a nice window. I'm very bullish on what personal agents look like in the enterprise. In the same way that email and Slack and these tools transformed how we communicate at work, I think personal agents will also be a big part of that. I think everyone will have their personal agent that does tasks for them, that delegates to other personal agents. Your agent will talk to other agents at work. I'm really excited about that.
--:-- But then how do we secure that in a world of increasing cybersecurity risk and increasing concerns at large enterprises around what's happening with everyone's agents taking action in all their computers and using their browsers? That tension between what employees and consumers want out of their personal agents at work and what enterprises need from a governance perspective is really interesting. I think there's going to be a lot of opportunity in that tension.
--:-- Sarah Guo: Do you think there will be a separate personal and work agent?
--:-- Jake Stauch: I think so. In the same way that I have a Gmail and I have a corporate email.
--:-- Elad Gil: If there's a premise that exceptions or access control is going to get complex and require a different level of scale, and that you're just going to have more automations overall because if you make it easy, people are going to extend in all different places—are there core design principles beyond "code generation should be the actual form of automation" that you have put into the product today that reflect that?
--:-- Jake Stauch: I think the biggest one is the software development lifecycle—something we take for granted in an engineering context. There's actually a dev environment, then there's staging, then there's prod, and there are gates between them. Maybe you've got branches that people can develop on that won't harm the other systems.
--:-- Elad Gil: Sorry, just to be clear, that's not how IT automation works today.
--:-- Jake Stauch: It does in some cases, but I think most new AI products are not embracing that. They're basically saying, "Let's build these cool agents and ship to production." But this idea that you have to actually follow an SDLC process if you're a large enterprise—a large financial institution is not just going to ship new agent capabilities without a review and user acceptance testing and all of that.
--:-- Jake Stauch: I think that's something we think about a lot: what is going to convince the largest companies in the world to adopt this technology, not just what is powerful and cool and something we're excited to get in the hands of customers.
--:-- Sarah Guo: It feels like it takes a lot of AI companies, particularly if they're vertically centric, a year or two to figure this out, sometimes longer. It's interesting to watch the whole industry now adopting both the forward-deployed motion and starting to think through software development lifecycle compliance and other aspects of the enterprise in a way that was almost missing. Some companies just look at what enterprise software takes and decide they don't really want to build that. You look at the story of a lot of these companies that have gone down market—sometimes they do look at the opportunity to go and build for the enterprise, and it requires a whole different set of building principles and products that you have to build. Some just don't choose to do it, and they cede that ground to other companies. We've chosen to say, "Give me your gnarliest problems and we're going to solve them and make AI transformations possible in this crazy world of risk and governance concerns."
--:-- Elad Gil: When you have your deployment in customer environment engineers and then ongoing automation specialists, and then you have the core platform, how do you decide what is actually brought into the core platform versus what stays in every individual customer's environment?
--:-- Jake Stauch: When we deploy these software engineers, what's really important is we actually never build custom features. We don't build bespoke product capabilities. Luckily for our category, no one asks us to. Everything that people ask for is something that every other customer is going to want at some point. Maybe they impact the prioritization, but it's still something we've always wanted in the platform.
--:-- Now, because it is a code-gen platform under the hood, what you build in Serval is very, very unique. The password reset workflow for one company is going to be very different—shockingly—than the password reset for another. Same with offboarding and all these different automations internally. They look very different. But that's just baked into the core primitives of the platform. Serval almost becomes the IDE for the IT department and other departments to build their automations and manage them at scale.
--:-- What we do is if we look and see the same customer building the same automations over and over again and having to customize them and build them in the product in the same way, we start to consider, "Should that be a first-class feature in the product?" One example: we saw a lot of customers building workflows to detect knowledge-based articles that were conflicting or out of date. We said, "We should just do that automatically." We had customers that would set up automations to reclaim old licenses—licenses for software you're paying for that no one is using, let's automatically cancel those subscriptions. We just decided to build that in the platform. It's a way for us to kind of test the waters of what should be really first-class features versus what should stay as automations in Serval.
--:-- Sarah Guo: Our enterprise customers, as it relates to just general spend on code generation for software teams, have gone through a journey this year. They're like, "We're going to let people try it. Then we're going to token max and we're going to restrict that and figure out who's actually doing this productively." How do you think that reflects for your business, and how do you think about the actual intelligence cost?
--:-- Jake Stauch: I think we're still watching this. I do see our CIO customers trying to pull back from that initial token-maxing era. But I also think we're so early in the adoption curve that it's not going to be a clampdown in aggregate. It's actually going to continue to be a pretty dramatic rise because most companies are still figuring this out for the first time. They are all going to go through this cycle of rapid early adoption and then a pullback of spend and trying to measure and calculate ROI.
--:-- What I hear from our customers all the time is not really a concern about the long-term budget impacts. They realize there's going to be a short-term pullback, but they're still planning on increasing spend. What they need, though, is more proof of ROI. That, I think, is what's going to transform in the industry: people looking for how can I prove that this spend is giving me something. That is hard. I don't know that anyone's really figured out how to measure what you're getting from the tokens that you're buying. That is going to be a huge opportunity for us to show the work that's being done and try to equate it to work that would have been done in other platforms, with other cost models, or by labor. That's what our customers ask us for: "Help me prove that this was worth it." That's the assignment. That's what we have to do.
--:-- We try to price on workflows run because I think it's a better reflection of actual work being done versus just tokens, which is more like verbosity and inefficiency. But I think we're all going to figure out different models of proving that what we're doing makes sense.
--:-- Elad Gil: One of the things people have been talking about is a sort of SaaS-pocalypse where the idea is people will just build their internal SaaS software and all these categories will go away. A lot of people, to your point on calculating the ROI and tokens, are trying to sort out whether it's worth rebuilding these systems and whether you can actually do that effectively. What is your view on the SaaS-pocalypse or this sort of shift in SaaS versus AI?
--:-- Jake Stauch: I think the SaaS-pocalypse is a little overblown. I think what we're going to see is a consolidation of the market. There are a lot of SaaS companies that are essentially just CRUD apps with some workflow logic on top. Those are going to be the ones that are most at risk. But there are also a lot of SaaS companies that have deep network effects, deep data moats, or deep integration moats. Those are going to be harder to displace.
--:-- I think what's more likely is that we see a world where the number of SaaS tools per company decreases, but the ones that remain become much more powerful and much more integrated. AI is going to be the layer that ties all of these together. So instead of having 50 different SaaS tools, you might have 10, but each one is doing a lot more and they're all connected through an AI layer.
--:-- For us specifically, we think about this a lot. We're not trying to replace every SaaS tool. We're trying to replace the workflow automation layer that sits on top of all of them. ServiceNow is the biggest example of that, but there are others. We think that by making it easy to build automations that touch all of these different systems, we become the central nervous system of the enterprise. That's a much bigger opportunity than just replacing one or two SaaS tools.
--:-- Sarah Guo: I think all the companies I've talked to that tried to vibe code these tools internally, especially the main system record tools, not the kind of point solutions, but the main system record tools, have discovered that it's quite expensive to build and maintain enterprise software. Even if you can do it, it doesn't make economic sense in many cases unless your needs are very bespoke and unique. Buying off the shelf for things that are not your core competency has always been the build versus buy argument. I think that is as true now as it has ever been, because it still is hard to build enterprise software. A lot of companies are discovering this or relearning this: enterprise software is complex and you need people to maintain it. Use cases and edge cases, things that break down, people need to update it. Then you have to staff the team and you're paying for that labor. No one really wants to work on the internal tools team in many cases. So you have all these problems with doing that.
--:-- Sarah Guo: Where I do see it is in the point solutions. A lot of companies have hundreds and hundreds of SaaS applications. They probably need a couple dozen total. So I do think you're going to have a shrinking of the sprawl of different point solutions, and they're going to be eaten by the core systems of record and the core platforms. All the middleware, too, I think is just going to go away.
--:-- Sarah Guo: So I think there is going to be consolidation. There will be a SaaSpocalypse, but in that kind of point solution, smaller market cap type companies. I don't think that you're going to truly get away from the core systems of record: the ERP, the HRS, the ITSM, the CRM. I think those are going to stick around. They might transform. And what we're seeing with Serval is we are displacing those tools. So I think there will be a new generation of incumbents. But I don't see those categories just disappearing.
--:-- Elad Gil: I'm curious what mental model or what frame of mind you have for the IT user. They're not software engineers, but they understand their business. They understand what they're trying to automate. When you talk about vibe coding, I think, okay, but do you understand security architecture? Do you understand how this is going to integrate with governance? How do you square code generation and the challenges with vibe coding with building robust IT systems?
--:-- Jake Stauch: Yeah, absolutely. I think IT is really interesting because they're closer to a software engineer than a lot of people give them credit for, especially in large enterprises. The IT department is the engineering team in a lot of cases. They don't have a separate software engineering team. So IT runs all these systems and IT often does write code and manage the systems. We treat them as adults. We give them the tools that we would give software engineers. Now, we abstract away some of that complexity. We make it easier to use in some regards. We don't just hand them cloud code and tell them good luck. We build a platform that's dedicated to their use case. We don't expect them to be able to necessarily write code or understand what's being written, but we still give them access to it under the hood. You can always see the code and you can always treat this through a full software development lifecycle, but you also don't have to.
--:-- Jake Stauch: A lot of the product principles involve thinking with a few different personas in mind: an IT engineer who actually can and does write code, all the way through someone in the marketing department or the people team that doesn't want to see code ever, but still wants to have that flexibility of code gen under the hood because they've got a complex workflow they want to build that they don't want to drag and drop their way to. It comes down to having multiple personas in your head and also treating IT like an adult. I think a lot of the products that have been out there have actually been pedantic to IT, like, "Oh, you don't actually understand this stuff, so you need this very stripped down, more opinionated, simpler version." Actually, IT appreciates the complexity and appreciates the power of these platforms. We want to deliver that to them.
--:-- Sarah Guo: You guys built an agent that helps you build automations. Can you explain how that works?
--:-- Jake Stauch: Yep. One of the things every company discovered as they were building the first generation of these AI tools is that you could build such amazing, complicated systems. But then you had to train people to understand all of your product primitives and how these systems fit together. The first generation of this was we deployed all these FDs to explain to customers how to build on the platform. Actually, let me just build it for you. Then we realized it would be really nice if customers could just talk to an agent that understood their requirements and translated that into the product primitives. So we built this agent called Catalyst that does exactly that. It's an agent for managing the platform. It's an agent for admins to build out all those underlying automations, but also configure the rules around approvals, permission checks, build out different policies, update all these different systems, understand the entire deployment. Because Serval is all code gen under the hood, having the agent on top of that to help you manage the system just made a ton of sense. It's so powerful because now I don't need you to train yourself on how to use Serval. You can just talk to an agent that actually understands how to use Serval and how to build on it. That becomes your primary interface into the product.
--:-- Sarah Guo: I think one thing that is underappreciated and is cool is that the experience for agents in different products, and in code generation agents as well, being able to bring a user along as they are building a product or an implementation is actually raising the technical competence and ability of a huge number of users and turning people into more engineers.
--:-- Jake Stauch: Yes, absolutely. You have this blank slate problem in the old generation of automations: where do I even begin? I know I want to automate all these parts of my job, but where do I begin? Products like ours say, "Where do you want to begin?" and can start making suggestions. You can iterate with them. You can say, "I actually don't understand that. What are you trying to say? Can you explain this to me? Okay, I'm trying to do this. How would I even go about doing this?" They can give you that information. They can fetch the information. They can explain concepts to you. It becomes a build partner and you can achieve so much more than just prompting an agent to go and build something. It can actually walk you through the process and help you explore new opportunities that you wouldn't have even thought of.
--:-- Elad Gil: One thing that seems a little under-discussed right now is background agents. Could you tell us more how you think about that and what you think is coming in terms of this next wave?
--:-- Jake Stauch: Yeah, I always describe background agents to companies as the ultimate moonshot in enterprise software. These agents that go off and do work for you and magically discover all your problems and solve all your problems. It would be amazing if we could get that deployed and it could work and it could follow our permissions. I think no one's really figured out how to do this well. On the one hand, you want these to be very autonomous, to go off and figure stuff out and then solve problems without somebody babysitting them. On the other hand, autonomy is very scary in the enterprise. You want to see exactly what they can read. This is the OpenAI hugging face incidents or other things where agents are given a task and they do something unexpected in order to accomplish that task. What are the guardrails you can put in place? My best employees, what do you do with them? You say, "Here's the general high-level mission. Go off, don't talk to me, and go do it." And you'll come back and they'll have done amazing things. I want to be able to trust agents to do the same thing for me, where I can give them this high-level task and they can come back and solve this problem. I don't think we've figured out how to do that in a way that feels safe in the enterprise.
--:-- Elad Gil: What do you think is a way to make it safe? People have experimented with different things over time, everything from constitutional AI to certain forms of RL. I'm curious if you have a solution in mind.
--:-- Jake Stauch: In the large enterprise, it's not perfect, but it's an incremental escalation of capabilities. The agent goes off and does things and comes back and says, "Hey, I would like to be able to do all this stuff. This would help me do my job." And you approve them. All these things are deterministic workflows. "I want to call this tool. I want to run this script that does X, Y, and Z. Can I do that?" You just approve these with certain permissions and guardrails around them. You say, "Okay, yeah, that's okay. That's okay. That's okay." Over time, the agent becomes pretty competent because it has a big enough toolkit that it can solve most problems. It's not perfect because it's annoying to have to babysit that process. But I think it's what you have to do, because otherwise it could theoretically just start writing code that could take action on all your systems. That is an unbounded security risk that I don't think anyone's going to be able to take anytime soon.
--:-- Elad Gil: Do you think, to your point of being excited about future capability, we're still missing some model capability here? The ability to be proactive and then predict what an individual wants and get it right is still very hard. You need a lot of preference data. I think about an employee working for you: what are Jake's actual priorities when he gives me this high-level mission? What does he mean? Do you think that's a model problem or it's entirely a guardrails and harness problem today?
--:-- Jake Stauch: I think it's both, honestly. Memory obviously has to be better and better because you want to understand the preferences. You want the context to be insanely large because I don't want to have a bunch of different chats with you. I want to have one single chat that goes back all the way to our very first conversation, and you remember the key points of that that are relevant. Then I think the permissions problem will always remain the biggest thing. Even if I feel like you are a very valuable, smart agent and you can do all these things for me, how much do I trust you to just run wild and not have to babysit your actions? You have to do that with people too, right? Anybody who works for you has to come back and check in and say, "This is what you wanted and here's three options." I think it's a very normal process.
--:-- Sarah Guo: The company is about two and a half years old. You have gotten very quickly to at least an enterprise mature product. How big is the team today?
--:-- Jake Stauch: We are 200 people today.
--:-- Sarah Guo: 200 people. You were 20 in January. It's been an incredible ramp year. That's quite a year. Do you expect to continue hiring so fast? And as somebody who clearly believes in AI, why? And then, you've been part of startups before. How are you operating very differently over the last two years versus the previous 10?
--:-- Jake Stauch: The hiring velocity, I think, will increase even from where we're at today. We believe that there is unlimited demand for work and there's unlimited competition and there's unlimited ambition. Even though every single employee does so much more than what they could do a year ago, we have even bigger ambitions for what they could do. Because every employee is so much more productive, we actually want more of them, not less of them.
--:-- Sarah Guo: Given your time at Verkata before, how are you operating differently over the last two years than the previous 10?
--:-- Jake Stauch: I think what's so unique about Serval and the category we're in and the moment we're in is the TAM is unbounded. The addressable market is really more about the size of our ambition than it is about fundamental limitations on how big this business can get. Because we are automating essentially all kinds of work for every single employee in the world, the TAM is as big as we want it to be. It's really about how fast can we grow and what is our ambition, making decisions based on that versus, "Okay, here's what the market looks like. Here's what we think we can actually do." That's a different set of challenges. It's not strictly better because now there's so much open space and so many more decisions we have to make about what we actually want to focus on. But it is really fun to not be constrained and hemmed in by what the market allows for.
--:-- Elad Gil: Given that scope is very broad, how do you organize the cadence or the EPD team?
--:-- Jake Stauch: We are embedded with customers. Even the engineers that are not necessarily deployed are in the customer Slack channels. We are talking about customers nonstop. They're hearing from customers every single day. Most engineers are on at least one customer call a week. I think most of them are actually on at least a customer call a day. It's like immersion in a country to learn a foreign language. You become so immersed in your buyer and your customer's mindset that you understand what they want. You understand how they're thinking about things. You are adapting to their preference changes in real time. As they're moving with the market and starting to talk about different products they're trying out and things they're excited about, our engineers are constantly aware of all this and able to adapt very quickly. We've decentralized product development really to the engineering team. We don't have a product function because they're the ones talking to customers every day and they're able to rapidly adapt to whatever customers are saying today and this week versus what they're saying a month ago.
--:-- Elad Gil: That sounds like a natural way to keep pace with a ripping market and changing priorities for your customers. How do you and Alex and leadership then make sense of it? In a traditional enterprise software company, you'd have architectural review, a quarterly roadmap, you'd prioritize the roadmap, train all your customers on the new launches, and communicate that things aren't going to change too fast because they can only absorb it so fast.
--:-- Jake Stauch: It's a big challenge. Our product ships and changes faster a lot of times than anyone can keep up, including myself. We just try to make sure that the engineers we hire are really product oriented. That's one change we've made. They have to really understand our customers and we want to put them in front of our customers because they're going to have to make so many decisions at the edge. Not every single product decision can be bubbled up to leadership every single time. That is a fundamental change: you have to be willing to push a lot of the product decision making down below. You have to be willing to move fast and also undo and revert if you go in the wrong direction and kill your darlings. You have to be willing to change course on things that you thought were fundamental to the product and switch gears when your engineers bring you something that shows how the customer sees things now and what we should do differently. One of the things I'm proudest of at Serval is how many times we've decided to erase things we've done in the past that we were very committed to and just say, "You know what? The market's changed. The customer's changed. We're going to do things entirely different." I think that is what's enabled us to keep up: that ability and willingness to reinvent ourselves over and over again. It's one of the most durable things in this era.
--:-- Sarah Guo: Thanks, Jake. Thanks for joining us. Find us on Twitter at NoPriorsPod. Subscribe to our YouTube channel if you want to see our faces. Follow the show on Apple Podcasts, Spotify, or wherever you listen. That way you get a new episode every week. And sign up for emails or find transcripts for every episode at no-priors.com.
