The AI-Native Workflow Market Has a Verification Gap. Serval Is Betting $1B on Crossing It.

Catalyst is a workflow engine that writes TypeScript. It reads your ticket history. It generates automations. Then it asks you to review the code before it is allowed to run. That is not a description of an enterprise IT product. That is a description of a smart-contract audit. The parallel is deliberate. In both cases, the asset is code. In both cases, the risk is human error. And in both cases, the only way to know whether the system is safe is to verify the output. Serval, the company behind Catalyst, has raised $127 million and carries a $1 billion valuation. It claims customers who bought ServiceNow's AI actually deploy less than 10% of it. ServiceNow denies the claim publicly. Neither side has opened its internal usage telemetry to outside review. The ledger never lies, only the interpreter does. But in this case, there is no shared ledger. The product itself is deceptively simple. Catalyst ingests a company's support ticket history, identifies repetitive patterns, and drafts workflow definitions as TypeScript code. It also generates form schemas, access policies, and dashboards. After a human approves the drafts, Catalyst spins up background agents that monitor live systems and propose additional fixes. The company's marquee customer, Ramp, says workflow construction is now 50% faster and the tool has expanded from IT to roughly ten internal teams. Mercor, an AI recruiting platform, is the second public reference. That is the entirety of the public commercial evidence. I have spent two decades building models that verify claims against actual transaction flows. In 2017, I audited a multisig wallet contract and found an access-control flaw in the initWallet function that could have exposed $31 million to hijacking. That experience taught me a simple rule: every line of code is a promise, and promises must be audited. Catalyst is a machine that writes promises. It generates a hundred small promises for every workflow it creates. The question is not whether the code compiles. The question is whether the logic matches the intent of a business process that is rarely documented anywhere outside a few stale wiki pages. Serval's choice of TypeScript is more important than the marketing materials suggest. Strong typing and git-friendly versioning mean the generated workflows can be treated as engineering artifacts, not black-box macros. That is a deliberate wedge into the developer-conscious buyer segment. Low-code tools have always been sold as empowering to business users, but they create a maintenance cliff when the drag-and-drop logic breaks. TypeScript pulls the work back into the engineering lane where it can be reviewed, tested, and rolled back in a CI/CD pipeline. That is significant. It is also the source of the greatest unspoken risk: Catylist is effectively re-creating the traditional implementation consultant as a language model. In the absence of noise, the signal screams. The signal here is the gap between what enterprises say about AI workflow automation and what they actually run in production. ServiceNow is not a trivial opponent. It has two decades of prebuilt integrations, a CMDB that maps every infrastructure relationship, and a partner network that installs its platform across the Fortune 1000. Its AI features may be underused, but the platform itself is embedded in processes that cannot be forklifted without inviting regulatory and operational exposure. Serval is attacking the outer rim of that market. The customers it names are technology-native companies, the kind that will try any new tool once. That is a legitimate wedge. It is not proof of a paradigm shift. The financial structure of the bet is straightforward. A billion-dollar valuation with an estimated ARR between ten and thirty million dollars implies a forward multiple of 33 to 100 times revenue. Hot AI SaaS deals trade at those multiples in bull markets. They compress quickly when collections slow. The real backstop may be the exit route. ServiceNow paid $2.85 billion for Moveworks, an AI service-desk company, in late 2025. That acquisition set a price floor for an AI-native ITSM startup. If Serval continues to add logos and demonstrate expansion within those accounts, a second acquisition is plausible. But the technical moat is not as deep as the narrative suggests. The underlying large language model is not disclosed. The inference infrastructure is apparently outsourced. The real proprietary asset is the private ticket history that Catalyst processes and the behavioral feedback loop that results from each approved workflow. That is a data flywheel, not a foundational technology breakthrough. In a world where model providers are increasingly commoditized, data aggregation is the only durable piece. The problem is that data flywheels are slow to build and easy to lose. A single enterprise customer with a strict data-residency requirement could sever the loop entirely. Correlation is a whisper; causation is the shout. The industry correlation between "AI agent market size" and "actual productivity gains" is weak. DataM Intelligence projects the enterprise AI agent market will grow from $6.65 billion in 2025 to $142 billion by 2035, a 35.5% CAGR. That forecast assumes a smooth adoption curve for tools that generate code, access policies, and background agents that can modify production systems. I have built enough stress tests to know that such curves rarely materialize without a major incident resetting expectations. The 2020 stablecoin crisis taught me that fixed mechanisms unravel precisely when liquidity is most stressed. Enterprise IT systems are no different. When an AI-generated workflow attempts to restart a service because it misinterpreted a log spike, the failure is not a programming bug. It is an incentive gap. No language model is constructed to understand the political and regulatory context of a financial institution's change-management committee. I have watched the AI-native workflow space long enough to recognize the shape of the first real damage. It will not come from an adversarial attack or a malicious prompt. It will come from automation approval fatigue. The human-in-the-loop design is the safest layer in the system. But it is also the most fragile. Once an administrator has approved fifty generated workflows without a visible error, the review process becomes a rubber stamp. The semantic distance between "the workflow passed its test" and "the workflow matches the intent of the business unit" is enormous. The AI cannot see the Slack thread where a manager approved the nuance. The auditor cannot see the prompt that shaped the code. The blockchain community solved one part of that problem by making every transaction permanently visible. Serval's code is visible in git, but only after it is generated. The generation itself is a black box. Whales don't chase hype. They chase liquidity, and enterprise liquidity is measured in compliance certificates and board-level approval processes. Serval has not publicly disclosed SOC 2 Type II, ISO 27001, or FedRAMP status. If those certifications are missing, a major cloud provider or financial institution will find them missing before the first procurement call ends. That is not a critique. It is a milestone queue that every SaaS company must pass. The order of that queue determines the valuation story. What should a skeptical observer watch over the next twelve months? Not the funding announcements. Watch the quarterly deployment numbers. If Catalyst can show that a meaningful percentage of its generated workflows actually stay in production past the first month, its "deliverable" economy will start to look like a real market. If instead the customers treat it as a trial playground and pull the agents back after a production scare, the billion-dollar narrative will shift from valuation to liability. The ledger never lies, only the interpreter does. In this case, the ledger is the set of production logs. Can an AI that writes TypeScript be audited the way we audit smart contracts? Technically, yes. Culturally, not yet. Every generation loop needs a verification loop — not for syntax, but for meaning. Until that verification loop exists as every stage, the 10% deployment rate at incumbents is a floor, not an outlier.