Tallinn, 6:12 AM. The notification wasn't a token delisting or a governance proposal. It was a TypeScript commit. An AI agent—trained on months of internal IT tickets—just wrote a production workflow for a payments company, turned it into a pull request, and waited for a human to hit approve. No drag-and-drop canvas. No low-code wizard. No integration specialist scheduling a 40-hour implementation with ServiceNow.
Ramp says workflow construction is 50% faster. The tool spread from IT to about ten teams, from finance to legal. The company behind it, Serval, has raised $127M total, with Sequoia leading a round at a $1B valuation, and just dropped Catalyst on an industry that thought ServiceNow's grip was absolute.
You saw the Moveworks acquisition, right? $2.85B for a chatbot service desk. That was December 2025. Catalyst is different. It doesn't help IT answer tickets. It removes the human from the ticket loop entirely. This isn't a SaaS update. This is a structural shift in who writes enterprise automation. And if you think this doesn't matter in crypto, you're already behind.
Why Care Right Now? Enterprise ITSM has run on the same passive model for two decades: ticket comes in, triage happens, human fixes, ticket closes. ServiceNow built a $200B+ market cap on that loop. But the AI-native playbook flips it. Catalyst's pitch is simple: analyze ticket history, find repeated patterns, draft workflows, generate code, and create background agents that keep watching your systems for the next problem. The human doesn't build anything. The human reviews.
This is a direct strike at the platform layer, and the timing is deliberate. DataM Intelligence puts the enterprise AI agent market at $6.65B in 2025 and projects $142B by 2035, a 35.5% CAGR. That forecast is wide enough to be meaningless, but it tells you where capital is flowing. Meanwhile, ServiceNow's own AI push is real—but slow. Its Level 1 Service Desk AI can still mostly resolve tickets; it rarely writes new automations. Serval is betting that AI-native generation beats AI-assisted configuration. That bet is why the round exists.
The Technical Dissection: What Catalyst Actually Does Catalyst's pipeline is a full agentic loop: parse ticket history, cluster repeated patterns, generate workflow definitions, produce forms and access policies, and spin up background agents that monitor connected IT systems. Then it ships everything as readable TypeScript. Not YAML. Not a visual flowchart. TypeScript.
That language choice says more than the product docs do. TypeScript is strongly typed, git-friendly, and CI/CD-ready. It means the generated workflow can go through code review, version control, rollback, and standard software engineering processes. It also tells you who the target user is: somebody with engineering muscle. The IT admin is no longer a widget-dragger. The IT admin is becoming a code architect who reviews AI output.
The human-in-the-loop part is the part most people will miss. Every generated artifact starts as a draft. Nothing touches production until a human approves. That's a deliberate safety design, and it's the only reason an enterprise legal team would even consider touching this. But here's the uncomfortable question: how long does that human review stay rigorous?
I've spent years watching exactly this pattern play out in crypto. When a tool works consistently, trust builds. Then review becomes rubber-stamping. Then the human is a signature, not a safeguard. The same thing will happen here. The draft-to-review-to-release loop is sound—on day one. Day 400 is the problem.
The Hidden Risks Nobody's Reporting Let me flag the issues that most coverage is skipping.
First, the model layer is a black box. Serval doesn't disclose which foundation models drive the ticket analysis or the code generation. That means core cost, latency, and scalability are external dependencies. In crypto terms, it's like building a DeFi protocol on a chain where the sequencer is a secret. It can work beautifully until it doesn't.
Second, the data flywheel is narrower than it looks. Catalyst learns from ticket history. But a huge chunk of real IT decisions never lands in tickets. They happen in Slack threads, in meetings, in half-written Notion docs. If the training signal only comes from formal tickets, the AI will optimize for the problems that were logged—not the ones that matter.
Third, there's the integration layer. The background agents need to read APIs, databases, logs, and eventually write access policies. That requires deep connectors. ServiceNow has over 1,000 prebuilt integrations accumulated over 20 years. Catalyst is starting from near zero. The counterargument is that AI can generate new integration code quickly if the connector APIs are accessible. Fine. But a generated integration is only as secure as the permission model underneath it. And we haven't seen any security validation.
Which brings me to compliance. This is the structural gap that could cap Serval's upside. No SOC 2 Type II, no ISO 27001, no FedRAMP roadmap in the public materials. For tech-native companies like Ramp and Mercor, that's tolerable. For banks, healthcare systems, and government entities, it's a hard stop. Enterprise procurement won't care if the AI writes beautiful TypeScript if the vendor can't prove the audit trail.
The commercial math is even more aggressive. Serval hasn't disclosed ARR. Based on the customer base—maybe 10 to 20 enterprise accounts with contract values between $100K and $1M—a reasonable estimate is $10M to $30M in annual recurring revenue. At a $1B valuation, that's a 50x multiple. In classic SaaS, that's absurd. In AI-hype cycles, it's normal. But it means Serval needs to hit $50M to $100M in ARR within the next 18 to 24 months just to justify the current price tag. That's an enormous execution load.
The Narrative War Has Already Started Serval claims that customers buying AI products from ServiceNow are actually deploying less than 10% of what they purchase. Shelfware, essentially. ServiceNow denies it. This isn't a technical argument. It's a battle for CTO and CIO mindshare, where the weapon is a single devastating stat.
Here's what that deployment-rate claim actually reveals: the hardest problem in enterprise AI isn't generating output. It's getting organizations to trust that output in production. The same lesson shows up in crypto all the time. A governance proposal can be flawlessly written and instantly dead if the community doesn't trust the multi-sig. The deployment gap is the trust gap.
The Contrarian Angle: Code Was Never Law Here's where my crypto background kicks in. We spent 2021 and 2022 chanting "code is law." Then we watched DAOs discover that the real power sits in the upgrade keys, the multi-sig admins, and the privileged roles. It was never actually code that governed. It was whoever controlled the authority layer beneath the code.
The alpha isn't in the timeline—it's in the review queue that sits between the AI's draft and the production network. In this new enterprise AI stack, the AI generates the workflow, but the human approval gate is the control point. That means the entire security and governance model reduces to a single question: who gets to approve, and under what conditions?
Right now, Serval's answer is "the human in the loop." But in practice, that human is a bottleneck. And in practice, enterprises will start to automate the review step itself. Approval workflows will get delegated to less senior people. Escalation thresholds will get lowered. And eventually, the AI will be reviewing its own drafts with a checkbox in between.
We've seen this movie. It's the same drift that turned careful DAO voting into "snapshot sentiment = executive action." The mechanism feels safe because there's a visible human step. The actual risk is that the human step becomes ceremonial.
The Permissions Problem Is Bigger Than the Code Problem The deeper issue is that Catalyst generates access policies. That means AI is writing identity and permission rules. In crypto, we call this the attack surface. If an AI agent holds the credentials to read your IT systems and write new access policies, then compromising that agent is equivalent to compromising the entire company.
In crypto, speed is a feature. In enterprise AI, speed is a liability. The faster the AI can generate proposals, the faster a compromised AI can generate catastrophic ones. The countermeasure is the same one we preach in DeFi: least privilege, formal audit trails, and explicit kill switches. We just don't know yet whether Serval has built those controls into Catalyst. It's not public. Red team reports aren't released. The security section is a blank page.
What the Big Guys Will Do Everyone is watching ServiceNow's countermove. It has the balance sheet, the compliance maturity, and the installed base. It also has the classic innovator's dilemma: a fully AI-native workflow generator would cannibalize its implementation revenue and its partner ecosystem. The global SIs—Accenture, Deloitte, the consulting giants—make enormous money installing and customizing ServiceNow. If the AI writes the workflows, what do the consultants sell?
That tension is Serval's opening. But it's also the trap. The moment ServiceNow decides to fully embrace AI generation, it can outspend, out-sell, and out-integrate. Microsoft is even scarier. Copilot Studio plus Power Platform plus Microsoft 365 distribution is a one-two punch that could make Catalyst irrelevant in the mid-market before Serval scales.
Nobody in the trade press is talking about this loud enough: the real race isn't Serval versus ServiceNow. It's every AI-native automation layer versus the platform incumbents that control the distribution and the compliance paperwork. The innovator can win with speed for about 12 to 24 months. Then the giants catch up unless the innovator has a defensible data moat.
Serval's moat is real but unproven. The data flywheel—ticket history, generated automations, production outcomes, model refinement—only compounds if the customers keep feeding the loop. Get 100 serious enterprise customers with petabytes of structured IT history, and the model quality becomes genuinely hard to replicate. Get stuck at 15 flashy fintech clients, and the moat is a puddle.
My Takeaway I've been reading whitepapers at speed since the 2017 ICO boom. The alpha was never in the headline. Check the authority model. Check who can upgrade the contracts. Check who holds the kill switch. It's the same exercise for Catalyst, just wearing enterprise clothing.
So here's your watchlist: One, does Serval publish SOC 2 Type II or any independent security audit in the next two quarters? Two, does ServiceNow ship its own native AI workflow generator and price it below consulting fees? Three, and most importantly, which executive takes the fall when an AI-generated fix triggers a production outage?
I learned this in the 2017 ICO trenches: the fastest tool is only as good as the person with the admin key. We demand transparency from DAOs about authority boundaries. We demand audit trails from DeFi protocols. The enterprise is about to learn the same lesson, the hard way.
The narrative is already ahead of the numbers. The numbers are always watching. Watch the testnet, not the press release.
