Some startups pride themselves on having prestigious financial backers, but having prestigious customers is just as important. This is a main point of pride for Serval, an enterprise AI company that announced a forty-seven million dollar Series A funding round. The round was led by Redpoint Ventures, with participation from other venture firms like First Round, General Catalyst, and Box Group. Even more impressive than the investors is the company’s list of clients, which includes major AI players like Perplexity, Mercor, and Together AI.
Serval uses agentic AI models to automate IT service management. The company has a unique approach that uses the power of agentic AI while avoiding many of its common problems. One agent is used to code internal automations for everyday tasks, like authorizing software or provisioning a device. The founders see it as a kind of vibe-coding tool, overseen by an IT manager, that does most of the work independently. A separate help desk agent responds to user requests by calling those tools on command, following the rules established for each tool.
Serval’s CEO, Jake Stauch, says the key was to make the process of building a tool as simple as possible. The goal is for clients not to feel the marginal cost of building these automations. The company wants to make it easier to automate something forever than to do it manually just once.
Splitting the task into two agents, one to build tools and one to use them, also gives managers a way to monitor permissions. When an automation is created, the manager sets rules for when it can be used. This provides an extra line of defense against overeager help desk agents.
Enterprise clients are very aware of the risks of a rogue AI system. This is part of why Serval decided against a single all-purpose Help Desk Agent. The company wanted to avoid a scenario where someone could ask to delete all company data and an AI agent would comply. Instead, the agent will state that it does not have a tool for that request but can offer other tasks like resetting a password.
Because the tools themselves are deterministic, they can include extremely complex permissions. These can require certain multi-factor authentication processes or only allow actions within a specific time frame. Any time those rules need to be changed, an AI agent is ready to dive into the codebase and make the alteration.
This is a new approach to the common problem of how to oversee agentic AI systems. The goal is to have full visibility and control over what an AI agent is doing. This is achieved by using Serval to build those tools and customize the permissions and approvals behind them.