Serval's Bet: Keep ServiceNow's Primitives, Take Humans Out of the Build

Jake Stauch is founder and CEO of Serval, a two-year-old company he positions as an AI-native ServiceNow. Host Pat Grady of Sequoia's *Training Data* asks the obvious question: why does the world need another one? Employees need help at work, Stauch says. The category name is enterprise service management; the job is getting someone a permission, a password reset, a laptop without waiting for a ticket to wander the building. The ideal is ask and receive, instantly. That only happens with automation, and AI is the tool that can finally build it.

Grady remembers meeting ServiceNow founder Fred Luddy around 2007 or 2008, when the product looked like a step-function over Peregrine and Remedy. Luddy's insight was that enterprise software is workflows on a database, with a flexible configuration engine on top. IT people loved it. Stauch does not argue with the history. "They got it right." Serval is also workflows on databases. Those are still the right primitives. The failure mode is maintenance. Building a workflow often needs dedicated developers and takes weeks to months. By the time it ships, the business process has already moved. Automation runs behind reality, more now than ever. The same tax hits the asset database: consultants or internal engineers manually keeping IT records current.

Serval's move is to keep the primitives and take the human out of the assembly. A codegen engine takes a natural-language description — steps, permissions, approvals, logic — and turns it into code. The workflow appears almost immediately. Describe which data to pull from which systems and the same engine writes the fetch-and-refresh code. Grady quotes a line from their first meetings that still interests him: if you want enterprise automation to actually happen, building the automation has to be as simple as, or simpler than, the workflow being automated. Stauch still believes it. Put yourself in IT. Someone wants a password reset. Path one: open Google Workspace, find the user, click reset. Path two: open a builder, drag a trigger, drag a response, wire logic. In that moment everyone clicks reset. Make the automation easier and there is no reason not to build it.

Too easy creates its own mess. Grady asks about "slop automation," the cousin of vibe-coded junk. It is real. Someone ships the twentieth password-reset workflow this week, a near-copy of the nineteen before it, and the model no longer knows which one to run. Serval's answer is another agent on top of the product, with full memory of every workflow ever built. Ask it for something that does X, Y, and Z and it can say you already have nineteen of those, delete ten, bucket the rest, add these approvals. It is meant to be an expert in Serval itself, translating a business request into an implementation instead of another duplicate.

Grady has heard from every direction that Stauch is unusually customer-obsessed. Is there a North Star metric, or is it anecdotes synthesized into a gut feel? Stauch says the latter, taken to an extreme. He is in every customer Slack channel. Most of those customers, he thinks, get a message from him every day. Grady flags the scale for listeners who might assume that means four logos. Serval has 100 customers, many of them large enterprises. Living in all of those threads is overwhelming, and he sometimes wonders if it is a waste of time. He still thinks there is no substitute, especially as implementation has gotten fast. More of a startup's moat, he argues, is customer insight — actually understanding what people want. A differentiated read on the customer is worth more than a product advantage, "which is copyable overnight."

That answer sits inside the wrapper panic that started when ChatGPT shipped in November 2022. Grady puts the two schools of thought on the table: foundation models will do everything, versus application companies exist to close the gap between raw capability and a customer's actual problem. Stauch's guiding principle is simple enough to print on a wall. You have to be happy when the new models come out. Whatever you are building should get better when the labs and hyperscalers ship, not become obsolete. For Serval, that means the product *is* the boundaries. The interesting question is no longer whether Opus or GPT-5.5 can do amazing things. It is whether they can do the specific thing you want, inside an enterprise that cares about security, without raising risk.

So the team spends its energy on boring, old-school enterprise software: permissions, approvals, limited API scope, visibility, audits, reporting, logs, alerts — everything that lets a company get comfortable letting models run. Capabilities are practically unlimited. Comfort is the scarce resource. That is the application-layer job as he sees it: not competing with the model's IQ, but making the IQ deployable company-wide.

The architecture splits the work in two. An admin agent helps IT build tools and skills and decides what the help desk may do and know. A help desk agent talks to employees and resolves problems, but it can only call tools that admins have explicitly built, published, and wrapped in approvals and permissions. End users can ask anything. The help desk gets to use full reasoning. Every tool it touches still sits inside a fence the admin drew. That two-pronged split, Stauch says, is what lets you "let the help desk agent run wild" without handing it the keys to the company.

Under the hood they use OpenAI and Anthropic, with eval suites on everything and constant experiments on new releases. The split has been stable for a long time: GPT models are still best at talking to the end user, calling the right tools, and answering in the right tone. Codegen automation prefers Anthropic — Sonnet and Opus. New releases are not plug-and-play. Prompt tuning and infrastructure built around an old model's quirks stop making sense when you swap. Evals are automated; the rest is not as automated as he would like. They slow-roll across customers. Sometimes they upgrade and then downgrade. The new model is smarter and misbehaves in less predictable ways; the old one is a bit slower or dumber, and they know the guardrails will hold.

Cost is not the current obsession, for a unit-economics reason. Serval is not in the business of reselling tokens. Automations compile to TypeScript. Once a password-reset script exists, the next request runs the code; it does not regenerate it. Over time users generate less code because a library of automations covers more of the long tail. Codegen is the expensive step, and they do less of it than outsiders assume. He tells the team to spend more money and use the best models; optimizations can wait. The place cost will start to matter is background agents — long-running jobs that mine historical tickets and device logs and propose problems nobody filed. Those can run away if nobody is watching.

Grady asks the founder nightmare: OpenAI or Anthropic notices ServiceNow is a system of record and ships a first-party version. Stauch admits any answer sounds naive against infinite money and AGI. History, he says, still favors the smaller company; OpenAI and Anthropic themselves are the proof that incumbents with infinite resources lose to focus. Every VC asked the same "won't Google just do this?" question at his first company. Google does a lot of things, and doing all of them well is hard. ITSM is a bad use of a lab's best people. In the past couple of months, he says, Anthropic has added more ARR than ServiceNow has in twenty years. Even a successful ITSM push would take years to match what the rest of the portfolio prints in months. He would not be surprised by a simple mid-market or SMB sketch. Mastering enterprise service management would be a prioritization mistake he does not expect them to make.

Serval now has AI-native logos and large enterprises. Stauch expected their needs to diverge. The surprise is how alike the pain looks, from a few hundred employees to a few hundred thousand. The real difference is how many people it takes to decide. A few-hundred-person company has an IT leader who can say how onboarding and password resets work. At a few hundred thousand, nobody even knows if that person exists, so you live in committees. That, more than product fit, is the implementation bottleneck — and it is why the labs are building consulting and deployment teams. Coordinating humans is the rate limiter on adoption.

Ask the AI-native crowd for a compliment and Stauch reaches for a sad picture. A customer spent much of the day in ServiceNow provisioning someone else's access to Cursor: an ancient ticketing system used to hand out the cool AI tool, while the IT person's own stack stayed stuck in the past. Serval, he says, lets that person spend time building again. They got into IT because they loved technology. Large enterprises talk at a higher altitude, about employee experience. In a small company, even messy IT still answers. In a big one you can send a ticket into the abyss, hear nothing, send another, confuse the queue with two tickets from the same human, and block someone for weeks. There the product changes what it feels like to be an employee, because the problems run deeper.

Inside Serval he forces the team to use the product for everything, including snacks and office supplies. The example he likes is a Slack channel called Dream Team Draft. Recruiting is person-first: employees post the best people they have ever worked with, plus a LinkedIn URL. Serval runs automations that drop the profile into outbound and nurture, and does other things his marketing team will not let him describe, all aimed at making Serval hard to miss. It is a long game. These people are not quitting tomorrow. The employee's job is to say this person is great and stop. Culture matches the machine. Candidates tell him the office is unusually kind and high-energy, which he claims he was not even selecting for. It is a poor place for training, mentorship, onboarding programs, or a clean org chart. He does not actually know who reports to him. They hire people who can produce on day one and who like that ambiguity. There is no promised climb up a ladder that does not exist.

Compared with Verkada, his prior company, the operating system is different. Every department and every role starts under suspicion: maybe it does not need a person. AI gets the right of first refusal. Sometimes they rediscover that a function still needs humans; they start by assuming it might not. Solutions engineering went all the way to zero. So did SDRs, which he calls a less controversial cut. Account executives are not necessarily more technical than the last generation, but each of them has Serval. Product questions get instant answers. Mid-call, the system can generate decks, one-pagers, battle cards, comparison sheets. They expect more from AEs and do not staff an SE bench. Four deployed engineers help with pilot implementation. Enablement got surprisingly far on one product-marketing person and automated resources. RevOps lasted a long time with nobody, then they learned they did need someone — a smaller someone than the original plan.

The company is two years old. Early on Stauch felt useless. The CTO was building an obvious product; he was hiring, booking customer calls, and failing to sell. Then the boulder flipped: the business started dragging him. He is still in customers and recruiting every day. He is less in product than he wants to be. Direction now emerges from customers through forward-deployed engineers and into the codebase. He calls it gradient descent for product: a swarm of feedback, a week later a visibly better system. The cost is less time thinking about the future, which he says is exactly what an AI-company CEO should be doing. The biggest organizational difference is how often you have to uproot something you were sure of a few months ago — rename parts of the product, change how whole flows work. Engineers used to ship software meant to last twenty years. Now they ship software they hope lasts six months, then rebuild when the paradigm moves.

Grady's image is a sandwich: foundation models changing at high speed, customers changing slowly, Serval as the buffer so the downpour of capability does not drown the buyer. Stauch takes the job. Meet customers at their business problems, use deployed engineers to discover how those problems map onto Serval, and on the other side absorb the latest model advances. Educate them that the way you would have done it three months ago is already wrong. They are also building an agent that asks the end user what the business problem is and then just does the work, so the customer never has to track the implementation path as it mutates.

His most contrarian AI take is a crack opening inside companies. Individuals want autonomy: a Claude agent that can do everything and touch everything. The organization wants control, and security and IT are right to be nervous. Consumer products are built for autonomy; enterprise products are built for control; both instincts now live in the same building. Grady reaches for shadow IT. Employees brought iPhones while companies issued BlackBerrys they could lock down. CISOs' first instinct was protect, protect, protect. Eventually they noticed the unapproved devices were a leading indicator of what people needed to be productive, and they embraced them. Stauch thinks the movie is running again. Companies that say yes as the default will run well ahead of companies that say no. The yes-camp will eat security incidents and learn from them. They will still get there first. The risks, Grady notes, are real, and some of those leaders will also be the ones who get burned.

The biggest issue at Serval, still, is hiring. Every AI company he talks to is in the same bind: the pitch is that AI automates the work, and they are all hiring more than ever. People are the last moat. Better people in the room is the only thing that keeps you ahead, and the bar is harder to hold as he meets some hires only at the interview and the onboarding. Grady's frame is accelerating change: cost down, capability up, so optimize for agility, which means the smallest possible number of the best possible people. Stauch's mantra is two words, repeated: fewer, better. How do we make this department fewer and better? A future Serval might be unrecognizable. That rewrite is plausible overnight with a small team. It is not plausible if you have to convince thousands of people that go-to-market and the product just changed.

If Serval ever becomes a monster company, he does not want the story to be revenue or headcount. He wants the impact to be obvious: meaningful work unlocked. Every profession, he says, has a gap between the job you imagined and the job you got. Kids want to be firefighters and astronauts, then discover how much of the day they dislike. Serval's ambition is to be the tool that closes that gap by eating the repetitive, menial work that was never why you signed up. They have done versions of that for IT. If automation spreads across the company, people get back the work they actually enjoy. He does not want the epitaph to be that Serval automated away a bunch of IT jobs. He wants people to say it made work better.