ServiceNow's New AI "Passport" Promises a Kill Switch. Its Own Customers Aren't Convinced Yet
On October 15, ServiceNow renamed its AI Control Tower the Super Intelligence Control Tower and introduced AI Passport: a cryptographically signed credential that defines an agent's identity, the actions it's permitted to take, the systems it can touch, and the human who owns it. The credential is checked at runtime, not just at deployment, which is the part ServiceNow wants enterprises to notice. In a demo cited by SiliconANGLE, a third-party customer-service agent configured to process refunds under $25 let an $18 request through automatically and blocked a $40 request at the moment it tried to execute, kicking it to human review. Nvidia is the launch partner, wiring AI Passport into its Open Agent Safety Platform so the credential checks happen inside Nvidia's own secure runtime.
The pitch is graduated control rather than binary on/off. SVP Nenshad Bardoliwalla called it a "dimmer switch": an agent can be paused by revoking its session tokens, restricted to a narrower set of tools and endpoints, or deactivated outright by killing its credentials, with every step logged and an agent restorable later with a recorded justification. That's a meaningfully different claim than most agent-governance tooling has made so far, which has mostly meant dashboards that show what agents are doing after the fact, not controls that stop them mid-action. ServiceNow is betting enterprises will pay to turn observability into enforcement.
What ServiceNow's own customers are reporting
The harder evidence came two days later, in a Breaking Analysis field study run by Qualitate between September 15 and 21 and published by SiliconANGLE on October 17. Researchers interviewed 15 large enterprises that already run ServiceNow — a sample skewed toward the installed base by design, which makes the findings more notable, not less. Eleven of the 15 had installed, piloted, or were evaluating Control Tower. But 14 of 15 described actual runtime enforcement as external to ServiceNow, custom-built, mediated by a human, limited to a handful of approved paths, or simply unproven at any real scale. Only one organization, in financial services, had tested blocking behavior in production, and only for a single model and a single agent. One security practitioner in the study estimated that only 30 to 40 percent of an enterprise's total AI activity could realistically be routed through ServiceNow's control paths today — everything running on Azure OpenAI, third-party SaaS copilots, or shadow deployments sits outside it.
Twelve of 15 respondents also flagged a dependency problem: AI Passport's permission scope is drawn from ServiceNow's configuration management database, and the credential is only as reliable as that CMDB is current. A senior engineering manager at a North American financial firm told researchers the CMDB requires constant cleanup to stay trustworthy, which means the governance story quietly degrades wherever the asset inventory lags. Cost was the other recurring complaint: of eight respondents who shared cost data, seven said spending exceeded their original expectations, with one reporting a 30 percent overrun driven by partners, data normalization, and specialist staffing rather than licensing fees.
"More like a governance orchestration layer than the actual governance of AI." — a study respondent, describing ServiceNow's AI Control Tower
Why the gap matters for buyers now
ServiceNow isn't wrong that the problem is real. More than 500 customers went live with Control Tower within six months of its original launch, according to the company, and AI-related annual contract value topped $1 billion in the second quarter with a $1.5 billion full-year target — enterprises are clearly buying the premise that agent sprawl needs a governance layer, fast. NTT Data, one of the only named customers, is using Control Tower specifically to find shadow IT and overlapping agent deployments across its own organization, which is the mundane but real problem most large enterprises are quietly fighting right now.
The timing gap is the story for anyone actually budgeting an AI governance rollout: AI Passport, shadow-AI detection, and the AI Cost Management module are all still slated for "later this quarter," not shipped. That means the capability enterprises are being sold on today — runtime enforcement with a verifiable credential — is mostly a roadmap item layered on top of a discovery-and-dashboard product that's already live. The lesson from the field study isn't that ServiceNow's approach is hollow; it's that enterprises evaluating any vendor's agent-identity pitch right now should separate inventory and alerting, which multiple products can already do reasonably well, from actual blocking, which almost none of them can prove in production yet. The buyer-side test the researchers recommend — force one real policy violation through the system and verify it gets blocked, logged, assigned, and remediated end to end before scaling — is a reasonable minimum bar for any agent-governance purchase this quarter, not just ServiceNow's.