ServiceNow touts AI governance for its Autonomous Workforce

What happens when something goes wrong?

Fully autonomous AI agents are the long-term vision for most of the industry, but so far, such automation only rarely makes it into enterprise production. Reports frequently surface about rogue AI agents that tear down production cloud environments or threaten to delete email inboxes without authorization.

A research paper published by OpenAI researchers in September 2025 concluded that hallucinations in large language models are a mathematical inevitability. And a security researcher at AppOmni reported in November that ServiceNow's Now Assist AI agents were vulnerable to prompt-injection attacks when configured incorrectly, despite protection features, which ServiceNow addressed with documentation updates on proper agent setup.

AI governance features are common in AI agent orchestration platforms, as are platform vendors' claims that their agents are grounded in good contextual data and connected to deterministic, rules-based workflows and policies where necessary. Some early enterprise AI adopters have found that smaller, more specialized AI agents working together tend to produce better results than a single larger agent handling multiple tasks. And the initial ServiceNow L1 Specialist is handling the kind of low-level tasks some enterprises had become comfortable with during a previous wave of AIOps tools.

Furthermore, ServiceNow officials say the company is addressing AI governance worries head-on as it launches its new agents. Autonomous Workforce AI agents are designed to "know what they don't know" and escalate to a human if they encounter that scenario, a company spokesperson said in an email to Informa TechTarget this week. When a company first sets up Autonomous Workforce, it can set its own thresholds for that escalation, along with policies regulating AI specialists' behavior, the spokesperson said.

"OpenAI's research is right. Hallucinations are mathematically inevitable in large language models. That's exactly why we didn't build AI specialists on top of a language model and hope for the best," the email statement said. "Other vendors are putting AI in charge and hoping governance catches up. We are building governance first. The AI specialist is running inside a system that has been managing enterprise risk, compliance and approvals for two decades."

It will be interesting to see how effective ServiceNow's control measures are to make sure that there isn't any improvisation happening if the input is bad.

In a scenario where the IT service desk manager sees frequent escalations or handoffs to humans, ServiceNow's AI Control Tower flags those patterns and generates remediation recommendations within that manager's UI. The manager can then create or revise knowledge base articles to provide feedback to the AI specialist to refine its behavior. The human manager can also intervene in the AI specialist's workflow at any time within their existing UI, according to the ServiceNow spokesperson.

Analysts look for real-world proof, pricing

As always, automation mechanisms must demonstrate success in the real world, according to Will McKeon-White, an analyst at Forrester.

"If you don't have good inputs, then you can't get good outputs -- it will be interesting to see how effective ServiceNow's control measures are to make sure that there isn't any improvisation happening if the input is bad," McKeon-White said. "I will be very interested to see how the AI model calls are architected to ensure supremacy of the governance system, how effectively smart triggers work and how well organizations can figure out those thresholds on their own."

Forward-deployed engineers employed by a vendor but embedded within a customer's organization -- from Moveworks, which was acquired by ServiceNow last year and relaunched this week as EmployeeWorks -- "might be ServiceNow's secret weapon" for helping enterprises set up autonomous AI specialists, McKeon-White added.

"Moveworks had one of the first multi-model systems online that I saw, before anybody else was even talking about it," he said. "And they were doing forward-deployed engineering, before people were talking about forward-deployed engineering."

Even if AI governance controls do work as advertised, production guardrails won't be enough to guarantee AI safety, said Keith Kirkpatrick, an analyst at Futurum Group.

For example, workflows in IT service management, including service desk operations, were flagged as most likely to be successful for early adopters of the ServiceNow AI platform last year, since service management teams were often forced to clean up knowledge base documentation to use AIOps tools. But many enterprises still haven't completed such an exercise, especially one that's comprehensive across the organization, Kirkpatrick said.

"The ability for the organization to define thresholds for escalation is a good control mechanism that may make it more useful for organizations that want to use autonomous agents while ensuring a pre-defined level of customer experience," he said. "But organizations must ensure that their data management, organization and access are clean. This is an enterprise issue, not a ServiceNow issue."

Another open question about ServiceNow's Autonomous Workforce is its long-term cost to customers, since ServiceNow has not publicly disclosed how it will price the product, said Roy Illsley, an analyst at Omdia, a division of Informa TechTarget.

"If you can fire up an agent, which is the equivalent of a human doing an L1 role, ServiceNow have got to pitch that at a price that's less than the human," Illsley said. "The fact that [ServiceNow hasn't specified a price yet] says that they're obviously looking and hoping to find out from the private betas that are going on what people are prepared to pay based on the functionality they're delivering."

Beth Pariseau, a senior news writer for Informa TechTarget, is an award-winning veteran of IT journalism. Have a tip? Email her.