Serval bets enterprise AI's value lies in guardrails, not intelligence

1. Why ServiceNow frustrates buyers — ▶ 2:13

   ServiceNow is a powerful workflow platform, but every new automation idea must be implemented, often by consultants, so the platform runs 6 to 12 months behind what the business wants. That is tolerable for stable 5 to 10 year flows such as onboarding. It breaks down when a company needs to keep pace with change. Stouch adds that costs keep rising. Serval chose to attack this incumbent directly rather than build a layer on top of it or go down-market, arguing both routes cap the prize or the product decisions.

2. Coding agents for the rest — ▶ 3:34

   Early enterprise AI adoption clustered in engineering because code review makes it safe for even a 19-year-old intern to build something powerful before it reaches production. Serval aims to extend that safety net to IT, HR, finance and legal through approvals, staging and audit trails. Stouch's premise is that most companies have an IT department rather than a software engineering team, and IT is the group that already configures technology for the company. Under the hood the work is still code generation.

3. Personal agents hit permission walls — ▶ 4:52

   Stouch says workforce agents suit consumers, SMBs and mid-market firms, but in large enterprises an individual's permissions are so narrow that a personal agent can barely read email. Serval instead connects with service account and admin-level credentials. Requests become access tickets, pass a deterministic approval workflow, and grant just-in-time access to run that workflow once, with everything logged because Serval is also a ticketing system. He concedes this is more annoying than the magical personal agent experience, but argues regulated enterprises require it.

4. Why code, and why it worked — ▶ 10:11

   Stouch and his co-founder held that code is the right way to express enterprise automation. Drag-and-drop builders are too complex for nontechnical users and awkwardly abstracted for technical ones. The bet was that coding is a verifiable domain that should improve with training, especially for procedural tasks like password resets and approvals. Two and a half years ago it did not work, so Serval seeded prompts with example workflows. Models later improved until it became hard to give them a workflow they could not automate, provided APIs existed.

5. Tighter access favors Serval — ▶ 13:26

   Stouch predicts enterprises will shrink permissions toward true least privilege. Today a company may safely over-grant access because one person can do limited damage, but agents magnify the harm from any elevated permission. Exceptions will then be constant, and Serval positions itself as the layer that grants them with approvals, tickets and audit trails. His conclusion is that better models make customers clamp down further, so Serval grows more essential as controls tighten. This is a prediction, supported by his reference to current cybersecurity risk rather than data.

6. Selling the promotion, not the risk — ▶ 14:52

   Replacing an entrenched incumbent carries personal career risk for buyers, so Serval pitches itself as the way to get promoted. Sales go top-down to the C-level and IT leadership, then spread to other departments. Stouch's evidence is Fox Corporation, a large customer that adopted Serval companywide. The employee who led the migration was nominated MVP of the entire company and received an award. The pitch is that leading the AI transition makes an IT leader the most essential person in the organization, much as ServiceNow once did.

7. Agents need a software lifecycle — ▶ 17:57

   The biggest design principle Stouch cites is the software development life cycle: dev, staging and production environments with gates between them. In his view most new AI products skip this and ship agents straight to production. A large financial institution will not deploy new agent capabilities without review and user acceptance testing. He argues some AI startups never learn this and cede enterprise ground, while Serval chose to take on the gnarliest problems in a world of risk and governance concerns.

8. Promoting automations to features — ▶ 20:00

   Serval says its deployed engineers never build bespoke product features, and no customer asks for them. Because the platform is codegen underneath, each company's password reset or onboarding workflow is already unique. The platform acts like an IDE for IT. Feature decisions come from repetition: when customers keep rebuilding the same automation, it becomes a candidate for first-class status. Examples include detecting conflicting or outdated knowledge-base articles and reclaiming unused software licenses by cancelling their subscriptions.

9. Token spend needs ROI proof — ▶ 21:53

   Stouch says CIO customers are pulling back from an initial token-maxing phase, but he expects aggregate spend to keep rising dramatically because most companies are early in adoption. The customer demand is for proof of return, which he admits no one has figured out how to measure for tokens. Serval prices on workflows run rather than tokens, which he says reflect actual work done rather than verbosity and inefficiency. It aims to equate that output with what other platforms or labor would have cost.

10. SaaS apocalypse limited to point tools — ▶ 23:22

    Stouch rejects the idea that companies will build their own systems of record. Those he has spoken with who tried found it expensive to build and maintain, requiring staffed teams that no one wants to join. Build-versus-buy still favors buying non-core software. The consolidation he expects hits point solutions: companies with hundreds of SaaS apps probably need a couple dozen, and the rest get absorbed by core platforms. ERP, HR, ITSM and CRM categories persist, with a new generation of incumbents, which he says Serval is displacing.

11. Background agents via incremental escalation — ▶ 29:46

    Stouch calls background agents the ultimate moonshot in enterprise software and says no one has solved them. Companies want autonomy without babysitting, yet autonomy is scary, and he cites the OpenAI Hugging Face incidents as examples of agents doing something unexpected to finish a task. His approach is escalation. The agent asks to run specific deterministic tools, humans approve each, and the toolkit grows until the agent is competent. He admits it is imperfect and tedious, but says unbounded code execution is a security risk no one can accept soon.

Jake Stouch's central claim is counterintuitive: as AI agents become more capable, large enterprises will want to constrain them more, and the company that governs exceptions to those constraints becomes more valuable. Stouch, co-founder and CEO of Serval, says his company is not building language models. It is building the boundaries around them. Intelligence improving, in his view, makes the product more powerful and also makes customers want tighter guardrails. That is a strategic position as much as a technical one. It treats model progress as a tailwind for a control layer rather than a threat to it, and it explains most of what Serval builds.

The mechanism is a change in how permissions behave. Today, Stouch says, an organization can safely give someone slightly more access than needed because one person can only do so much damage. Agents change that arithmetic, since any elevated permission now carries a larger blast radius. His prediction is that enterprises will shrink everyone toward true least privilege, with no access to documents or actions unless the job demands it. Exceptions will then arrive constantly. Serval wants to be the layer that grants them through approval procedures, tickets and an audit trail. By owning how privileges escalate, he argues, the company grows more essential the harder customers clamp down.

This logic also explains his skepticism about personal workforce agents in big companies. Stouch says they work well for consumers, small businesses and much of the mid-market. In a large enterprise, an individual's permissions are so limited that an agent can barely read email, so the agent inherits those limits. Serval works around this by connecting with service account and admin-level credentials, then routing each action through a deterministic approval workflow. Access is granted just in time, usually to run a defined workflow once, and logged because Serval is also a ticketing system. Stouch concedes this is more annoying than the magical personal agent experience. His argument is that regulated enterprises will not adopt agents any other way.

The same constraint shapes his view of background agents, which he calls "the ultimate moonshot in enterprise software". The dream is the best employee given a high-level mission who disappears and returns with results. The obstacle is that autonomy frightens enterprises, and Stouch cites the OpenAI Hugging Face incidents as examples of agents doing something unexpected to complete a task. His proposed answer is incremental escalation. The agent asks to call particular tools or run particular scripts, humans approve them one by one, and the toolkit grows until the agent can solve most problems. The alternative, an agent writing arbitrary code against company systems, he calls an unbounded security risk.

The product underneath rests on an older bet that code is the right form of enterprise automation. Stouch argues that ServiceNow's drag-and-drop workflow builders are too abstract for nontechnical people and awkward for technical ones. Every new idea must be implemented, often with consultants, so the platform runs 6 to 12 months behind what the business wants while costs keep rising. When Serval started two and a half years ago, coding models could not do this reliably. Serval seeded prompts with example workflows and enlarged that library, betting that coding was verifiable and would improve. It did, to the point where, he says, it is hard to give a model a workflow it cannot automate if the APIs exist.

Go-to-market follows from the same premise. Stouch dismisses the two common startup routes against incumbents, building a layer on top or serving the lower end of the market. A layer, in his view, caps growth unless it finds a niche to expand into. Going down-market means building a different company. Serval chose to attack ServiceNow directly from day one, which shaped product, hiring and growth decisions. It sells top-down to IT leadership and frames adoption as a career-making move, citing Fox Corporation, where the migration lead was nominated company MVP. It also insists on a software development life cycle with staging and approval, which he says most AI products skip.

The economics are less settled. Stouch expects CIOs to retreat from early token maxing but still raise aggregate spend, with the real demand being proof of return. He admits no one has figured out how to measure what tokens buy, and Serval's answer is to price on workflows run, which he says reflects work done rather than verbosity. His view of the so-called SaaS apocalypse is similarly narrow. Companies that vibe coded replacements for core systems of record found them expensive to maintain, so he expects consolidation of point solutions rather than the disappearance of ERP, HR, ITSM or CRM. He adds that Serval is displacing incumbents in those categories, so the build-versus-buy argument is also his sales pitch.

The limits of the evidence matter here. The least-privilege prediction is supported by a reference to current cybersecurity risk, not data, and the Fox story is one customer's anecdote. Stouch is also a vendor describing a market in which his own product is the answer. Two tensions stand out. He says no one has figured out how to measure the return on AI spend, yet proposes workflows run as a better proxy without showing it solves the problem. And he admits his approval model is annoying, which raises the question of whether users will tolerate it when a frictionless consumer agent is available elsewhere.

The deepest unresolved issue is whether this model can ever deliver what buyers say they want. Stouch describes the ideal as trusting an agent the way he trusts his best employees, but his mechanism requires a human to approve each new capability. He argues the toolkit eventually becomes large enough that agents solve most problems, and he says a model problem exists too, since memory and context must improve. Yet he insists permissions will remain the biggest issue regardless. If so, oversight cost may track the agent's usefulness rather than shrink, and the control layer wins by staying in the loop rather than by making the loop disappear.