Serval CEO Jake Stauch on the Promise and Peril of Autonomous Enterprise AI Agents

Jake Stauch, co-founder and CEO of Serval, describes autonomous enterprise agents as the "ultimate moonshot in enterprise software." These agents are envisioned to magically discover and solve problems, but Stauch notes that no one has yet figured out how to implement them effectively.

The primary challenge with unbounded autonomous agents is the significant security risk they pose. If an agent becomes competent enough to write code and take action on systems, it presents an unpredictable security threat that enterprises are currently unwilling to accept. This concern is amplified by incidents where agents, given a task, perform unexpected actions.

Large enterprises require rigid rules, approvals, and process-heavy guardrails to adopt AI safely. Serval aims to provide these necessary boundaries and infrastructure, enabling organizations with complex political and procedural landscapes to implement AI tools securely and at scale.

The company is named Serval, after the African wildcat, reflecting its ambition to rapidly tackle ServiceNow and extend code generation capabilities to IT and broader enterprise teams.

Serval offers enterprise automation and service management through codegen agents, aiming to enable users to ask questions, get help, and automate undesirable job aspects. The company directly competes with ServiceNow, which Stauch identifies as the historically large legacy provider in this space.

Stauch identified this market opportunity while working at Vicata, where he sold enterprise physical security products to IT teams. Through numerous conversations with IT professionals, he consistently found ServiceNow topping lists of high-cost, disliked systems.

IT teams frequently view ServiceNow as an expensive, legacy platform that is difficult to customize and keep up with their needs. Serval seeks to provide a faster, more adaptable alternative to address these long-standing frustrations.

Legacy platforms like ServiceNow are acknowledged as robust, customizable, and powerful for mapping enterprise workflows. Fundamentally, they are workflow automation platforms that can achieve a great deal. However, implementing new ideas on these platforms is challenging.

Customizations often require bringing in consultants, resulting in a significant delay of six to twelve months between identifying a business need and deploying a solution. This lag is tolerable only for long-term, static workflows that are not expected to change frequently, such as a major onboarding process.

The core frustration stems from these platforms' inability to keep pace with the rapid demand for automation and their increasing cost over time. Businesses find that the platforms cannot evolve quickly enough to meet their changing needs, leading to escalating expenses and dissatisfaction.

Personal virtual assistants, while effective for consumers and small to mid-sized businesses, are largely ineffective in large enterprises due to limited individual permission sets. Employees' personal permissions are often so restricted that even if an agent were allowed (which is often not the case), it could barely perform meaningful actions, such as reading emails.

The extensive processes and approval requirements within large organizations severely limit what a personal agent can do. Serval addresses this by working with IT teams that can leverage service account credentials and admin-level permissions. This allows Serval to create a structure where elevated permissions can be accessed.

To ensure secure execution, Serval employs strict deterministic approval cycles and workflows. As an underlying ticketing system, it provides an audit trail for all actions. When an agent requests to perform a task, it generates access requests that go through an approval process, ensuring all actions are documented and logged, thereby providing just-in-time access for deterministic workflows.

Enterprises often struggle with adopting new AI technologies due to existing people, processes, and workflows. Stauch notes that companies have barely scratched the surface of AI's potential, largely because they lack a deep understanding of how to integrate it into their specific operational contexts.

Serval addresses this by deploying specialized personnel on-site: a 'deployment strategist' and a 'deployed software engineer'. These individuals work directly with customers to understand their business problems, implement Serval's solutions, and guide them through the integration process.

Recognizing that customer demand for automation is virtually unlimited, Serval established a dedicated team called 'Pro-Serval,' or 'automation engineering.' This team provides ongoing support and services, helping enterprises identify what can be automated, streamline processes, navigate their systems, and put solutions into production.

Stauch mentions that Serval, founded two and a half years ago, is one of the few companies directly attacking and replacing enterprise systems of record for large organizations, including those typically handled by ServiceNow.

Jake Stauch views ServiceNow fundamentally as a workflow automation platform. He criticizes traditional drag-and-drop workflow builders as a "painful level of abstraction." He argues they are simultaneously too complex for non-technical users and annoyingly abstracted for technical experts, making them inefficient for enterprise automation.

Stauch and his co-founder believe that code is the superior method for expressing automation in an enterprise setting, advocating for all automations to be built in code. The historical barrier to this approach was that few people could write code quickly enough for effective automation.

The advent of AI has changed this dynamic, making it possible for virtually anyone to write code. This realization became the core insight for Serval: if code is the correct way to build automation and AI enables anyone to write it, then a superior automation platform can be created. This strategy positions Serval to challenge massive legacy software companies like ServiceNow.

When Serval started two and a half years ago, early coding models were largely ineffective, often requiring manual prompting and seeded examples for basic functionality. The founders, however, made a strategic bet on the rapid improvement of coding agents.

Their theory was that coding agents would excel in deterministic, procedural, and verifiable domains. Tasks such as resetting passwords or obtaining approvals are well-defined and verifiable, making them suitable for AI-driven automation where performance could "hill climb" effectively.

As large language models (LLMs) improved, the difficulty of automating defined workflows decreased significantly. Stauch states that it has become challenging to find a workflow that cannot be automated today, provided the necessary APIs exist.

Enterprises are increasingly moving towards 'least privilege' access models to mitigate cybersecurity risks. This trend is driven by the understanding that giving individuals more permissions than strictly necessary can expand the potential damage in case of a breach.

The rise of AI agents exacerbates this concern. If employees have access to powerful agents, any slightly elevated permissions can lead to significantly greater potential damage, a risk that is currently being observed in cybersecurity.

Stauch anticipates that large enterprises will continue to tighten individual permissions, narrowing what employees can do to a bare minimum necessary for their job functions. Serval becomes essential in this environment by managing the exceptions to these strict policies. It provides a layer for granting necessary elevated access, setting up approval procedures, and maintaining a comprehensive audit trail with associated tickets for every action.

Stauch believes that consumer trends often provide a preview of what will happen in the enterprise. He is bullish on the future of personal agents in the workplace, envisioning them transforming work communication in the same way email and Slack have.

He anticipates that everyone will eventually have a personal agent capable of performing tasks and delegating to other agents, leading to agents communicating with each other within the workplace. However, this raises concerns about securing these agents amidst increasing cybersecurity risks and enterprise concerns over agents taking actions on computers and browsers.

Stauch expects a clear division between personal and work agents, similar to having separate Gmail and corporate email accounts. This reflects an interesting tension between employees' desire for agent-driven efficiency and enterprises' need for robust governance and security.

A core design principle for Serval is to apply a rigorous Software Development Life Cycle (SDLC) to AI automation. Stauch notes that while engineers take concepts like dev, staging, and production environments with gates between them for granted, many new AI products currently bypass these crucial steps.

For large enterprises and financial institutions, simply shipping new agent capabilities without proper review and user acceptance testing (UAT) is unacceptable. Serval focuses on building technology that convinces the world's largest companies to adopt AI, prioritizing robust processes over merely powerful or cool features.

Stauch observes that the broader industry is starting to adopt forward-deployed models and formal SDLCs, along with compliance and other enterprise-grade aspects, which were notably missing in earlier AI development approaches.

Stauch notes that many companies avoid building enterprise-grade software because they perceive it as too complex and requiring a fundamentally different set of building principles and products. These companies often cede the enterprise market to others.

Serval, however, has specifically chosen to tackle the "gnarliest" problems within enterprises. This includes addressing complex issues involving risk, governance, and the stringent requirements for AI transformations in highly regulated environments.

Stauch argues that companies must decide whether to commit to building for the enterprise market or concede that ground. Serval's approach involves a core platform with deployment engineers and automation specialists, ensuring that customer-specific needs inform product development without creating bespoke features for individual clients. Instead, all requested capabilities are designed to benefit the broader customer base, impacting prioritization rather than introducing unique, one-off features.

Stauch observes that CIOs are starting to pull back from the initial 'token maxing' era of AI adoption, where there was less scrutiny on consumption. However, this is not an aggregate clampdown but rather a move towards more deliberate, ROI-driven spending.

Despite short-term pullbacks to measure and calculate ROI, enterprises are still planning to increase their AI spending in the long term. The primary demand from customers now is for concrete proof of return on investment, which Stauch believes will transform the industry.

Serval addresses this by pricing based on "workflows run" rather than raw tokens. Stauch argues that this metric better reflects actual work being done and efficiency, rather than verbosity or inefficiency. A key challenge for AI vendors is demonstrating the tangible value of their solutions compared to traditional labor costs or other platforms.

Stauch addresses the concept of a 'SaaS apocalypse,' where internal AI tools might replace external SaaS solutions. He notes that companies attempting to build and maintain internal enterprise software, particularly core systems of record, have found it to be prohibitively expensive and often lacking economic sense unless their needs are highly bespoke.

The 'build versus buy' argument remains highly relevant today. Stauch emphasizes that enterprise software is inherently complex, requiring dedicated teams for maintenance, handling numerous use cases and edge cases, and continuous updates. He notes that many companies are rediscovering the challenges and costs associated with maintaining internal tools.

Stauch predicts a shrinking of the sprawl of different point solutions. He believes that many of the hundreds of SaaS applications used by companies will be absorbed and consolidated by core systems of record and platforms. While ERP, HR, ITSM, and CRM systems will persist, they are likely to undergo significant transformation, becoming more centralized.

Serval has developed an agent called Catalyst to manage its platform, addressing the challenge of training users on product primitives and system integration. This agent eliminates the need for extensive user training by acting as an intelligent intermediary.

Catalyst is designed to translate user requirements into platform configurations. It can build underlying automations, configure rules for approvals, perform permission checks, develop different policies, and update various systems, all while understanding the entire deployment.

Because Serval's platform is entirely codegen under the hood, having an agent like Catalyst on top makes managing the system highly efficient and powerful. It serves as the primary interface for administrators, allowing them to interact with the platform naturally without needing to master its technical intricacies.

Stauch highlights that agents solve the 'blank slate problem' in automation, where users often don't know where to begin. Instead of facing a daunting empty canvas, users can now state their desired automation, and the agent can immediately start making suggestions.

This interaction fosters an iterative collaboration, allowing users to ask for explanations, seek clarity on concepts, and explore new possibilities they might not have considered. The agent effectively acts as a "build partner," raising the technical competence and ability of users.

By providing information, fetching data, and explaining complex ideas, AI agents help users achieve significantly more than simply prompting an agent to build something. They guide the user through the process, essentially turning more people into engineers.

Stauch describes background agents as the "ultimate moonshot in enterprise software": agents that autonomously discover and solve problems. However, he acknowledges that no one has yet figured out how to implement this well due to the inherent tension between desired autonomy and the fear of unpredictable actions within an enterprise.

The concern stems from incidents where agents, given a task, perform unexpected actions to accomplish it, akin to 'OpenAI Hugging Face incidents.' While employees desire agents that can operate with a high-level mission, similar to a trusted human employee, the enterprise demands strict guardrails and visibility into what agents are doing.

Enterprises prioritize a safe, incremental escalation of agent capabilities. This means carefully controlling what agents can read and do, rather than granting them broad, unmonitored access. The current approach involves agents requesting specific permissions for deterministic workflows, which then require human approval.

This system ensures that all actions are reviewed and documented, providing the necessary safety and auditability for enterprises to gradually increase their trust in AI agents without incurring undue risk.

For advanced agent performance, Stauch emphasizes the need for significant improvements in memory to understand user preferences over time. He also calls for "insanely large" context windows, enabling a single, continuous conversation that remembers all relevant key points from the very first interaction.

Despite advancements in AI models, Stauch believes that the problem of permissions and trust will remain the biggest hurdle. Even if an agent is highly capable and valuable, the fundamental question persists: how much can an organization trust it to operate autonomously without constant human oversight?

He compares this to managing human employees, who also require check-ins and present options. This highlights that the need for oversight and approval is a normal organizational process, which must be adapted for AI agents to ensure safety and alignment with enterprise objectives.

Serval has decentralized its product development, empowering its engineering team to act as the primary product function. This is achieved by embedding engineers directly with customers; they are active in customer Slack channels and participate in customer calls daily.

This immersive approach allows engineers to understand customer mindsets, preferences, and evolving market needs in real-time. This direct engagement enables rapid adaptation to customer feedback, often within days or weeks, rather than months.

By eliminating a traditional separate product function, Serval ensures that engineers are constantly aware of customer requirements, allowing them to quickly pivot and adapt their products. This organizational structure is designed to keep pace with a rapidly changing market and shifting customer priorities.

Leadership encourages teams to "kill their darlings" and pivot based on customer feedback, fostering an environment of continuous learning and responsiveness. This contrasts with traditional enterprise software companies that often rely on quarterly roadmaps and slower change absorption by customers.